In January, I published a piece on creativity in the sciences. I’ve been exploring serendipity, surprise, and how different types of research come about. I’m also interested in what it means to categorize research in different ways, whether as a two-dimensional basic-versus-applied spectrum or through Donald Stokes’ quadrants. This led me to Dean Simonton’s formal definition for creativity. In the post, I explore what it means to write an equation for creativity and offer a somewhat contradictory example to his definition.

Writing is an inherently dignified human activity by Celine Nguyen. I enjoyed reading Nguyen’s reflections on two years of maintaining a writing habit. This particularly resonated with me:

Early on, whenever I felt discouraged by my mediocre writing, I would cheer myself up like this. First, I would find a newsletter or blog I admired: stylish, well-written, distinctive in voice and approach. (And popular: they often had thousands of readers.)

Then, I would go into the newsletter’s archives and scroll down to the very first post I could find. It was always more raw, unpolished, and amateurish than the writing I was familiar with. I can’t describe how reassuring this was! I could see how people had become—through persistent and publicly-observable attempts—the writers that I knew and loved.

I do this too. Whenever I find a blogger with an interesting take on biology or biotech, I look for their first post and see how their writing and views have evolved. One of my writing resolutions for this year is to write more opinion pieces.

The golden age of vaccine development by Saloni Dattani for Works in Progress. An incredible essay detailing the key pieces that drove vaccine development from its conception to present times. The golden age of vaccine development is ahead of us! Essays like this give me the chills. Scientific progress is incredible, everything we do rests on the shoulders of giants, and there is infinitely more progress we can all contribute to. Also, I did not know this:

Jenner’s vaccine was spread by arm-to-arm transfer, a process that involved collecting fluid from cowpox (and sometimes horsepox) pustules with a lancet, then literally scratching it into the skin to produce immunity to smallpox in the first recipient. Subsequent recipients would develop their own pustules, from which fresh fluid was taken and scratched into others, continuing the chain of vaccination.

Is replication pro-progress or anti-risk? by Jordan Dworkin. Dworkin makes compelling points about how to think about the benefits of replication. I especially like his distinction between discovery and translational research. In my mind, replication is most valuable where individual studies or a small group of studies influence public policy or the general public narrative. Those are translational studies. Discovery research is a strong-link problem, and we should try to avoid treating it like a weak-link problem. Dworkin says,

Drug development is the canonical example. A paper reporting potential preclinical benefits of molecule X on cell line Y in a specialty journal may not generate substantial follow-on research or advance a broader subfield. But when that genre of paper is incorrect as often as it is correct, preclinical academic literature becomes an unreliable source of leads for biotech startups and pharmaceutical companies, and our system of knowledge translation trades efficiency for redundancy.

Framing papers as “correct” or “incorrect” is interesting to me, and perhaps not exactly what I look for in studies. This could be my discovery research bias, but I would be more interested in understanding the “why” of an unreplicated study. What can we take away, what might have caused it to not replicate, what confounding factors might have been missed? In any case, I’m no expert on what matters in replication, and here is Stuart Buck’s back and forth with Dworkin.

Subscribe now

What I didn’t expect about being a funder by James Özden. An interesting reflection on what you don’t know before becoming a grantmaker. I wish more people wrote posts like this about their own fields. For most professions, you don’t know what it’s really like unless you actually do the thing.

Why do research institutes often look the same? In Asimov Press by Samuel Arbesman. A great point to bring into metascience discussions. Many new research organizations start with a bold vision to be different from traditional academia, but over time they adopt familiar academic structures. There are many reasons for this, but one I hadn’t considered is the tax code. Arbesman writes,

Due to my involvement in the space of non-traditional research organizations, I speak with many people who are thinking about building new institutions. A common question that I get asked is whether to go non-profit or for-profit. This decision will impact the kind of people or organizations they approach for fundraising, the regulations they will need to adhere to, and so forth, and these things should not be taken lightly. Yet, is it not odd that our tax codes have reduced the complexity in how researchers think about science? We imagine the vast and high-dimensional space of outlier research institutions, and then are forced to collapse it into these two categories because of tax implications.

I had always thought of tax implications as downstream of decisions about organizational structure, not the other way around. Also, Overedge is a great catalog of new research organizations.

Building brains on a computer by Maximilian Schons in Asimov Press. A comprehensive essay on what it will take to truly emulate the brain. This is already a condensed form of the 175-page report Schons and team have written. They outline three core capabilities needed to emulate the brain: recording brain activity, reconstructing a brain wiring diagram, and modeling the brain by combining these data. Of these, I agree that recording brain activity is the biggest bottleneck. Recording every single cell (including neurons, astrocytes, glia, and the various neurotransmitters and neuromodulators) in the brain lags behind current wiring and modeling capabilities. Schons estimates 20 to 50 years to emulate the human brain. I would guess it will be closer to the upper end.

Horizontalization in biotech by Corin Wagen. An interesting observation about the transformation in the current biotech and drug discovery scene.

horizontalization is a natural response to increasing market size and complexity. As the market became big enough that there could be “a database company” or “an ads company,” it became advantageous for these capabilities to become their own firms rather than stay part of a single monolithic ur-company.

I liked this post because it reminded me of a constant battle in neuroscience projects I’ve been involved with: to buy or to build. You can choose the tried and true National Instruments and program your data acquisition instruments to converse with your animal behavior and other recording equipment, or you can build a custom hardware and software stack meant to be more flexible. The former is more rigid but reliable; the latter is more flexible but will probably break more often and require bespoke software.

There should be ‘general managers’ for more of the world’s important problems by Nan Ransohoff. I could not agree more with this. Why don’t the world’s biggest problems, of which there are many, have champions?

In my opinion, there should be ‘general managers’—GMs—for problems like these. These are founder-types who feel personally responsible for delivering a specific outcome (vs field-building generally); hyper-competent leaders who will pull whatever levers necessary to achieve the defined outcome. Most companies wouldn’t let an important initiative go unmanned or without a ‘directly responsible individual’ — why are we OK not having GMs for even more wide-reaching problems?

The 100x research institution by Andy Hall. An excellent report on one person’s use of AI in automating their research. This is another space where I wish people wrote more concretely about how they’re incorporating AI into their workflows, especially in science work that includes the lab bench. When I ask my coworkers how they use LLMs, the most common answer is email. Most people are only using it to write emails, which is sad. I have a post planned about what it’s been like starting research in a new field and how I’ve been using LLMs to help me learn faster.

How to be less awkward by Adam Mastroianni. As always, a super insightful piece. I feel like it was written for me.

School is way worse for kids than social media by Eli Stark-Elster. Very interesting. What problems are we trying to address? What policies are we implementing? This was well said:

Simple solutions are easy to think about and apply. But many important problems aren’t simple, which means that the simple solutions are often wrong. Social media, like fear of witchcraft and immigration, is yet another all too obvious answer to a much more complicated question: how did our children become so sad? School is also not the only right answer; as always, the truth is multi-faceted. But the data shows that it is one especially important facet.

Brendan Foody on teaching AI and the future of knowledge work on CWT. I enjoyed the second half of the conversation much more than the start. One moment got me thinking about training LLMs with current experts. Tyler asked,

If you could model a much older poet — William Wordsworth, Blake, John Milton, Rilke — some of my friends say there are no truly great poets left anymore. The best poets were way back when. Is it a goal to model the older poets and figure out what they would think, and rather than having Larry Summers and Cass Sunstein come in, that you have some AI-generated model of John Milton?

Should we be trying to capture the actual GOATs as opposed to current experts? Sure, I could be an expert in, say, biomedical sciences, but would it be better to use my expertise to train a Watson or a Crick, and then use those avatars to provide feedback to the models? It would be much easier for me to evaluate whether an agent is personifying Watson than to actually be Watson. I also found this analysis of what makes great poetry by Hollis Robins to be very interesting

Beyond the endless frontier by Ben Reinhardt. I’m intrigued to see how the NSF’s Tech Labs initiative unfolds. It’s a unique opportunity to experiment with new models of funding. But as Reinhardt writes, it’s critical that the NSF does this right to prevent Tech Labs from becoming more of the same, i.e., funding to professors at universities. I especially liked Reinhardt’s emphasis on taking risks on innovative ideas with scrappy teams rather than picking teams based on strong track records:

It’s likely that this call won’t fully address the “cold-start” problem. That is, if the program selects for teams that look like they have the best chance of getting the farthest during the tech lab, it is going to heavily favor pre-existing teams that have already gotten funding in one way or another to do as much work as possible. That precondition already puts a constraint on the novel impact that the tech labs can enable because there will be a selection effect for work that already looked promising to a more traditional funding source. It’s likely out of the scope of this initial effort, but in order to maximize the value of the tech labs program, TIP should also look at funding organizations that are able to start ambitious efforts “from scratch” to get them to a place where they would be a good fit for a full tech labs proposal. Another approach to this “cold-start” problem would be to accept a few earlier-stage teams into phase 0 based on potential and then judge them on whether they can make shocking amounts of progress in a short amount of time compared to other more mature teams.

The BBN Fund by Eric Gilliam. I’m excited for this new initiative to build more BBNs. You can read about what BBNs are here. Gilliam says,

This past year, I’ve embraced the role of a ‘field strategist’ for the BBN ecosystem. In this period — Stage 1 of the modern BBN experiment — I sought to verify that there was both demand for BBNs from ARPA-like funders and a supply of top researchers eager to found BBNs. Thanks to the UK’s Advanced Research + Invention Agency (ARIA), both have now been resoundingly verified. To provide just one data point in support: according to ARIA’s most recent fiscal year data, a small set of scrappy BBNs won more in ARIA funding during the year than every lab at the University of Cambridge combined. Stage 1 of the modern BBN experiment is now complete.

It is now time for Stage 2 of the experiment: building a “Convergent Research for BBNs.” The BBN Fund’s objective will be simple: seed a modern ecosystem of BBNs and work to maximize their overall technical ambition. If successful, we will forge a new pathway for today’s best applied, ambitious researchers to pursue ambitious R&D agendas — as Convergent Research has done with FROs.

Building science when execution is the bottleneck by Henry Lee. Lee argues for investing in automating scientific instrumentation since the scientific enterprise is bottlenecked by how fast humans can run experiments.

Experiments must be set up, calibrated, monitored, documented, cleaned up, and repeated. Errors propagate, context is lost, and human availability becomes a bottleneck. The distance between thinking and doing remains wide.

Lee further argues that general-purpose lab instrumentation does not scale for the flexibility needed to carry out various experiments, and instead we should invest in building highly specialized equipment. I largely agree with this argument given the current state of our technological capabilities. But is there a world, perhaps sci-fi for now, where a general-purpose lab robot (perhaps my robot clone) could execute experiments on highly specialized autonomous instruments?

The ground truth: crystal symmetry by Lewis Martin. A great post highlighting the importance of verifying data quality as datasets grow ever larger. Trained human experts are still needed to validate data quality. This reminds me of the post about AI radiologists and keeping humans in the loop.

Just one scientific publication this time, because it’s a big one:

A cool methods paper, Correlative voltage imaging and cryo-electron tomography bridge neuronal activity and molecular structure by Jung et al., combines voltage imaging to measure neuronal electrophysiological patterns with cryo-electron tomography to visualize molecular structures within the same neurons. Their goal was to link the internal molecular architecture of neurons with their electrical firing patterns, i.e., correlating structure with function.

The experimental approach is elegant: researchers grow neurons on tiny metal grids that fit under both optical and electron microscopes. First, they electrically stimulate the neurons while recording their responses with a fluorescent voltage dye. This allows them to group neurons into three categories based on how strongly they respond to stimulation. Immediately after imaging (within 15 minutes), they flash-freeze the grids to preserve the cells in their natural state. They then use cryo-electron tomography to capture nanometer-scale images of structures inside the cell body. Finally, they analyze the distribution and structural states of ribosomes (the cell’s protein-making machines) in each neuron.

While I think the development of this new method is excellent and correlating live function measurements with molecular structure is very cool, I have a few questions:

Their reasons for studying ribosomes, instead of molecules that actually drive electrical activity, don’t seem grounded in biology. Voltage-gated sodium and potassium channels directly control neuronal firing, but they’re too small and sparse to reliably visualize with current cryo-ET techniques. Ribosomes are large, abundant, and easier to image at high resolution; this seems like a “looking under the streetlight” problem. The connection between electrical responsiveness and ribosome structural states is indirect at best, requiring multiple inferential leaps (activity to translation changes to ribosome states, which occur over hours, not minutes). The paper demonstrates technical feasibility (we can correlate imaging with cryo-ET) but picks a molecular target based on what’s technically tractable rather than what’s biologically most relevant.

They also don’t actually know what types of neurons they’re looking at. There are no measurements of neuronal maturity, cell type, or health beyond the voltage imaging. The clusters could reflect differences in cell age, damage, or identity rather than true electrical properties. Over half the neurons (55%) don’t respond to stimulation. Are these immature, damaged, or just different?

The measurement and analysis of firing patterns also seems a little strange. They measure fluorescence changes from a voltage-sensitive dye, and the signal could vary based on dye uptake, not just electrical activity. Imaging at 100 Hz (every 10 ms) is too slow to properly resolve action potentials that last 2-3 ms. Their “decay parameter” measures fluorescence amplitude at 60 ms after the peak, but action potentials are completely over by 3 ms. They might be measuring network reverberation or noise, not spike kinetics. Only 5-8 neurons per group for cryo-ET also seems too small. What does intragroup variability look like for ribosomal states? Would that account for all these differences?

Overall, I think this is a really cool technique! It’s directly measuring function and structure within the same cell populations. I’m just cautious about what conclusions we can draw from their choice of biological application.

For reference, a recent review of advances in cryo-electron tomography.

That’s it for today folks! I will have a couple bigger updates to share next month. Thanks for reading.

Cover image: Molecular Clarity—Discovering What’s Possible with Cryo-Electron Tomography. Thermo Fisher Scientific. Source.

Optogenetics transformed circuit neuroscience. It enabled researchers to probe specific brain networks, such as those involved in memory, addiction, or Parkinson’s, and provide causal mechanistic explanations about their function. While optogenetics is far from being ubiquitous in the clinic (one does not simply genetically modify humans), it has a lot of potential. A 2021 paper showed the use of optogenetics to partially restore vision in a blind patient.

It’s rare enough for a scientist to transform their field once, but Boyden did it again ten years later. This time, instead of pond algae, he used baby diapers to develop a method called expansion microscopy. Typically, when scientists have tried to see small structures within biological cells, they have developed new imaging tools with higher magnification capabilities. Boyden was the first to flip the equation and ask: can we make the object itself bigger instead of making the microscope better? Inspired by the remarkable absorbing and swelling properties of a baby diaper, his lab developed a method to enlarge the brain. This method is now being used to map every neuron connection across the mouse brain.

Boyden has a method, which he calls Tiling Tree, to search the idea space for all possible ways to solve a given problem. This tool for thought enables him to think outside the box and search for ideas that might otherwise be missed. The tiling tree method contributed to the creation of both optogenetics and expansion microscopy.

If you have a system to generate creative solutions, is that still considered creative?

The creativity literature might say no, that creative ideas cannot be generated by systematic analysis. Papers on creativity suggest that the creative process cannot be algorithmic but rather relies on heuristics. This kind of creativity has of course played a role in many of the major scientific discoveries and innovations. Famously, Archimedes stepping into a bath and noticing the water level rise led to his eureka moment about buoyancy and displacement. Kary Mullis has said that the core idea for the Polymerase Chain Reaction (PCR) came to him while driving.

But intuitively, to me, both optogenetics and expansion microscopy are creative scientific innovations borne out of systematic analysis. This is what I’m calling the Boyden paradox. Is Ed Boyden creative or uncreative? And why does that even matter?

To answer whether Boyden is creative or not, we first need to define what creativity actually means. If we can understand what makes ideas truly creative, we might be able to generate more of them. We might even be able to engineer creativity and therefore discovery itself.

Subscribe now

Defining creativity

I’ve been reading some literature on creativity, specifically the work of Dean Simonton, professor emeritus in the Department of Psychology at UC Davis. Simonton is known for his research on genius, creativity, and the factors that contribute to eminence across domains of science and art.

Simonton proposes that creativity requires three factors: originality, utility, and surprise. An idea is creative when we maximize all three. Critically, each factor is necessary but not sufficient, meaning that you need all three together.

Let’s define what each of these terms mean:

Originality: Simonton defines originality in terms of the idea’s initial probability (p). How likely was the individual to think of this idea when first encountering the problem? Original ideas have a low initial probability, meaning it’s unlikely that this particular person would come up with this idea when they first start thinking about the problem. On the other hand, unoriginal ideas have a high initial probability, meaning they’re among the first ideas that come to mind.

Since p (a probability) ranges from 0 to 1 (where 1 is highly probable and 0 is highly improbable), originality is defined as the inverse: originality = 1 - p. A routine idea with p = 1 gives originality = 0, while a low-probability idea with p approaching 0 gives originality approaching 1.

Utility: The idea’s final utility (u) indicates its usefulness, effectiveness, or value once the creator considers the solution finalized. Utility is a continuous variable between 0 and 1. While some problems have binary outcomes (u = 1 if it works, u = 0 if it doesn’t), many ideas have partial utility, where u falls somewhere in between when the solution works, but not perfectly.

Surprise: Surprise is defined in terms of the creator’s prior knowledge (v) about the idea’s utility (not the idea itself). At the moment of conception, does the creator know if it will work? Prior knowledge, v, is also between 0 and 1. If v = 1, the creator already knows the utility in advance; if v = 0, the creator is in the dark about the idea’s future utility.

Surprise is the inverse: surprise = 1 - v. When v is low, you learn something new by trying the idea, regardless of whether it succeeds (high utility) or fails (low utility). Surprise adds the dimension of time. It’s the change in knowledge through the creative process. This change in understanding from unexpected outcomes is what Simonton calls surprise or serendipity.

But why these three factors? Where do they come from?

When we call something creative, we usually point to work that stands out as something different from what came before. That’s originality. But difference alone isn’t enough. We also expect creative work to accomplish something, to solve a problem or produce value. That’s utility. There’s also a sense that the creator didn’t just mechanically apply known techniques. There’s an element of discovery, of the creator learning something through the process. The outcome wasn’t obvious or guaranteed beforehand. That’s surprise.

These three factors emerge from how we use the word “creative.” We don’t call routine work creative, even when it’s useful. We don’t call random novelty creative if it serves no purpose. And we usually don’t call the straightforward application of expertise creative, even when it produces excellent results. Creativity seems to require all three: trying something unusual that works in ways that weren’t fully predictable.

Regardless of whether you agree that these three factors exactly describe creativity, let’s assume it does and continue with Simonton.

Simonton formalizes his concept of creativity through the following equation:

Creativity = originality · utility · surprise, or

C = (1-p) · u · (1-v)

To make this concrete, Simonton identifies 8 possible combinations when p, u, and v approach either 0 or 1. I found this exercise helpful for comparing different types of ideas:

There are seven ways to be uncreative:

Case 1: Originality = 0, utility = 0, surprise = 0 (irrational repetition)

No originality, no utility, no surprise. This is doing something habitual that you know doesn’t work, like checking your jacket pocket for car keys for the tenth time when you know they’re not there. “The definition of insanity is doing the same thing over and over and expecting different results.”

Case 2: Originality = 0, utility = 0, surprise = 1 (problem finding)

The idea is routine but you’re surprised it has no utility. A home example: you make sourdough with your usual starter, but this time the bread doesn’t rise. Nothing novel about your process, but you expected it to work. This surprise reveals hidden problems – maybe the starter went bad, or the temperature was off. A lab example: a routine calibration failing can reveal instrument problems. Problem finding isn’t creative itself, but it can lead to creative solutions.

Case 3: Originality = 0, utility = 1, surprise = 0 (routine, habitual ideas)

This describes most of our day-to-day thinking. Ideas we’re likely to generate that we know have high utility: your usual route to work, routine cell media exchanges, your toddler’s bedtime routine. Habits that let us navigate everyday life. Nothing extraordinary.

Case 4: Originality = 0, utility = 1, surprise = 1 (lucky guess)

Nothing original, but you’re surprised by the positive outcome. An action that usually has no utility ends up having unexpected value. You buy lottery tickets every week with no wins. One day you win big. Buying ticket number 100 wasn’t creative, your expectation of payout was low, but you’re surprised when it pays off.

Case 5: Originality = 1, utility = 0, surprise = 0 (rational suppression)

High originality but low utility, and you knew ahead of time it wouldn’t work. Like trying to expand the brain by stretching it with your hands. It would be an unusual idea, but you’re confident it won’t work. You use expertise to rule this out, consciously or unconsciously.

Case 6: Originality = 1, utility = 0, surprise = 1 (blissful ignorance)

This one confused me. It suggests an original idea where you were surprised it didn’t work. Meaning you thought it would succeed but it failed. Overconfidence from expertise, perhaps. But Simonton calls this “blissful ignorance” where an idea is never tried. This reveals how parameter values matter: are we setting them to exactly 0, or approaching 0? There’s some inconsistency in Simonton’s treatment.

Case 7: Originality = 1, utility = 1, surprise = 0 (irrational suppression)

An original idea with high utility that you know will work, yet you rarely do it (low initial probability). Deliberately not doing something you know is good for you: exercise, eating healthy. A science example might be sticking with traditional methods when better ones exist, like continuing to use outdated statistical methods, or avoiding better model organisms.

One way to be creative:

Case 8: Originality = 1, utility = 1, surprise = 1 (maximum creativity)

Finally, the creative case. After seven ways to be uncreative, it’s clear why we need to maximize all three factors. Low initial probability (originality) indicates that the idea isn’t obvious. High utility indicates that the idea needs to work. Low prior knowledge (surprise) indicates that you don’t know ahead of time it will work.

Expansion microscopy as an example

Let’s see how expansion microscopy satisfies all three criteria:

Originality: When everyone else was working on making microscopes better, Boyden thought to make the object itself bigger. This had low initial probability; It isn’t among the first ideas that come to mind when thinking about imaging small structures. Most microscopists would naturally think about improving lenses, detectors, or light sources, not inflating the sample. Originality ~1.

Utility: The technique works. Scientists are using expansion microscopy to map brain networks at single-cell resolution. Groups beyond Boyden’s lab have adopted it. For instance, E11 Bio is using it to map every neuron connection in the mouse brain. The method has also proven useful across many applications. Utility ~1.

Surprise: Boyden didn’t know ahead of time whether swelling brains would work. There was a significant chance the effort failed – the brain could have torn apart or the cellular structure could have distorted. There was genuine uncertainty about whether this approach would preserve the precise spatial relationships scientists needed to see. Only by trying it did they learn it worked. Surprise ~1.

With all three factors maximized, expansion microscopy scores high on creativity: C = (1-p) · u · (1-v) → 1.

Resolving the paradox

A key component in the creativity field (yes, there is a creativity field) is that problem solving must be heuristic rather than algorithmic for ideas to be creative. Once you have a procedure to generate solutions, it ceases to be creative.

But this is exactly what Boyden has done. He believes that “creativity, and problem solving, like any other skill, can be learned and taught.” In his tiling tree method, you start with a problem and systematically map out all possible solutions. He claims this forces you to consider solutions you would normally not think of.

Boyden’s group uses this method regularly. It played a role in both optogenetics and expansion microscopy. And he’s not alone. Fritz Zwicky, a Swiss astronomer, developed his own systematic method for investigating complex problems. Using this approach, he coined the term “supernova,” described neutron star formation, and was the first to propose dark matter.

To me, this is engineering at its best. Taking the process of discovery and reverse-engineering it to become reproducible. Meta-creative, if you will: becoming creative about your own creative process.

This seems to directly contradict the creativity literature, which excludes systematic ideation as creative. But maybe Simonton and Boyden aren’t actually in conflict. Maybe they’re describing different aspects of the same process.

Boyden’s Tiling Tree Method doesn’t eliminate the need for uncertainty or a flash of insight. Instead, it’s a systematic way to find the uncertain ideas worth pursuing. The tree maps the solution space comprehensively, forcing you to consider branches that seem unlikely or unreasonable. But you still have to choose which branch to climb without knowing if it will bear fruit.

The method lowers p (makes you think of unusual ideas) without raising v (you still don’t know which will work). This preservation of uncertainty is crucial. As Michael Nielsen has argued, scientific breakthroughs often require a willingness to pursue ideas that don’t yet make rational sense. A perfectly rational actor would only pursue ideas with high known utility. But those are precisely the uncreative ideas in Simonton’s framework. Creative breakthroughs require taking bets on low-v ideas.

In this sense, Boyden isn’t necessarily engineering creativity itself. He’s engineering the conditions that make creative insights more likely. He has even talked about this process as engineering serendipity. He’s systematizing the search process while preserving the essential uncertainty that Simonton argues is necessary for creativity.

Some unresolved thoughts on Simonton’s framework for creativity

I find the qualitative concepts of originality, utility, and surprise useful for discussing creativity, but I’m unsure about the quantitative aspects. Simonton offers an equation, but provides no method for assigning numbers. How original is expansion microscopy? We know it has high originality, but is that 0.8 or 0.99? What would the difference between those numbers even mean?

Similarly, we know expansion microscopy has utility, but it also has limitations like uneven expansion in certain tissue types, for instance. Does that make u = 0.95 instead of 1.0? How much does each limitation reduce the utility score?

More fundamentally, how would one actually apply this equation? Consider a thought experiment: two people in different rooms independently solve the same problem. One describes the solution in two pages of detailed prose; the other gives a succinct one-paragraph explanation of the same solution. The longer description has lower probability of occurring than the shorter one (fewer people would spontaneously write exactly those two pages). Does that make the verbose solution more creative, even though both produce the same result?

Another aspect is expertise. Let’s take another situation with two people in different rooms solving the same problem, this time let’s assume they each come up with a similar one-paragraph description, except now one individual is a domain expert while the other is not. Given that the domain expert probably has a higher likelihood of coming with that solution, does that naturally always make a novice more creative?

Finally, how does creativity in the arts fare with this equation? How do I quantify the initial probability or surprise at the utility of Picasso’s work? Does this framework suggest Picasso was creative, and was he more or less creative than Matisse? There’s a line in the movie Midnight in Paris where a character in the 1920s claims, “Matisse is the greater painter, but Picasso is the greater artist.” While completely fictional, the distinction is telling: “artist” implies creativity, while “painter” suggests skill or technical mastery. Can Simonton’s equation capture this difference?

Can we engineer creativity?

Maybe you can’t engineer the creative idea itself since that would raise v and kill the surprise. But perhaps you can engineer the process that generates creative ideas. You can create systems for exploring the improbable. You can develop algorithms for finding heuristic problems. Boyden’s Tiling Tree doesn’t guarantee which branch will succeed (keeping v low), but it systematically forces exploration of low-probability solutions (lowering p for unusual ideas).

Why does any of this matter?

First, I’m interested in understanding how scientific discoveries happen. Creativity plays a crucial role in major breakthroughs, from penicillin to CRISPR to optogenetics. If we can develop a robust theory of the creative process, we can build environments that enable creative outputs.

Second, generative AI presents an interesting test case. We often ask whether ChatGPT is creative. Can we apply Simonton’s equation to formally evaluate it? Based on the complications we’ve explored, any straightforward application seems problematic: a more verbose model that adds filler text would rank higher in originality simply because longer outputs have lower initial probability.

If we can develop a robust, quantitative theory of creativity, analogous to what Shannon did for information theory, that could transform how we build AI systems. What would an LLM or image model be capable of if we could optimize with a rigorous measure of creativity?

I don’t have the answers to all these questions, and I’ve only scratched the surface of Simonton’s work, let alone all the other literature and commentary on creativity. I find Simonton’s framework useful for thinking about it, even if I have questions about the applicability of the math.

I’m very much thinking out loud in this essay. If you made it this far, and are interested in discussing any of these ideas, please DM me or drop a comment!

My interest in exploring Simonton’s research comes from some very interesting conversations with Jeff Tsao. Thanks to Gavin Brown for all the discussions and feedback on the drafts of this essay.

Cover Photo: Photo of Picasso working on Guernica. Source.

14. Arjun Raj’s essay on quitting projects in science. I fully agree with this. Scientists should quit projects more often. Given that projects can take 2-4 years to complete, you don’t get to work on many over the course of your career. Make what you work on worth your while. I made a significant pivot in my research direction earlier this year. There were many reasons why I left my old project behind (after working on it for three years), but one was a lack of excitement about the long-term research direction it was taking me toward.

15. I saw this post on Marginal Revolution the day after this fMRI paper went viral. The virality stemmed from fears that everything we’ve known about fMRI has been wrong. This is not entirely true. A few points that were on my mind:

    1. The viral tweet misstates the paper’s claims. It suggests the fMRI signal was misaligned with underlying neural activity in 40% of people. What the paper actually reports: the BOLD signal in 40% of brain voxels (the 3D equivalent of pixels) showing a significant change exhibited opposing changes in oxygen metabolism. This corresponds to ~22% of all gray matter voxels analyzed.

    2. The paper compares indirect measures of neural activity. It measures cerebral blood flow, cerebral metabolic rate of oxygen, and oxygen extraction factor, in addition to the BOLD signal. All of these only indirectly measure neural activity. The paper compares two indirect measures to each other without directly measuring neural activity itself. Previous papers have measured EEG alongside fMRI, providing a more direct comparison. Beyond neural activity (the firing of neurons), other processes can contribute to hemodynamic changes, including astrocyte activity and neuromodulators. What the paper does demonstrate is that hemodynamic responses are more heterogeneous than canonical models assume, particularly for negative BOLD signals. It’s a valuable contribution to understanding physiology, even if I think the claims about “neuronal activity” are overstated.

    3. This paper fits within a rich history of neurovascular coupling research. There’s a long history of research on neurovascular coupling and what hemodynamic responses in fMRI indicate. This is one important paper, but it didn’t emerge in a vacuum. Thousands of scientists have tackled neurovascular coupling and fMRI interpretation. My own PhD work was part of this field. Knowing when and how to situate new papers within prior literature is crucial.

16. Continuing my literature review of high-resolution cellular structure with cryo-electron tomography (cryo-ET), I’ve been doing a deep dive into the work of Julia Mahamid and Bronwyn Lucas. Mahamid and Lucas have both advanced in situ cryo-ET to visualize macromolecules within intact cells rather than as isolated protein structures. Both use correlative approaches to bridge fluorescence microscopy with nanometer-resolution structural detail to reveal organizational principles across cellular compartments. I found this review to be a helpful broad overview of the current state and potential of cryo-ET.

17. Mahamid has a recent bioRxiv preprint capturing ribosomal complexes through the entire translation process, from initiation to recycling, within cells. This study is a major technical achievement in visualizing the protein synthesis machinery operating inside bacterial cells at near-atomic resolution. While scientists have studied translation extensively in test tubes, this work captures the full complexity of how ribosomes actually function within the crowded, dynamic cellular environment. The Mahamid group previously demonstrated atomic resolution imaging of ribosomes within bacterial cells.

18. To image cells under the electron microscope, they need to be thinned to ~200 nanometers. Cells, however, are much larger, in the micrometer range, and are thinned using a technique called focused ion beam (FIB) milling. Cell samples are placed inside a vacuum chamber, and an ion beam mills away the tops and bottoms of cellular regions of interest. This milling process can introduce artifacts in cellular structures that, if unaddressed, can impact downstream structure determination. This paper by Lucas provides an in-depth analysis of how the ion beam damages cellular structure during FIB milling. This isn’t particularly glamorous work, but it highlights the complexities and care required when resolving structures at the atomic scale within intricate cellular environments.

19. I recently read An Immense World by Ed Yong and loved it. It was eye opening to learn the specifics of how different animals sense the world around us. Scallops have mini DMDs for eyes! I am currently reading two books: Brian Potter’s book The Origins of Efficiency and Vaclav Smil’s Energy and Civilization: A History.

20. A few other misc December readings: English has become easier to read, Vibecession: Much More Than You Wanted To Know, The Boring Part of Bell Labs.

21. Finally, some year-end things:

    1. It’s been exciting to start blogging about science and metascience this year. I’ve been publishing scientifically for over a decade, but 2025 was the first year I wrote more broadly about scientific progress. From writing for Asimov Press to becoming a writing fellow at the Roots of Progress Institute and attending the Progress Conference, 2025 was the year I became an active participant in the Progress community. This is the most intellectually stimulating and ambitious crowd I’ve been part of, and I spend most of my time around PhDs. I’m looking forward to contributing to these conversations in 2026.

    2. This year also saw a pivot in my research from systems neuroscience to structural cellular biology. This came with a family move from Seattle to Madison. I’ve thoroughly enjoyed diving into the world of cryo-electron microscopy and tomography. It’s been fascinating transitioning from a lab where most instruments were custom builds to a field with standardized instrumentation. Standardized instruments don’t mean standardized workflows, though. Generating the perfect vitrified sample for imaging under the electron microscope is very much an art form. Every sample type, every cellular system requires its own optimization. My conviction for in situ cryo-electron tomography is strong. Understanding molecular machinery in its native context to bridge cellular and structural biology feels like the future, and I’m excited to be part of it.

Happy New Year from Bangalore! I hope you stick around for a lot more of my writing in 2026.

Cover image: Magic Mirror by M. C. Escher. Source.

Science is being rebuilt from the ground up. Across new institutions, tools, and funding models, a new generation of scientists and builders is rethinking how scientific and technological progress happens.

At the front of that movement is metascience, the study and redesign of how science operates. In the roughly six years since Patrick Collison and Tyler Cowen kicked off the “Progress Studies” movement, the metascience community has been marching forward across all dimensions. By establishing bold new formats of scientific institutions such as the Arc and Astera institutes, developing ambitious funding models such as the U.K.’s Advanced Research and Invention Agency (ARIA), and supporting a new frontier of Focused Research Organizations (FROs) through Convergent Research, the metascience community is making its mark outside traditional university models of scientific research.

These initiatives signal that the metascience movement has reached a turning point. What began as a conversation about the stagnation of science has evolved into an ecosystem of builders experimenting with new ways to fund, conduct, and distribute science. The excitement at this year’s Progress Conference underscores our movement’s collective realization that progress not only depends on increasing understanding but also on implementing change in the machinery of discovery.

Subscribe now

AI for science

Artificial intelligence (AI) for science was one of the defining themes at Progress Conference 2025. Everyone from OpenAI CEO Sam Altman to Michael Kratsios, director of the White House Office of Science and Technology Policy, spoke about leveraging AI infrastructure to advance science. As a life sciences researcher who has tried to integrate AI into my experimental workflows but still spends long hours at the bench, I often find these discussions either too abstract or detached from the manual bottlenecks and slow feedback loops of research, but the talks at this year’s conference felt refreshingly concrete.

I was particularly struck by Seemay Chou’s presentation on re-engineering biological data generation for a world with AI agents. Chou, co-founder of Arcadia Science and the Astera Institute and board chair at The Navigation Fund, argued that as machines become active participants in research, science itself must be redesigned around AI agents as the primary operator. To make AI systems genuinely productive, she said, we need radically open, high-quality datasets that capture the full complexity of nature rather than curated subsets optimized for publication.

To put this vision into practice, Chou’s team is building what it calls the Protein Data Bank 2.0. The original Protein Data Bank has served as the central repository for three-dimensional protein structures for more than 50 years. Now numbering over 200,000 entries, it has been fundamental to breakthroughs like AlphaFold. Over 80% of the Protein Data Bank’s structures were determined using X-ray crystallography, and during the interpretation of X-ray diffraction data, much of the so-called “background” diffraction, which appears as fuzzy, low-intensity scattering that reflects alternative protein conformations or dynamic motion, is typically discarded, even though it contains rich structural information. Chou’s team, spanning Astera and several academic collaborators, is re-engineering this pipeline to retain and reinterpret that lost information.

The technical details matter less than the philosophy behind them: deliberately questioning decades-old conventions that simplified complex biological data for ease of human understanding. With AI as a research partner, we don’t need to strip away that complexity anymore. The messiness of biological data, all the signals buried in noise, may be exactly what provides context and helps machines uncover deeper structure. When I listen to Chou’s vision, I don’t just see it as modernizing crystallography; I see it as reimagining how humans and machines will work together to make sense of nature.

I also see this as a broader shift in the culture of science. In academia, experiments are typically designed to produce a cohesive story for publication, which often results in incomplete datasets or the absence of essential metadata needed for reproducibility. Throwing AI at those fragmented datasets is unlikely to yield insights; it will simply amplify the gaps. By treating the dataset — and not the paper — as the scientific output, Chou’s group is rebuilding the incentive structure from the ground up.

Their collaboration has taken a radical step by foregoing journal publications. Instead, all labs involved have agreed to release open datasets, lab notebooks, and even Slack transcripts as living records of discovery. The aim is to make the work so useful that it spreads by adoption, not citation. It’s a bold experiment in both infrastructure and culture, one that could redefine how research is done in the age of intelligent machines. There are, of course, many efforts to integrate AI into the analysis stage of science, but I’ve seen far fewer instances where the machine is treated as a true partner.

New research organizations

If Chou’s work represents a bottom-up reimagining of how scientific data is generated for the intelligence age, Anastasia Gamick’s vision takes a top-down view of how entire research systems must evolve to support it.

As the co-founder of Convergent Research, Gamick has spent the past four years designing and launching Focused Research Organizations (FROs), which are time-bound, startup-style nonprofits that fill the institutional gaps left between academia, industry, and government. She argues that progress stalls not because we run out of ideas, but because we run out of tools, such as shared datasets, instruments, and platforms that make new fields possible. By “pre-standardizing” the way research is organized, we have inadvertently optimized for individual breakthroughs instead of building the trunks of the technological tree that sustain all future branches.

FROs target the kinds of projects unlikely to receive traditional funding — the medium-scale, coordinated ones that are too engineering-heavy for academia, too pre-competitive for venture capital, too modular for mega-projects, and too focused for ARPA agencies. One example is E11 Bio, a neuroscience-focused organization building a brain connectomics pipeline using molecular, optical, and computational techniques that crush costs and enable brain-wide scaling. A project like this would struggle to find a home in academia, where grants typically support small, hypothesis-driven studies rather than multiyear engineering efforts to build core technologies. By investing in infrastructure, E11 Bio aims to make brain mapping orders of magnitude cheaper and faster, laying the foundation to address unanswered questions: how brain circuits are organized across the mammalian brain and how structure gives rise to function across the whole brain.

The early lesson I take away here is that, by systematically laying out the “trunks” and “branches” of the technological and scientific tree, we can uncover more FRO-shaped opportunities than expected, with lots of low-hanging limbs. I think the key is that FROs don’t replace the existing research ecosystem — they augment it.

I’m curious to see how the growth of these trunks plays out in the intelligence age. At least for now, true embodied intelligence, like the kind that can manipulate pipettes or align optics, is likely to emerge more slowly than cloud-based intelligence. Our ability to use AI agents for hypothesis generation, experimental design, and data interpretation will occur before robots take over the physical lab bench. That makes the human task clearer: to build the instruments, datasets, and collaborative platforms that these disembodied intelligences will need to operate effectively. In that sense, the FROs of today may be laying the groundwork for environments where both humans and machines can make discoveries.

Most of the new research models emerging from the progress community lean toward engineering-heavy capability building. That makes sense to me — AI will only unlock real discovery once the underlying tools, datasets, and platforms exist. But I also find that much of the conversation right now is concentrated on that side of the spectrum. Less attention is being paid to the exploratory, curiosity-driven work that has historically produced many of science’s paradigm-shifting breakthroughs.

We can’t do everything at once. It’s probably wise to start by building the right tools, and this new wave of alternative research structures is still in its early days. Still, I often wonder whether in our focus on infrastructure and capability, we risk overlooking the kind of open-ended inquiry that drives the biggest leaps.

Jeffrey Tsao, a senior scientist at Sandia National Labs, offered one way to address this gap.

We often think of progress as a one-way street: scientists uncover new knowledge, which engineers then turn into tools and products. Tsao’s core claim is that the reverse path, where we use existing technologies and real-world systems as sites of discovery, can be just as powerful. By studying how technologies actually work in practice, we can uncover new scientific principles and generate knowledge that immediately connects back to use. He notes that this dynamic once thrived in the great corporate research labs of the 20th century, such as Bell Labs, Xerox PARC, and GE Research. Scientists at these firms investigated fundamental questions that emerged from the technologies they were building. When those labs disappeared, we lost not only the discoveries and innovations but also a unique model of how science and engineering can drive each other in a continuous loop.

Tsao’s vision resonated deeply with me. As a biomedical engineer, I’ve always been drawn to the intersection of technology and discovery, the place where a new instrument can make an old question answerable or unlock a whole suite of new questions. The most meaningful work, to me, is not building tools for their own sake but developing technologies that enable specific, previously unreachable insights about how life works. Progress isn’t just about identifying where science has stalled but about designing the institutions and mechanisms that help it move again.

Alternate funding models

Building new research organizations is only half the challenge. The other half is designing the incentive systems that allow them to thrive. Even the most visionary scientists can stall if trapped in rigid funding cycles or narrow review panels, and across the progress movement, a growing number of builders are asking: What if we redesigned the structure of funding itself?

One answer comes from Ilan Gur, the founding CEO of the Advanced Research and Invention Agency (ARIA), an R&D funding agency that has secured £1 billion from the U.K. government to engineer progress by rethinking how research is funded, not just done.

For Gur, progress is not an abstract concept but a systems-engineering problem. The challenge is not a shortage of ideas, scientists, or money; it’s that our massive, bureaucratic funding ecosystem scatters these resources too thinly to create the catalytic conditions where innovations happen. ARIA’s answer is to compress that reaction. Program directors are empowered to define bold “opportunity spaces” and then recruit top talent across disciplines and institutions to pursue them with flexible, milestone-driven grants. Gur’s formula is deceptively simple: align people, environment, and resources, and “magic happens.”

From my own experience, I’ve come to believe Gur’s equation captures something fundamental. People, environment, and resources are not interchangeable ingredients but sequential catalysts. You can have money and infrastructure, but without the right people, the system stays inert. The wrong environment can smother the creativity of brilliant people. Only when all three align does progress really ignite. I’ve seen how this sequence holds true. When the right minds are brought together and given the freedom to explore — the “free play of free intellects,” as Vannevar Bush once called it — resources become accelerants. Gur’s philosophy formalizes that intuition into institutional design. Instead of forcing people to fit pre-set programs, ARIA builds environments around them.

Tom Kalil, CEO of Renaissance Philanthropy, is also rethinking funding and incentives. He advocates shifting from “push” grants to “pull” mechanisms, which are like funding models that pay for outcomes rather than intentions, much like NASA’s partnerships with SpaceX or Operation Warp Speed. Caleb Watney, cofounder of the Institute for Progress, proposed X-Labs: large-scale, flexible grants to independent research organizations designed to explore bold, interdisciplinary questions that universities and startups can’t. Both approaches share a conviction that funding should act less like paperwork and more like propulsion by rewarding results, scaling what works, and giving science the institutional freedom to experiment with itself.

Ultimately, I find that what’s missing from today’s research funding ecosystem is not just more money, but incentives that encourage healthy competition between funders. Research funding in the U.S. remains remarkably one-dimensional. Most public science is funded by the National Institutes of Health (NIH) and, to a lesser extent, the National Science Foundation (NSF), agencies that overwhelmingly fund small, principal investigator-led grants designed for an earlier era of individual inquiry. This structure rewards safety over exploration and leaves little room for institutional or methodological experimentation. Unlike in venture capital, where investors compete to identify and back the most promising, high-risk ideas, science funders face no comparable pressure to seek out unconventional bets or refine their own models.

The result is a stagnant funding ecosystem that selects for conformity rather than creativity. A truly dynamic research economy would mirror the process of evolution: many funding mechanisms, each taking different approaches, competing for the best ideas and talent. If we can build that kind of diversity in our funding landscape, with real competition and variation, we might restore the adaptive spirit that science depends on.

Looking ahead

One question that still feels unresolved to me is how we can best judge and support curiosity-driven research in a world that’s moving away from traditional academic structures like journals. Efforts to reform science have generally focused on building tools, platforms, and datasets, which are projects with clear deliverables that can be measured and funded in short cycles. But curiosity-driven research doesn’t fit neatly into that framework. For all their flaws, journal publications have at least provided a mechanism to evaluate and reward the early stages of long-arc inquiry.

Pull mechanisms like those proposed by Tom Kalil work well when success can be defined up front. But for exploratory work, where the outcomes are unknown by definition, that model might not be ideal. Michael Nielsen, a research fellow at the Astera Institute, and Kanjun Qiu, cofounder and CEO of AI startup Imbue, have proposed a “Century Grant Program” as a thought experiment — it would fund research for a hundred years through an endowment model. It’s a beautiful idea, but who, exactly, would be willing to take that bet, and on whom?

As we experiment with new funding structures and build the next generation of research institutions, this is the next challenge we need to face: How do we create the patience, trust, and evaluative frameworks needed to sustain slow, curiosity-driven science in a culture that prizes speed and measurable outputs? How will we recognize progress when it unfolds over decades and might only become visible long after its creators have moved on?

I left the conference feeling optimistic. The metascience movement is no longer just diagnosing the problems of modern research; it is building a new era of discovery. From AI-enabled data generation to FROs to new funding architectures, the community is learning by doing. If the past few years have shown anything, it’s that when the right people are given the freedom to build, entirely new possibilities for science — and for human progress — can begin to emerge.

1. My deep dive into The Making of the Electron Microscope was published in Asimov Press last month. This piece was challenging to write: it weaves together multiple inventors, the scientific environment that made the microscope possible, and its uneven path to commercialization, not to mention that I began the project knowing very little about electron microscopy. I always love working with the Asimov team. I feel myself transforming as a writer every time I go through their editing rounds. (The cover photo for this post is used in the Asimov piece and credited to David Goodsell.)

2. As a Roots of Progress Institute Fellow, I attended the 2025 Progress Conference in October. I was honored to write an essay reporting on the metascience track for Big Think’s special issue, The Engine of Progress. I wrote about how, since the official launch of Progress Studies in 2019, the metascience community has grown from diagnosing the stagnation of science into experimenting with new research organizations, funding models, and policies. I love that the metascience community within the progress movement is truly a group of builders taking real action.

3. I highly recommend reading the other fellows’ reports on the different conference tracks: Afra Wang on American Dynamism, Laura Mazer on Longevity, Jeff Fong on Policy, Grant Mulligan on Climate and Energy, and Anton Leicht on AI Protopia.

4. I’ve also been excited about Seemay Chou’s work at Astera in rebuilding the X-ray crystallography analysis pipeline. The Diffuse Project, which includes a team at Astera as well as investigators across universities, has intentionally deprioritized journal publication in favor of building comprehensive datasets as the primary output. Here’s the blog post. Here’s the Diffuse Project website. I hope we see many more efforts like this. High-quality, well-curated “ground truth” data is a bottleneck for building many more useful AI systems in biology.

5. That said, creating more datasets doesn’t solve everything. Stuart Buck makes an important point about which data we collect and how it’s structured. Data without the right metadata, like in the Childhood Cancer Data Initiative, can be far less useful than it looks. As Buck writes: “A takeaway here: If NIH wants to upscale a pediatric cancer database so that it is most useful to AI researchers, it needs to work closely with top medical and AI researchers who will have the best on-the-ground insights as to what would make the database actually useful.”

6. I’ve also appreciated some more measured takes on AI in biology. I found Claus Wilke’s essay, We still can’t predict much of anything in biology, a refreshing take on how far we’ve come and how far we still have to go in using AI to “solve biology.” Ruxandra Teslo writes about Eroom’s law and the need for predictive validity in our AI agents for drug discovery. Similarly, I came across Amarda Shehu’s essay, From Blue Sky to Ground Truth: Predictive AI for Life.

7. I enjoyed reading Beyond the Endless Frontier: Rebuilding Corporate Research for a Stronger American Future by Jeffrey Tsao. I had the opportunity to meet and spend some time speaking with Jeff at the Progress Conference. Jeff makes the case that the scientific method is incomplete, that industrial research labs like Bell Labs, Xerox PARC, and GE once created unique environments where real-world problems sparked foundational discoveries. Without such labs today, this crucial method of discovery, what Donald Stokes called Pasteur’s Quadrant research, is largely missing.

8. The Institute for Progress recently relaunched their Macroscience Blog. “We are relaunching Macroscience with a broader focus and community of writers, including the scientists and policy entrepreneurs shaping the future of American science. Macroscience will explore ideas for how to improve science and policy, and it’ll feature writers who challenge assumptions and start friendly arguments.”

9. I enjoyed Owl Posting’s podcast episode with Ellen Zhong, a computer science professor at Princeton working on machine learning for cryo-EM structure prediction. What stood out to me most was how grounded her answers were in biological reality. One thing I wanted to add, today’s high-end electron microscopes cost 10-15 million dollars.

10. Lots of insights in John Collison’s podcast with Eli Lilly’s CEO Dave Ricks. I’m amazed at the fact that Eli Lilly’s R&D budget is one-third of the entire NIH budget! This was also the first time I had heard in a well articulated way why clinical trials are so expensive.

Now, onto some scientific papers I’ve been reading:

For some context: in my day job, I study cellular mechanisms of neurodegeneration in Alzheimer’s and Parkinson’s disease. In both conditions, misfolded proteins (tau in Alzheimer’s and alpha-synuclein in Parkinson’s) are defining pathological features. Traditionally, the structures of these proteins are studied by extracting and purifying them from brain tissue, then applying techniques like nuclear magnetic resonance or cryo-electron microscopy to determine their atomic structure.

But to truly understand how these proteins damage cells, we need to see them in their native context, which is embedded inside real neurons, interacting with membranes, organelles, and other proteins. Cryo-electron tomography (cryo-ET) is currently the only method capable of imaging molecular structures in intact tissue at sub-nanometer resolution. Setting this up in human tissue is extraordinarily difficult, and it’s what I’m building toward in our lab.

Here are a few papers that have influenced my workflow:

11. In this paper, René Frank’s lab at the University of Leeds investigates the cellular architecture surrounding amyloid-beta and tau proteins in Alzheimer’s disease in what I think is an impressively ambitious technical undertaking. The team begins with flash-frozen blocks of human brain tissue, which are thawed to room temperature and kept in preservative media to prevent structural degradation. The tissue is sectioned into thin slices (~100 micrometers) and fluorescently labeled to highlight amyloid deposits. These labeled sections are then high-pressure frozen to lock native cellular structures in place.

    Next, the frozen sections are imaged using fluorescence microscopy to identify amyloid-rich regions. This correlative step is essential without which locating pathology under an electron microscope would be a needle-in-a-haystack problem. Once regions of interest are identified, the samples must be thinned further. Because electrons can only pass through extremely thin specimens, the ~100-micrometer sections are trimmed down to ~70 nanometers using a diamond knife, over 1,000 times thinner than the original tissue slice. Finally, the ultrathin sections are placed in a transmission electron microscope, where images are acquired across multiple tilt angles and computationally reconstructed into three-dimensional volumes. From these reconstructions, individual tau filaments are digitally segmented and thousands of copies are aligned and averaged to reveal their molecular structure within the native cellular environment.

    This study shows that amyloid-beta plaques and tau filaments in Alzheimer’s disease have distinct, spatially organized structures, and that tau filaments adopt location-specific conformations within native tissue. This study links molecular structure directly to cellular context and pathology.

12. There are two main ways to flash-freeze (vitrify) biological samples for cryo-EM and cryo-ET. The first, and more common method, is plunge freezing, where a sample is rapidly dropped into liquid ethane cooled to about -180C. Because the cooling happens so quickly, water in the sample does not have time to form ice crystals and instead freezes into amorphous (glassy) ice. This matters because crystalline ice is disastrous for structural biology: ice crystals expand and physically tear apart delicate cellular structures. The main limitation of plunge freezing is sample thickness, where only very thin samples, typically less than ~10 µm, cool fast enough throughout to avoid crystal formation. Thicker samples freeze too slowly in their interior, leading to crystalline ice and structural damage.

    The second method, high-pressure freezing, is designed for thicker samples. Here, the specimen is rapidly cooled to liquid nitrogen temperatures (~-180 °C) while simultaneously being pressurized. Under high pressure, water is prevented from expanding as it freezes, which strongly suppresses crystal formation even in much thicker tissue. The downside is that high-pressure freezers are expensive, complex, and far less widely available than plunge freezers, putting them out of reach for many labs.

    These practical constraints have pushed researchers to innovate around hardware limitations. In this paper, Yi-Wei Chang’s group at the University of Pennsylvania demonstrated that even a 100 µm-thick brain section can be plunge frozen and later processed for cryo-ET by thinning it down with a plasma focused ion beam. Using a xenon plasma beam, they were able to mill away most of the frozen tissue and produce electron-transparent lamellae just a few hundred nanometers thick. This approach effectively reopens plunge freezing as an option for much thicker biological samples by shifting the burden from freezing to post-processing.

13. One of the major challenges with plunge-freezing thick brain tissue is that the cold does not penetrate the interior fast enough to prevent ice crystals from forming. Instead of vitrifying into amorphous ice, water in the center of the sample partially crystallizes, damaging fine cellular and molecular structures. A common workaround is to soak the sample in cryoprotectants, which are chemicals that slow ice nucleation and crystal growth. But not all cryoprotectants are equal: some preserve structure at the cost of massively increasing osmolarity, while others risk chemically perturbing proteins or membranes. In this paper, Weier et al. systematically benchmark different cryoprotectant cocktails to identify mixtures that can help vitrify brain tissue up to ~100 µm thick by plunge freezing alone, without destroying physiological structure.

I have many more papers I could share, but these summaries ended up much longer than I intended, so I’ll stop here and pick up again next month.

Finally, I’ll be attending Foresight Institute’s Vision Weekend in SF this upcoming weekend.

Happy Thanksgiving!

I’ve been thinking back recently to my PhD prospectus, which I completed years ago, and how it became the moment when I first felt myself shifting into the mindset of a real scientist. On the surface, the prospectus (or proposal) is just an hour-long presentation to your thesis committee explaining what you hope to accomplish for your dissertation. But it was the first time I had to articulate, in a deep and defensible way, what actually motivated my research. What gap in knowledge was I trying to fill? Why would any of it matter? At the time, it felt like another requirement everyone had to get through. Only later did I understand how much the structure of the prospectus quietly shapes the way you learn to think, at least it did for me.

For some context, I did my PhD in Biomedical Engineering at Boston University. In most departments across most universities, PhD students are required to complete a prospectus. It usually includes a written proposal, often in the NIH grant style with specific aims and a research strategy, and an accompanying oral presentation outlining the work you plan to do for your dissertation. These are submitted to your thesis committee, which includes your advisor and four other professors.

Different departments have different timelines. BU’s BME program required us to complete the prospectus by the end of our third year. In other fields, like computer science, students didn’t need to do it until six months before graduation. In disciplines where research depends on physical experiments, like biology or engineering, departments tend to require it earlier to ensure students have a clear plan and enough time to execute. In more theoretical fields, the experiments may be computational and short enough that a later prospectus still works.

Structurally, the proposal is often divided into three aims that map onto a larger scientific question. You introduce the big problem you are trying to solve, then describe three aims that collectively get you closer to answering it. For each aim, you lay out the experiments or analyses you plan to run, why those are the right approaches, and what results you expect to see.

Of course, I had talked through these questions many times with my advisor, and I could explain my work in broad strokes to friends and family. But until the prospectus, I had always implicitly trusted that my advisor understood the bigger picture and that I was working on something worthwhile. The prospectus forced me to answer a harder question: Did I believe in and understand the research strongly enough to convince five experts that they should believe in it, too? And you can’t convince anyone of anything until you’ve convinced yourself.

Connecting the dots from what I did every day at the lab bench to how that might create meaningful knowledge to push the field forward, and maybe even help the world in a concrete way, was transformative. It made me think ambitiously. I suddenly saw a whole arc of research far beyond what I could realistically do in the three years I had left. But I wanted to do it anyway. And so, I proposed the entire thing.

What the project was isn’t nearly as important as what happened next. I gave my presentation, it went smoothly, my committee congratulated me, and then promptly told me to cut half of it. One member even joked that I’d outlined the work of two, maybe three, PhDs. The exact work isn’t important because every one of my peers got the same feedback. It seems to be almost universal: PhD students propose wildly ambitious projects, and thesis committees guide them toward something more realistic.

Why do students over-propose? One reason is that students fear coming across as unproductive so end up proposing more to look good. I don’t find that very compelling as the primary driver. I think most students want to do the work they outline, or at least think they want to. A more convincing explanation is simply being naive about how long it takes to do research, especially in the biomedical sciences. The planning fallacy is deep-rooted in science, from graduate students to PIs. But PhD students might be especially vulnerable to this because they just haven’t failed enough yet to know how long their work might take.

But I think the deeper reason is that, at the prospectus stage most of us simply haven’t yet learned how to think deeply about a scientific problem. I don’t mean that we don’t know what’s important, or that we can’t outline the logical steps of a project. I mean something closer to understanding what it really takes to know the truth in science and to get to the bottom of a natural phenomenon. I’m still learning this myself, seven years after my own prospectus. It’s a lifelong process.

And that’s why we propose broad, ambitious ideas: they feel exciting. Broad claims feel consequential. Narrowing a question, cutting away good ideas for a single sharp, answerable problem, requires a kind of discipline and experience you just don’t have yet. The prospectus is, in many ways, the first real lesson in that. In the traditional academic path, it’s also the first formal moment when the scientific community gives you feedback on how you think as a scientist.

This implies that the prospectus serves a very important role in the development of a scientist. And that’s what makes it a powerful but underrated teaching tool. The committee’s job is not just to keep you from overworking yourself. It’s to teach you how to frame a problem, how to identify the core question, how to think from multiple angles rather than chase multiple directions. It’s one of the rare moments in graduate school where you are forced to move from “I’m doing experiments” to “I’m building a coherent research program.” I wish I could go back and appreciate this, instead of treating it like another box to check on my way to adding Dr. to my name.

At the same time, I sometimes wonder about the message embedded in the ritual. When every student is told some version of “you’re doing too much,” does that gently nudge people toward safer, smaller science? Does the instinct to protect students from failure also unintentionally train them to think less ambitiously?

Ambition and feasibility are both essential scientific virtues. The trouble is that graduate training tends to over-index on feasibility, because feasibility keeps students sane and ensures they graduate. But the scientific enterprise thrives on the opposite: people who are willing to try things that don’t fit neatly into a 3-5-year timelines. If the earliest training moments consistently reward the safest version of the idea, then young scientists may absorb the wrong lesson, not that ideas must be rigorous, but that ambition is something to be contained. We need training structures that can teach focus without extinguishing the spark that leads people to ask the biggest questions.

Doing a PhD is not the only path to developing this kind of depth. Any sustained effort, paired with real feedback, can reshape how you think. Looking back, I’m grateful for the prospectus because it forced me to see what it actually means to pursue a question with depth. Beneath all the logistics and forms that made me a “PhD candidate,” there was also a shift from doing the work to understanding it.

Ever since the invention of the microscope, scientists have wanted to expand the depths of what they could see into plant and animal life. Biology spans a vast spatial spectrum from angstroms to centimeters. On the large end we have whole organisms, which themselves vary in size from single-celled bacteria to large mammals. At the small end we find proteins and individual amino acids that build up these protein chains. In between we have everything from viruses to sub-cellular organelles. From Antonie van Leeuwenhoek and Santiago Ramon y Cajal to Ed Boyden and Richard Henderson, we have traveled the spatial scale of millimeters, micrometers, nanometers, angstroms, and are now closing in on the sub-angstrom world.

Subscribe now

In my previous piece, Zooming into biological structure, I wrote about how scientists have built a mosaic of microscopy techniques to study biological structure. I also recently published a piece with Asimov Press, Making the Electron Microscope, which talks about the early history of the electron microscope and a detailed description of how it works. This current essay, which is meant to be a standalone piece, focuses on the more recent history of electron microscopy. I outline how an amalgamation of sample preparation techniques, computational advances, and electron detector technology all came together to revolutionize structural biology, yet again.

Today, cryo-EM allows us to resolve biological structures at the scale of individual atoms. In 2020, a group of scientists at the MRC Laboratory of Molecular Biology in Cambridge, UK, resolved the structure of a neuronal membrane protein at atomic resolution. In 2024, a group at the University of Leeds determined the structure of tau protein and amyloid-β within postmortem tissue from Alzheimer’s disease brains to near-atomic resolution. Efforts are now underway to use these three-dimensional structures to design drugs that target specific domains within the structure.

Progress in science often depends not on one breakthrough, but on the slow alignment of many. In the case of electron microscopy, the revolution came only after decades of parallel advances in chemistry, physics, and computation finally converged.

For the purpose of this essay, I will only briefly describe how the electron microscope works. For full detail I refer you to the Asimov Press piece. At its core, the electron microscope works much like a light microscope, but instead of photons it uses a beam of electrons to probe the sample. At the top of its tall metal column sits the electron gun that emits electrons accelerated to nearly the speed of light. Magnetic lenses made of coiled wire focus and steer this beam, just as glass lenses bend light. When the electrons pass through a thin sample, they scatter according to the arrangement of atoms inside, carrying with them information about the sample’s structure. These scattered electrons are magnified through a series of lenses and projected onto a detector, forming an image. Because electron wavelengths are thousands of times shorter than visible light, the microscope can resolve features far smaller than anything seen with optical microscopy, down to individual atoms under the right conditions.

If you freeze water fast enough, it will turn glass-like

While electron microscopy promised theoretical resolutions at the scale of atomic bonds, in practice this proved extremely difficult to achieve. For decades, it remained a niche tool with several major limitations. One of the biggest challenges was the water present in biological samples, a problem that’s hard to avoid, since most of biology is made of water. As I described in my Asimov piece:

“For instance, from the start, electron microscopy for biology faced a water problem. Because the microscope operates under high vacuum, liquid water evaporates instantly, leaving delicate biological samples collapsed or distorted. To avoid this, aqueous samples had to be dried, fixed, or stained, which produced recognizable images but with obvious artifacts, such as shrunken cells, ruptured membranes, and structural distortions that no longer reflected the living state. Through the 1940s and 1950s, embedding samples in resins and the use of ultra-thin sectioning made cellular ultrastructure visible, while freeze-drying and early cryogenic sectioning offered partial preservation of hydrated material, though the results were still plagued by distortion.”

Another major limitation was radiation damage. The high-energy electron beam that revealed such fine detail also destroyed the very sample it was imaging, burning through delicate biological material within seconds. Scientists realized that cooling the sample below freezing reduced molecular motion, which slowed heat buildup and therefore limited radiation damage. The problem, however, was water itself: when it freezes, it expands and forms crystalline ice, which distorts biological structures.

The breakthrough came with the realization that if you freeze water fast enough, the molecules don’t have time to form a crystal lattice. Instead, they solidify into an amorphous, glass-like state in a process called vitrification. This is exactly what Jacques Dubochet and his colleagues achieved in the 1980s at the European Molecular Biology Laboratory. They demonstrated that water could be rapidly cooled into vitreous ice, thereby preserving biological samples in a lifelike, hydrated form, without the destructive crystallization of ice.

Today, samples are typically vitrified by plunging them into liquid ethane chilled by liquid nitrogen to around -183 °C. In an instant, the thin film of water around the molecules solidifies into clear, glass-like ice. This technique not only reduces radiation damage but also preserves biomolecules almost exactly as they exist in life. They are frozen in motion, yet structurally intact.

If you average enough particles, you get three-dimensional models

Another major bottleneck for electron microscopy was the lack of three-dimensional information. For that, X-ray crystallography remained the gold standard. Since the 1930s, X-ray crystallography had enabled atomic-resolution structures of biological molecules, but it came with a key requirement: the ability to produce high-quality crystals. These crystals diffracted X-rays at specific angles to reveal the spacing of atoms within the crystal. However, many large macromolecules, such as viruses, ribosomes, or membrane proteins, could not be crystallized, leaving major regions of biology structurally unresolved.

In contrast, electron microscopy, despite its much higher theoretical resolution, produced only two-dimensional projection images of samples. For decades, electron microscopy was primarily a visualization tool for large complexes and cellular structures, but reconstructing accurate 3D models from these projections remained a formidable challenge. Without the regular, repeating patterns that crystals provided, early scientists lacked the computational methods to extract reliable 3D information.

That changed when Joachim Frank and his colleagues developed statistical and computational techniques to solve this problem. As I wrote in my Asimov piece:

“In parallel, computational techniques were improving. Beginning in the 1970s, Joachim Frank developed statistical methods for aligning and averaging thousands of noisy electron micrographs of individual macromolecules. This “single-particle analysis” transformed faint, low-contrast images into coherent 3D reconstructions. When combined with Dubochet’s vitrification method, the two advances gave rise to single-particle cryo-electron microscopy: Molecules suspended in vitreous ice could be imaged in random orientations and computationally combined into detailed three-dimensional structures.”

To unpack this a bit more: a single vitrified sample of purified protein may contain millions of identical macromolecules, such as ribosomes, each embedded in a thin layer of vitreous ice. Before freezing, these molecules move freely in solution. Upon vitrification, they are trapped in place, ideally in random orientations1, producing a field of particles viewed from many angles. Frank’s methods used statistical alignment and averaging to identify the orientation of each particle and combine these thousands of 2D projections into a coherent 3D reconstruction.

In the 1970s and 1980s, these images were recorded on photographic film that was later digitized for analysis. By the 1990s, CCD cameras began to replace film, streamlining data collection, though they still offered limited resolution. With these early tools, single-particle reconstructions typically reached resolutions of about 10 to 20 angstroms, enough to reveal the overall shape of large complexes like ribosomes, but not the positions of individual amino acid side chains.

While vitrification and computational methods had laid the foundation, the achievable resolution was still constrained by the quality of the recorded images. To break through that final barrier, a third component needed to fall into place.

If you detect electrons directly, you get a revolution

The stagnation in resolution was now due to the efficiency of the camera detectors. CCD and CMOS sensors were originally designed to detect light, not electrons. Light, in the form of photons, is converted into electrical charge and read by the sensor. Electrons, however, are not light. To overcome this limitation, scientists added a scintillator in front of the detector, which is a layer that converts incoming electrons into photons, which are then detected by the camera. But when electrons strike the scintillator, the emitted photons scatter in all directions, blurring the image and reducing resolution.

While digital cameras allowed faster feedback and easier image analysis, their scattering effect meant that photographic film still outperformed digital acquisition in terms of resolution. By the early 2000s, it was clear to nearly everyone in the field that this indirect detection method was fundamentally inefficient. The next leap in resolution would require detecting electrons directly.

Traditional CCD and CMOS cameras weren’t built for this task. They were designed for photons, which interact only with the surface of a sensor. High-energy electrons, by contrast, penetrate deeply into the silicon, creating trails of electrical charge that spread out through the material. In these older chips, the regions that collect the signal and the regions that read it out overlapped. Because the charge was generated deep inside the silicon, much of it was lost before reaching the collection layer. Even the charge that did make it to the surface often spread across several neighboring pixels, a phenomenon known as charge spreading, which blurred fine details and limited resolution.

The solution came with a new detector architecture called monolithic active pixel sensors (MAPS). These sensors separated the layer where the signal is generated from the layer that reads it out. They used a thin, lightly doped epitaxial layer2 near the surface of the silicon to collect charge efficiently without interference from the deeper circuitry. This design sharply reduced signal loss and charge spreading, allowing each electron to be detected cleanly and precisely. MAPS were first developed at the Rutherford Appleton Laboratory in the UK for particle physics experiments and published in 2000.

Around this time, the electron microscopy community began experimenting with a variety of new detectors to push performance further. In 2002, Richard Henderson and colleagues demonstrated the use of a hybrid pixel detector for electron microscopy, which separated the sensing and electronics layers but suffered from large pixel sizes that limited resolution. Then, in 2005, at least two independent groups (including Henderson’s) published promising results using MAPS for electron microscopy. These sensors could detect single electrons with excellent signal-to-noise ratios, but they still required engineering refinements before being integrated into commercial instruments.

By 2010, MAPS-based direct electron detectors (DEDs) had been optimized for stable, sustained electron microscopy use and incorporated into commercial microscopes. What made these detectors revolutionary was both their sensitivity to electrons and their speed. Instead of recording a single image on film, scientists could record whole movies, capturing many frames per second. This not only allowed more averaging to get sharper images, it also allowed them to correct for beam-induced motion, which are tiny shifts and vibrations caused by the electron beam that blur every image.

Just three years later, in 2013, the first generation of high-performance DEDs became widely available. For the first time, cryo-EM could capture movies of electrons hitting the detector and reconstruct macromolecules at near-atomic resolution. What had been a decades-long march of incremental improvements suddenly became a resolution revolution, reshaping structural biology and opening a new era of molecular discovery.

A new age in electron microscopy

Imagine a world where everything around you is just slightly out of focus. You can see the trees, but the leaves blur together: shapes without edges, outlines without detail. If you wear glasses, you know the moment that changes, the instant you put them on and the world snaps into focus. I still remember the first thing I said when I tried mine for the first time: “Oh my god, I can see every single leaf on that tree!”

That’s what the resolution revolution of the early 2010s felt like for electron microscopy. After decades of “blobology,” where proteins appeared only as vague, featureless shapes, they could suddenly see the contours of individual amino acids in exquisite detail. But these scientists weren’t looking at trees, they were looking at the machinery of life itself. And I imagine many of them had the same reaction: “Oh my god, I can see every single amino acid in that protein!”

This leap in clarity didn’t come from a single breakthrough. What I love about the story of cryo-EM is that it needed multiple independent lines of long-arc research directions to mature and align. The whole is truly greater than the sum of its parts. Dubochet’s vitrification preserved biological molecules in their native, hydrated state, allowing them to be imaged without distortions. The direct electron detectors were advances on camera technology that captured those fine details with minimal blurring. And Frank’s computational methods leveraged statistics to build clear images from noisy, low-contrast individual micrographs.

Cryo-EM needed all three because its power rests on a delicate sequence: the molecule must first be preserved in a lifelike state, then imaged faithfully, and finally reconstructed computationally. If any one of these steps fails, such as the ice too thick, the camera too inefficient, or the algorithms too crude, the final structure collapses into blur. In that sense, the resolution revolution wasn’t a single eureka moment but the culmination of decades of parallel progress in chemistry, computation, and instrumentation.

A fair question to ask is: did it really need to take that long to achieve atomic resolution? Frank was already developing single-particle averaging methods in the 1970s, and Dubochet’s vitrification techniques appeared in the early 1980s. Yet the so-called “resolution revolution” only arrived more than thirty years later.

Part of the answer lies in the detectors. In the 1980s and 1990s, digital imaging itself was still in its infancy. CCD and CMOS cameras were only beginning to replace photographic film, and while they offered convenience, they couldn’t yet match film’s resolution or sensitivity. The problem wasn’t in the lack of ideas, virtually everyone in the field knew that new detectors were needed. The problem was in the actual development of the hardware. Electron detectors had to evolve dramatically before they could capture images with the precision needed for atomic resolution.

But this delay was also not just technological developments, it was translational too. Even once new detectors like MAPS were demonstrated in physics labs in the early 2000s, it took another decade of engineering and iteration before they could be made robust, affordable, and stable enough for daily use in commercial electron microscopes. Moving a technology from a prototype on a lab bench to an instrument that thousands of scientists can rely on is slow, exacting work. It takes the persistence, and often invisible labor of scientists like Henderson to work alongside engineers to bridge the gap between discovery and adoption.

Looking ahead

This does not conclude my stories on electron microscopy!

For decades, most structures deposited in the PDB came from X-ray crystallography. That balance is now shifting. Over the past ten years, the number of structures solved by cryo-EM has risen sharply, while X-ray crystallography is starting to plateau. If current trends continue, 2025 may mark a symbolic crossover point, where cryo-EM and X-ray crystallography reach equal footing in the number of structures deposited this year in the PDB, and from there, cryo-EM might begin to dominate.

This shift has been driven in large part by the democratization of cryo-EM. The growth of national and regional centers that make cutting-edge microscopes and sample preparation accessible to scientists who might never have had direct access before. These shared facilities have lowered barriers to enter the field of cryo-EM and opened the door for an entirely new generation of structural biologists.

However, sample preparation remains a big, manual, bottleneck. Sample preparation, which requires plunge freezing the specimen into liquid ethane and then manually transferring it to liquid nitrogen, is a skill to be mastered. While it is certainly more automated today than when Dubochet first developed the method, it continues to be as much an art as a science. Producing the perfect thin, vitrified ice, free of contamination or crystal formation, takes remarkable skill.

If we want a future where AI systems will unlock new scientific discoveries, we need large, comprehensive, and consistent datasets for it to learn from. A bigger question may be: what would it take to make the cryo-EM pipeline so consistent and reproducible that an AI could navigate it end-to-end? This question deserves a post of its own.

And we haven’t yet touched the next phase of electron microscopy: cryo-electron tomography (cryo-ET), where scientists tilt samples inside the microscope to reconstruct entire cells and tissues in 3D, not just individual molecules. Cryo-ET is extending the reach of structural biology from purified proteins to the molecular architecture of life itself, captured in its native environment. With it, scientists are beginning to visualize how viruses assemble and mature, how misfolded proteins trigger neurodegeneration, and how organelles organize the crowded interior of the cell. It hasn’t yet reached atomic resolution, but it’s only a matter of time, and sustained effort by motivated individuals, before it does.

Ultimately, the structure of biology is valuable only insofar as it reveals function, and function rarely resides in isolated molecules. Cryo-EM and cryo-ET are bringing us closer to seeing how those molecules work together, interact, and organize into the living systems that make up our world.

My own background is in biomedical engineering and neuroscience, so much of my thinking is geared toward biological research. I have worked in academic labs and a non-profit research institute, and now I have transitioned back into the university lab. That mix of settings has made me think more deeply about what I value in research, and about how research environments shape the science that gets done. I outline some of these ideas here at a high level, and hope to explore each more fully through future writings.

These aren’t fully formed proposals. They’re more like ideas that have been bouncing around in my head, shaped by my experience across different research structures, and by reading people in the metascience space like Michael Nielsen, Ben Reinhardt, Eric Gilliam, Adam Marblestone, Jason Crawford, and Matt Clancy, to name a few. I outline three ideas below: the first suggests shifting how we hire scientists, away from narrow skills and toward broader talent; the second explores how different types of research goals require different organizational structure; and the third argues that institutional requirements in science funding are holding back small-scale research startups.

Integrating PhD and medical rotations into professional research to accelerate breakthroughs

During my PhD at Boston University, the first year was all about lab rotations. You tried three different labs, one each in the fall, spring, and summer, before committing to one lab. While learning new technical skills was part of the process, it was more about seeing how different groups worked, experiencing the culture, and asking yourself if you could spend five years there. You got the chance to make an informed choice, rather than being “assigned” to a professor from day one. That autonomy matters. Looking back, rotations are underappreciated.

The idea actually has roots in early industrial research. At GE’s Research Lab in the early 1900s, new hires were left to wander the floor for a few days to figure out what interested them. Irving Langmuir described how the director of the GE lab, Willis Whitney, would walk around asking, “Are you still having fun?” The ethos of picking a project that excites you and trusting that productivity will follow is powerful. GE still uses a version of rotations today for junior employees, moving them through various departments before settling into a permanent role. GE also has one of the higher employee retention rates, but of course, causal links here are fuzzy.

Another example of rotations comes from medicine. Medical students cycle through various specialties, like pediatrics, internal medicine, surgery, and others, before applying to a residency program. You see different patient populations, learn how teams operate in different clinical settings, and get a sense of what type of work aligns with your temperament and interests. In the end, you make a more informed choice about your specialty.

What if we professionalized this model more broadly, not just for PhD and medical students, but across research organizations? Instead of hiring for narrow technical skills, you hire for scientific talent. Right now, the system for hiring scientists in industry is largely skills-based. But that seems backwards. One of the most important skills in becoming an independent scientist isn’t how fast you can pipette liquids from one tube to another or how many animals you can run through a behavior rig. Those are consequences of good training, not the cause of scientific ability. The real skill is learning how to choose, evaluate, and solve meaningful problems.

By framing job opportunities more broadly, you naturally shift the emphasis from narrow technical skills to individual talent. You bring in great people, give them a structured way to explore, and let them choose what they’re most drawn to. I think that would create healthier environments and, maybe, bigger breakthroughs. Irving Langmuir went on to win the Nobel Prize in Chemistry for work that began as a curiosity-driven side project in the GE Industrial Labs.

Subscribe now

Curiosity-driven research and platform science need different organizational structures

Another distinction I’ve been chewing on is curiosity-driven research versus platform science. Curiosity-driven research is what most people picture when they think of academic science: chasing novelty, asking “how does this mechanism work?”, or developing new technology prototypes. It thrives on nimbleness and freedom.

Platform science (I also considered calling it research at scale) is different. It’s still fundamental science, but it comes after that first spark of curiosity. Expansion microscopy is a good example of something that began as a curiosity and evolved into a platform. Originating in Ed Boyden’s lab, the goal was to map neuronal connections in the mouse brain. Their strategy: Instead of developing new microscopes with higher resolution to see the connections of individual neurons, let’s just see if we can find a way to make neurons bigger and use our existing microscopes. Inspired by the remarkable absorbing and swelling properties of a baby diaper, the basic question they asked: can we take brain tissue that has been removed from the animal and make it bigger? This wild idea worked and has been adopted by many labs around the world.

Now, at E11 Bio, a Focused Research Organization (FRO), their goal is to scale up expansion microscopy to map the connections of the entire brain. FROs are (typically) non-profits that tackle large-scale, well-defined, and pre-defined scientific challenges to produce robust platforms, tools, or datasets that the broader research community can build upon. The emphasis is not on novelty for its own sake, but on making methods systematic, reliable, and scalable so they can power further discovery.

I love FROs for this. They’ve carved out a space to scale basic science, and they’re producing incredible resources. But I also worry they won’t fully satisfy the best curiosity-driven scientists. Platform projects require standardization, engineering pipelines, and bigger teams, which can be thrilling if you’re motivated by the scale itself, but perhaps a little unsatisfying if what excites you is the chase for new ideas and exploratory work.

I observed this push-pull between curiosity and platform during my time at the Allen Institute. Founded in 2003, the institute began as essentially an FRO (before the term FRO existed) to build the Allen Brain Atlas, an open-source map of gene expression and cellular anatomy of the brain. Over time, it recruited more curiosity-driven scientists who were either already in academia or would have otherwise gone into academia, and the institute attempted to do both large-scale data generation and exploratory science. But the needs of those two modes are different. Platforms need rigid engineering and stable pipelines. Curiosity needs flexibility, even a little chaos. Putting them together in one place often means neither side is being efficient in scientific progress. But we need both.

So maybe the solution is not to blend them, but to separate them more cleanly. Support organizations built for scale and organizations built for curiosity, rather than trying to make hybrids that stretch themselves thin. Then, we can build clear paths to transition an idea out of curiosity-mode and into scale-mode. What does an academic lab look like if you took it out of the university? Small-scale curiosity-driven research teams of 15-20 people without any specific profit or end goal in mind. Ben Reinhardt’s essay Unbundling the University is an excellent in-depth essay on how separating research from universities could enable small, independent teams optimized purely for discovery. I hope to unpack his essay in a future post.

Institutional affiliation in grant funding is a limiting factor for research “startups” in the life sciences

One of the reasons I suspect that spinning out small-scale research teams is challenging is the institutional requirement for government grant funding. If you’ve ever applied for NIH support, you’ll know the section: you have to detail your institution’s resources (an example I found from Mount Sinai). What equipment is available? What facilities does your university or institute maintain? How cutting-edge are the tools in your environment? The grant application doesn’t just evaluate you or your idea; it evaluates your research environment and institution itself.

For a capital-heavy, curiosity-driven research team, this is stifling. To secure funding, you need an institutional affiliation, but to set up an “institute” (here I only mean it as a small group of scientists), you need funding. This is where most individuals might pivot to a startup and secure VC funding if the idea can be spun to make a profit, or obtain an institute affiliation and start a lab in academia. Is there a third option? Perhaps obtain sufficient philanthropic seed money to establish the basics of an “institute” and then apply for government funding.

I’m not exactly sure when this requirement became so entrenched, but it also weighs heavily on early-career scientists. In biomedical research, landing funding can give you an edge in the academic job market, especially through transition-to-independence awards like the K99/R00. These awards typically cover two years of postdoc work followed by three years as an independent PI. But they also assume you already have an institutional anchor for your postdoc work.

If you don’t have a transition award, there are other early career awards available, but they all also require institutional affiliation. To secure a good position, it helps to come with funding. But to secure funding, you need institutional affiliation. When I looked for grants that would let me propose a project without an institutional home, the options were vanishingly few.

This setup narrows the pipeline. If you’re in a postdoc but want to pivot fields, or your PI isn’t supportive of transition awards, or you simply lose out in an extremely competitive cycle, your chances of securing an independent faculty position at a top research program is lower in the biomedical sciences. At this point, many great scientists pivot to industry positions. Further, the institutional requirement locks you into your current environment and discourages bold shifts in research direction. And what if you don’t want to be in a university setting, with teaching loads and administrative duties, but still want to pursue independent not-for-profit research? Under the current system, there are few viable paths forward.

I wonder if there’s a better way. Perhaps we set up communal science spaces with shared equipment where independent teams could operate. Suppose early-career scientists could apply for funding as individuals, with contingencies built in rather than prior requirements. You’d only receive the money if you secure a position, but you’d have the guarantee in hand while on the job market. That guarantee could enable riskier proposals, career pivots, or independent ideas that don’t map neatly onto your postdoc work.

Bonus: What does a laboratory look like in the age of AI?

This idea is much fuzzier and more of a fantasy right now.

A lot of AI-for-science projects are focused on literature review, hypothesis generation, or computational modeling. Those are very useful. But most scientists working at the bench spend very little of their time reading papers. The bottleneck is the lab itself: building microscopes, running animal experiments, prepping samples, growing and keeping cells alive. Most of my workday is there, not in the library. While LLMs have significantly improved my pace of learning through scoping the literature, it has only had a small effect on speeding up my experimental workflows. And that’s exactly the part of science that is least documented and least automated.

There are glimpses of change. FutureHouse is working on AI hypothesis generation and execution, Cultivarium is experimenting with smart glasses for documenting lab work, companies like Ginkgo and Medra are pushing automation for lab equipment. And Reinhardt points out that “AI scientists” won’t get far if most of science’s real work is invisible in undocumented benchwork. But all of these still assume the human-centered lab as the template.

So what if we designed labs for machines, not humans? Today’s labs, benches, pipettes, and optical tables are built for human bodies. But in a future where experiments could be autonomous, the lab might look completely different. Instead of humans conducting animal experiments, imagine animal housing systems equipped with microscopes that require minimal human intervention (some efforts have been made in this direction). Or, imagine modular wet lab rigs that reconfigure themselves for new protocols. That way, no human would ever have to set foot into a cell culture room. With this, we might finally experiment with biology at its own speed, rather than at our convenience.

Closing

In Michael Nielsen and Kanjun Qui’s essay “A Vision for Metascience,” they outline the process and culture of how we do science as an exploratory space, a playground for trial and error in building new social structures for research. What we want is not a single “right” structure but a thriving ecosystem, one that can rapidly generate and iterate on many ideas about how science could be organized. They say in their introduction:

“We argue that: (1) metascience is an imaginative design practice, exploring an enormous design space for social processes; (2) that exploration aims to find new social processes which unlock latent potential for discovery; (3) decentralized change must be possible, so outsiders with superior ideas can’t be blocked by established power centers; (4) ideally, change would align with what is best for science and for humanity, not merely what is fashionable, politically popular, or media-friendly; (5) the net result would be a far more structurally diverse set of environments for doing science; and (6) this would enable crucial types of work difficult or impossible within existing environments.”

In this sense, it’s about structural diversity. Distinct organizations for curiosity-driven and platform science could reduce institutional friction and provide each type of endeavor what it really needs. Funding models that decouple individuals from institutions might promote decentralization and lower barriers for early-career scientists. Creating an environment that makes rotations feasible might unlock talent in unexpected ways.

As Nielsen and Qui say, “Monoculture is the enemy of creative work.” The way forward is to multiply the environments in which science can happen, create an iterative feedback loop where we learn from our mistakes, and preserve the “free play of free intellects” that has always driven the most important breakthroughs.

If anyone is interested in chatting with me about these topics, please reach out!

This post is a little different. I am fortunate to be part of this year’s Roots of Progress Blog-Building Fellowship. The last few essays, and the next few to come, are products of the fellowship. This one grew out of an assignment to write a personal essay. Enjoy!

I love working with my hands. At work, that means building microscopes from scratch or manipulating rice-sized biological samples in liquid nitrogen. At home, it means carving intricate designs in wood.

You don’t need much to start woodworking. You need some space, you need YouTube, a few basic tools, and a bit of wood.

When I first began woodworking, we lived in a one-bedroom apartment and our office doubled as our “workshop”. For the really dusty jobs, we moved into the bathroom so sawdust wouldn’t sneak into our bookshelves. A few years later we upgraded to a home with a garage, and I was thrilled to discover the built-in pegboards, until it was pointed out that most American garages probably have pegboards.

The next thing you need is a teacher, and YouTube is full of them. One of my favorites is Paul Sellers, a hand-tool woodworker who has been teaching for decades. His calm, precise, British lecturing style is a joy to watch on its own. He even built cabinets that now live in the White House.

In terms of tools, if you were to, say, start with making a basic cutting board, all you need is a saw, a spokeshave, sandpaper, and some finishing oil. That’s it. If you want to get fancy, by adding a handle, you will also need a coping saw and a rasp.

Then there’s the wood itself. The decision to make is softwood or hardwood. Softwood cuts more easily, but hardwood looks far better when finished. Think pine versus mahogany.

I’m not going to write a tutorial here. For that, I’ll point you to Paul’s videos. Nothing I can say would beat just watching him work.

When we moved into a place with a garage, the first thing we did was build a workbench. There was something deeply satisfying about this: building the tool that allows you to create. This mirrors strongly with how I approach my science: building my instruments that allow me to answer questions in new ways.

Other projects followed: bar stools, coasters, spatulas. I stayed in the hand-tool world until I came across Neurowoodworks on X, posting wooden neurons. I was hooked instantly. After much back and forth, I was finally making neurons of my own.

At first, I thought a bandsaw would do the job, and we already had one at home. I learned the hard way that it would not. A bandsaw is like the highway: it’s made for long, straight cuts and gentle curves. It’s fast, powerful, and great if you need to rip a piece of wood down the middle. But try to cut the tight twists of a neuron with one, and it might feel like trying to parallel park a semi-truck (I can only imagine). The blade just doesn’t turn tightly enough, and before you know it, you’ve either broken your piece or veered way off your line.

The scroll saw, on the other hand, is like having a fine-point pen. Its thin blade moves up and down, so you can swivel your wood in place and steer through hairpin curves. That way you get the delicate branches of a dendrite without shredding your work. Not long after I found a used scroll saw on Facebook Marketplace.

When I am working with the scroll saw, nothing else matters. The sound of the motor pulsing the blade, the soft vibrations of the wood beneath my hands, the smell of freshly cut lumber. The minutes certainly slip into hours. The pull of the wood along the blade makes it feel less like I’m making something but more like being carried along with it.

Once I got through a few neurons, I started venturing into more creative projects. My most intricate piece yet is inspired by the Mandelbrot set. I hope to turn it into a bedside lamp for my daughter’s bedroom.

For me, woodworking is another outlet for creativity, learning, focus, and making something real and tangible. There is a deep satisfaction in seeing something take form under your fingertips, whether this is building microscopes, or typing code, or carving wood, a feedback loop between thought and action.

A timelapse of building our workbench:

Since then, I have been trying to learn more about the Alzheimer’s diagnostic space. This essay is part of that process. I am not an expert on Alzheimer’s research. I have worked for years at the intersection of neuroscience and microscopy, but not directly on Alzheimer’s. So this is me, sorting through what I’ve read, what I think I understand, and what still confuses me.

Subscribe now

Alzheimer’s affects roughly 40 million people worldwide, and there is no cure. With an aging population, it has never been more crucial to fight Alzheimer’s disease. Several drugs have been tested, but they are modest at best in slowing cognitive decline. One possible reason is that by the time clinical symptoms appear, which is when most patients are diagnosed, the disease biology is already far advanced. Detecting it earlier could give therapies a much better chance of making a difference.

An early and accurate diagnosis matters for more than just treatment. It shapes care plans, determines eligibility for clinical trials (often where the most promising therapies are available), helps rule out reversible causes of cognitive decline, and gives patients and families clarity as they navigate a frightening and confusing situation.

The problem is that our current diagnostic tools are either invasive (like spinal taps) or expensive (like PET scans). In practice, they are usually performed only after symptoms are obvious. That is why blood-based biomarkers are so appealing: they promise a cheap, simple test that could be used earlier and more broadly. And that is why the FDA’s approval in May 2025 felt like such big news.

But as I started reading more, I realized that flashy headlines can be misleading. So I wanted to understand: what exactly makes Alzheimer’s so hard to treat? How is it currently diagnosed? Why are existing tests inaccessible at early stages? And does this new blood test really bring us closer to early detection?

Alzheimer’s begins ~15 years before clinical symptoms appear

Pathologically, Alzheimer’s disease is defined by the accumulation of amyloid-beta plaques outside neurons and tau neurofibrillary tangles inside neurons, leading to widespread neurodegeneration (more on these below). Clinically, it presents as a progressive decline in memory, thinking, and daily functioning, beginning subtly and advancing to severe dementia.

While there are many hypotheses for what causes Alzheimer’s disease, the truth is we still don’t fully know. It’s not for lack of trying. We have tried. A lot. But the biology of how Alzheimer’s begins and unfolds is extremely complex, with feedback loops that play out over decades. With every paper I read, the disease only seems more complex, not less. What we do have a better handle on is how the disease progresses biologically, even if the exact causal links remain unclear.

One of the earliest changes we see is the buildup of a small protein fragment called amyloid beta (Aβ). In particular, the Aβ42 variant tends to deposit in the spaces outside neurons, forming plaques. Aβ is a byproduct of normal healthy processing, when the amyloid precursor protein (APP) is cut. APP is normally produced in healthy brains and may play roles in antimicrobial defense, neuronal signaling, and even plugging leaks in the blood-brain barrier after injury. The problem arises when Aβ is overproduced or not cleared efficiently. In that case, it begins to accumulate into plaques that gradually spread. Over the course of 15 to 20 years, this pathology can extend across much of the brain.

Even so, amyloid on its own turns out not to be a good predictor of cognitive decline. Some people have widespread Aβ plaques and show only subtle, even undetectable, deficits on neurological exams. What does line up more closely with neurodegeneration and symptoms is tau.

Tau is another naturally occurring protein in the brain, responsible for stabilizing the microtubules that form the cell’s internal scaffolding and transport system. But tau can go wrong. When it becomes abnormally folded, it can “seed” nearby tau proteins to misfold as well, creating tangles. These tangles disrupt neurons and trigger neurodegeneration. Memory loss is usually the first noticeable clinical sign of Alzheimer’s because tau pathology begins in the brain’s memory centers. Once it starts, tau can spread rapidly along neural circuits, driving degeneration wherever it goes.

Crucially, once tau pathology is underway, it no longer seems to depend on amyloid. This has potentially big implications for treatment: if you clear amyloid after tau tangles have already formed, you may not stop the damage. Tau pathology, neurodegeneration, and cognitive decline tend to track together in both time and space. That suggests two possible therapeutic windows. One is to target tau directly before neurodegeneration has set in, though this window may be very narrow. The other is to intervene earlier, by clearing amyloid before tau pathology begins.

Current treatments have been modest at best

The first drug to reach patients was aducanumab (Aduhelm). In June 2021, the FDA granted it accelerated approval on the basis that it cleared amyloid plaques from the brain. But the clinical data were inconsistent: one phase 3 trial suggested a modest slowing of decline, while a second found no benefit. The approval was deeply controversial, with several members of the FDA advisory panel voting against it or resigning in protest. By January 2024, Biogen withdrew the drug from the U.S. market.

The next drug to hit the market was lecanemab (Leqembi). It earned accelerated approval in January 2023 and full traditional approval that July, after a trial showed ~27% slowing of cognitive and functional decline over 18 months. For the first time, there was solid evidence that clearing amyloid could translate into meaningful clinical benefit, even if the effect was modest. A year later, in July 2024, the FDA approved donanemab (Kisunla). In the study, donanemab slowed decline by ~29–35%, with the largest benefits seen in people treated earlier in the disease course.

None of these drugs are cures, but they have shifted the field from no disease-modifying therapies at all to the first tangible, though modest, slowdowns in decline. The limited gains highlight two key issues. Timing: by the time symptoms appear, amyloid has been accumulating for over a decade, tau pathology is often entrenched, and neurodegeneration is already in motion. Starting treatment then may be too late to rescue neurons already lost. And target: most late-stage programs so far have aimed squarely at amyloid. Tau-directed antibodies are now advancing, with phase 3 trials ongoing, and it may be that targeting tau, or using combination therapies, will yield greater benefit.

Current diagnostics are used too late in disease progression

With Alzheimer’s, by the time symptoms become obvious, the disease has often been silently progressing for years and sometimes decades. Therefore, early detection will lead to a better prognosis. That’s where today’s practice collides with biology. Our healthcare system as it stands is built around reacting to problems rather than preventing them.

The Alzheimer’s Association Workgroup, composed of experts from the National Institute on Aging and the Alzheimer’s Association, has established guidelines to help clinicians diagnose Alzheimer’s. These guidelines emphasize shifting from a purely symptom-based approach to a biological definition of the disease. In this framework, Alzheimer’s is understood as a process that begins with measurable neuropathological changes, such as the accumulation of amyloid deposits, long before clinical symptoms appear. At present, however, presymptomatic testing is generally limited to individuals with a known genetic predisposition that puts them at higher risk and only in research settings.

Most people enter the system because of symptoms: a spouse or adult child notices that they repeat questions, don’t pay bills, or take wrong turns on a familiar drive. The clinician takes a careful history, screens cognition, checks mood and sleep, orders bloodwork to catch reversible contributors like thyroid disease or B-12 deficiency, and obtains structural imaging, such as MRI or CT, mainly to rule out other causes like stroke, tumors, or other injuries. These are appropriate first steps, but they are indirect for Alzheimer’s. They say little about amyloid and tau.

The most direct measure of Alzheimer’s is through Positron Emission Tomography. In PET imaging, a radioactive label is injected into the patient. These tracers are chemical compounds that specifically lodge themselves into Aβ sheets or tau tangles. By taking images with a camera, the physician can see the extent of plaque formation or tau spreading. There are currently three FDA-approved radiotracers for Aβ and one for tau. Most people have some low levels of plaques and tangles, but Alzheimer’s patients will show high levels of widespread distribution of these pathologies.

If PET imaging can be performed early, it has the potential to catch Aβ accumulation before the spread of tau pathology and neurodegeneration. However, PET scans, while accurate and sensitive, are expensive due to the cost of radiotracers. So, PET scans are only ordered when the physician is almost certain that the patient has Alzheimer’s.

A more recent form of testing is to perform a spinal tap. The fluid that surrounds the brain and spinal cord is called the cerebrospinal fluid (CSF). Small amounts of Aβ and tau naturally make their way into the CSF. By measuring these levels and comparing them to established healthy patterns, clinicians can detect the presence of Alzheimer’s-related changes with high accuracy. To get more specific, there are two major forms of amyloid-β: Aβ42 and Aβ40, both produced in the brain. Plaques deposited in the brain mostly consist of Aβ42. So, when plaques begin to form, the concentration of Aβ42 in the CSF falls, while Aβ40 remains relatively unchanged. Evidence from studies shows that the ratio of Aβ42/Aβ40 is consistently lower in Alzheimer’s patients than in healthy controls.

Similarly, tau pathologies can also be measured in the CSF. Tau is a normal protein that helps stabilize the internal scaffolding of neurons, but when it becomes abnormally phosphorylated (called p-tau) it tends to misfold and form tangles. Both total tau and p-tau circulate at low levels in the CSF, and their concentrations rise as tangles accumulate in the brain. Large studies and meta-analyses show that combining Aβ42/Aβ40 with p-tau makes CSF testing highly accurate, with over 90% sensitivity and specificity when compared against PET imaging or autopsy confirmation. In contrast, measuring Aβ alone achieves only around 80% sensitivity and specificity, meaning it produces more false positives and false negatives than the combined approach.

As with the PET scan, spinal taps are only performed when the patient has clear cognitive decline and other causes are ruled out. Performing the spinal tap requires a highly trained clinician and inserting a needle into the spinal canal. While any serious risks with the procedure are very low, people probably don’t want to do this unless necessary.

In real-world U.S. care, the accessibility of PET scans and CSF tests is still low. The rates vary between speciality clinics and primary care, but some data suggest a rate of 16% for speciality clinics and 10% in primary care settings. About 60-70% of patients receive structural imaging with CT or MRI. In short, the biological tests that could move diagnosis earlier and sharpen decisions were the least used.

We need a simple and accurate early diagnostic for Alzheimer’s

Blood tests, in contrast, are easy and common. That is why blood-based biomarkers have attracted so much attention. Small amounts of brain-derived Aβ and tau cross the blood-brain barrier, and with ultrasensitive assays, we can measure them. In May 2025, the FDA approved the first blood test to aid in diagnosing Alzheimer’s disease in adults 55 and older who are already being evaluated for cognitive impairment (which is a good start, but already late in the disease stage).

So how do these markers behave? On average, plasma Aβ42/40 is lower in people with amyloid plaques in the brain, but the change is modest, typically a 10–15% drop, compared to healthy controls. Notably, this 10-15% figure is in the advanced stage, and could be much lower in the early stages. Plasma p-tau levels, by contrast, show a much more robust increase with Alzheimer’s pathology and correlate strongly with both amyloid PET and tau PET scans. In large research cohorts, plasma p-tau can approach the diagnostic accuracy of CSF testing, often exceeding 90% sensitivity and specificity when combined with Aβ42/40.

But p-tau is a late signal. Phosphorylated tau reflects the stage when tangles are forming and spreading, which tends to occur closer to the onset of symptoms. If the goal is to diagnose Alzheimer’s in its earliest, pre-symptomatic phase, p-tau may not yet be elevated. In that scenario, we are left with Aβ alone, and Aβ by itself is not a strong predictor of disease progression. It flags the presence of pathology but not its severity, and its sensitivity and specificity are lower (around 80%) with more false positives and negatives compared to combined testing.

This means the current FDA-approved blood test is most powerful in people who already have signs of Alzheimer’s, not those in the silent phase we most want to reach. It is progress to have a blood test available at all, but the biology of the markers limits how early they can truly diagnose the disease. What we really need are biomarkers that change in lockstep with the very first shifts in Alzheimer’s biology, long before tau tangles and overt symptoms appear. Until then, blood tests will remain more of a tool for confirmation than for genuine early detection.

Where does this leave us?

Having accurate biomarkers for early diagnosis of Alzheimer’s, ones that are also simple and accessible, feels crucial. We are not there yet. In practice, only a small fraction of patients who present with memory concerns receive anything beyond neurological exams and basic imaging. PET scans and spinal taps remain limited by cost, availability, or invasiveness. Blood tests are promising, but they are still in their infancy, and the markers we have today are not perfect reflections of the earliest disease stages.

Why does this matter so much? Because diagnosis is not just about putting a label on a condition. An early and accurate diagnosis shapes care plans, allows patients and families to plan, and, perhaps most importantly, opens the door to intervening while the disease biology is still malleable. Alzheimer’s begins 15-20 years before symptoms, which means there is a long silent window where intervention could have the biggest impact. Without early diagnosis, we are always playing catch-up.

Of course, diagnostics are only one side of the equation. A test without an effective treatment is limited in value. The drugs we have so far offer only modest slowing of decline, and are used only after symptoms appear. Still, they show that disease modification is possible, and pairing them with earlier diagnosis may extend their benefit. That’s the hope of ongoing presymptomatic trials: to see if starting treatment before tau tangles and neurodegeneration set in can truly change the trajectory.

And then there’s the question of trust. The Alzheimer’s field has not been free of controversy. The high-profile fraud around a particular amyloid variant (Aβ*56) cast a long shadow and added to the public confusion. Fraud is bad, and ideally, we don’t want to be wasting resources and money. But it is also important to separate one fraudulent line of research from the thousands of independent studies that stand on solid ground. From what I understand, the overwhelming body of evidence on Aβ42, tau, and the biomarker cascade remains intact.

So, where does this leave us? We have a clearer map of how Alzheimer’s unfolds. We have our first disease-modifying therapies, even if modest. We have new blood tests that make biological diagnosis more accessible than before. None of these pieces is enough on its own. But together, they move us away from the old model of diagnosing late and treating symptomatically, and toward a future where Alzheimer’s can be detected earlier, targeted biologically, and perhaps one day, truly prevented.

Two leading figures stood on opposite sides of this question. Santiago Ramón y Cajal, a neuroanatomist and artist, drew what he saw under the microscope and argued that neurons were discrete units. Camillo Golgi, whose staining method made such drawings possible, believed the nervous system was one seamless network. In 1906, they shared the Nobel Prize in physiology, despite holding opposing views. The disagreement was like trying to decide whether a chain-link fence is made of separate pieces or one continuous sheet when you are only allowed a blurry photograph.

The connection point between neurons, what we now call the synapse, was invisible with the microscopes of their time. To Cajal and Golgi, the brain tissue under the microscope looked like a mesh, and the idea of a gap between cells remained purely theoretical. It would take another fifty years and major advances in imaging before scientists could directly see the synapse. Resolving this debate and embracing the concept of the synapse paved the way for understanding brain signaling and developing treatments for depression, epilepsy, and Parkinson’s disease.

Studying biology has always depended on our ability to see structure in finer and finer detail. Life spans an enormous range of scales, from nanometers to centimeters. At the large end are whole organisms, from single-celled bacteria to complex mammals. At the small end are proteins and even the individual amino acids that make them up. In between lie viruses, organelles (the structures within cells), and the molecular machinery that keeps cells running and allows them to communicate with one another.

Understanding biological structure in high-resolution detail allows us to better predict function, design therapeutics, and develop vaccines for diseases. In 1953, X-ray crystallography revealed the structure of DNA’s double helix, laying the foundation for molecular genetics and modern biotechnology. In 1969, Dorothy Hodgkin’s crystal structure of the insulin molecule not only showed how insulin is stored in the body, but also guided the rational design of insulin analogs that are still central to diabetes treatment today. More recently, cryo-electron microscopy has become an essential tool in drug discovery and informing the development of improved antibody-based therapies.

There is no one-size-fits-all instrument that can image the large and the small. Instead, the field has evolved a multitude of approaches to study structure at different scales. We have traveled the spatial scale of millimeters, micrometers, nanometers, angstroms, and are now knocking on the door of the sub-angstrom world. Getting to this point took centuries of work.

Subscribe now

Light Microscopy

From the 17th to the 19th century, light microscopy was biology’s only window into the fine structures of life. The earliest compound microscopes, like Galileo’s in the early 1600s, offered just 30 times magnification, which was enough to see insect legs, hairs, and the outlines of large plant cells, but still blind to red blood cells or bacteria.

That leap came from Antonie van Leeuwenhoek, a Dutch microbiologist and microscopist who turned microscopy into an art form. By the late 1600s, his meticulously crafted single-lens microscopes reached 300 times magnification with resolutions down to one micrometer. For the first time, blood cells, bacteria, and protozoa came into view.

But this microscopic world existed only in memory and sketches of its observers. For centuries, scientists had no way to capture what they saw beyond hand drawings. It was only in the late 1800s, after the invention of photography, that scientists first began experimenting with photographing microscopic observations.

Meanwhile, van Leeuwenhoek’s single-lens microscopes had a problem: a tiny field of view. Early compound microscopes with two or more lenses promised to fix this, but the added optics created severe color aberration, which created a rainbow halo around every specimen because different wavelengths bent at different angles. In the mid-1700s, lens makers solved this by combining different glass types into the achromatic doublet, which canceled out the color fringing.

By the early 1800s, achromatic lenses gave compound microscopes the resolution of van Leeuwenhoek’s instruments but with wider fields of view. Scientists could now image cells within whole tissues, like plant stems, animal organs, the kinds of specimens familiar to any high school biology student today. This clarity fueled the cell theory in 1839, establishing that all living organisms are made of cells and that the cell is life’s fundamental unit.

Then came the precision era. In the late 1800s, Carl Zeiss and Ernst Abbe transformed microscopy from artisanal tinkering into optical engineering. Using mathematical optics, they designed lenses systematically, pushing resolution to about 0.25 micrometers. Now, cell membranes, nuclei, mitochondria, and bacterial shapes all came into sharp focus.

By this time, the light microscope was reaching its limits. In 1873, Abbe mathematically proved that no matter how perfect the lenses, light microscopes could not resolve details smaller than ~200 nanometers, which is about 250 times thinner than a human hair. Much of biology, including synapses, viruses, and proteins, lies hidden below this line. The next leaps in imaging would require breaking the limits of light itself.

Into the Invisible

In 1895, X-rays were discovered. As physicists worked to understand their nature, they found that X-rays behaved like waves, and that crystals acted like 3D diffraction gratings when X-rays passed through them. In 1913, William and Lawrence Bragg, the father–son duo, realized they could use these diffraction patterns to deduce crystal structures mathematically, inventing X-ray crystallography. In X-ray crystallography, X-rays scatter through a crystal to produce a pattern of spots on a screen. Analyzing those patterns reveals the arrangement of atoms inside the crystal down to 0.3 nanometers in the best cases.

Soon, this powerful technique was applied to biology. In 1934, Dorothy Hodgkin and J.D. Bernal demonstrated that protein crystals, if carefully prepared, could also yield X-ray diffraction patterns, leading to the birth of structural biology. The next few decades delivered landmark discoveries: the double helix of DNA in 1953, the first protein structures (myoglobin and hemoglobin) in the late 1950s. Biology moved from describing life’s parts to understanding them as 3D machines whose shapes explained their functions.

But X-ray crystallography had limits. It required crystals, ruling out most large proteins, viruses, and cellular assemblies. And because no lens could bend X-rays, scientists only saw diffraction patterns, never direct images. The vast middle ground between small crystallizable molecules and whole cells imaged by light microscopy remained invisible.

The electron microscope filled that gap. Invented by Ernst Ruska in 1931, it used electron beams, with wavelengths far shorter than light, to achieve much higher resolution. By the late 1930s, researchers could see entire viruses like tobacco mosaic virus and vaccinia for the first time, proving these “germs” were real physical particles and validating the germ theory of disease. Early electron micrographs also revealed mitochondria, nuclei, and the internal architecture of cells’ structures that had been guessed at but never seen.

Over the next decades, electron microscopy kept advancing. New staining and sample preparation methods in the 1950s uncovered membranes and organelles in unprecedented detail. The 1960s brought scanning electron microscopy, producing dramatic 3D-like views of cell surfaces. Then came the revolution of cryogenic electron microscopy (cryo-EM) in the 1980s and 1990s, where samples were rapidly frozen, preserving them close to their native state. Today, cryo-EM can reach near-atomic resolution for proteins, viruses, and organelles, finally linking the atomic details revealed by X-ray crystallography to the cellular landscapes seen with light microscopy.

From Images to Insights

Over the past four centuries, biology has advanced hand-in-hand with the instruments used to visualize it. From the early single-lens microscopes that revealed blood cells and bacteria, to X-ray diffraction patterns that uncovered the atomic scaffolding of DNA and proteins, and then to electron microscopes that brought viruses and cell ultrastructure into view, each leap in imaging has opened a new layer of the biological world. Together, these approaches form a mosaic that spans scales from nanometers to centimeters, connecting the structure of molecules to the organization of tissues.

This is far from an exhaustive list. Even as electron microscopy and X-ray crystallography transformed biology, light microscopy never stood still. Techniques like confocal microscopy, light-sheet microscopy, super-resolution methods, and expansion microscopy have pushed past earlier limits to image cells in ever greater detail. Other approaches, such as nuclear magnetic resonance (NMR) and single-molecule spectroscopy, opened additional windows onto molecular structure and dynamics. Rather than replacing one another, these techniques have filled complementary niches, each expanding what could be visualized in space or time.

This essay has focused on static structure, which is the ability to freeze life at different scales and examine it in detail. An equally important frontier has been dynamic imaging: light microscopes adapted to look inside living cells, fluorescent proteins that allow individual molecules to be tracked in real time, and super-resolution methods that extend beyond the diffraction limit of light. Structural and dynamic approaches together provide a fuller picture of biology. Our ability to see into biology is inseparable from our ability to understand, and each improvement in technology reshapes the questions that we could ask about biology.

I first became interested in Irving Langmuir not through his discoveries, but through his citation history. While analyzing the long-term impact of Nobel Prize-winning research, I noticed that Langmuir’s papers had unusual trajectories: they remained lightly cited for most of the 20th century, then experienced a dramatic surge in citations starting in the early 2000s. That kind of delayed recognition, sometimes called a “Sleeping Beauty,” raised some interesting questions. What were these papers about? Why do they matter so much now? And who was this industrial chemist whose ideas were suddenly central to environmental science? This essay is the result of that curiosity: a look into Langmuir’s life, his science, and how a century-old theory found new purpose in a modern world.

Irving Langmuir won the Nobel Prize in Chemistry in 1932, but his most cited paper titled, “The adsorption of gases on plane surfaces of glass, mica and platinum,” saw the majority of its scientific impact nearly a century after publication. Over the past two decades, this once-overlooked work has become foundational in environmental science, particularly in the design of modern water purification systems. Why did a core theory in surface chemistry lie dormant for so long before rising to prominence?

Ironically, Langmuir's own passion for environmental conservation predated this scientific influence. In 1910, long before climate change or clean water were central global concerns, he became captivated by Lake George in upstate New York. Then a young chemist at the General Electric Research Laboratory, he spent his spare time canoeing between the lake’s many islands, hauling rocks to protect their shorelines from erosion. A nearby dam caused the rise and fall of water levels in the lake beyond its natural depth, leading to erosion of the islands. This was the start of Langmuir’s decades-long effort to preserve the lake’s health.

Now, nearly a century after he began his first riprapping efforts on Lake George, Langmuir’s research into the molecular interactions at surfaces is at the heart of efforts to decontaminate polluted water around the world. The Langmuir Isotherm, a model describing how gas molecules adhere (or adsorb) to solid surfaces, has become a key tool used in designing systems that remove pollutants from drinking water.

In a previous post on citation trajectories of Nobel Prize-winning work, I highlighted Langmuir as a striking case: recognized early as groundbreaking work, yet broadly adopted only many decades later. As I dug deeper into his life and the reasons behind this delayed impact, I discovered a scientist with an unusually diverse and dynamic career. Langmuir not only founded the field of surface chemistry; he also helped invent the modern incandescent light bulb, pioneered studies of ionized gas (he coined the term “plasma”), discovered atomic hydrogen, and developed hydrogen welding. Alongside colleagues at GE, he even launched large-scale experiments in cloud seeding to make it rain.

So, how does Nobel Prize-winning science lie dormant for 70 years before being broadly adopted by a field? This essay explores the enduring impact of Irving Langmuir’s work, the story of a scientist who followed his curiosity across disciplines, and the forces that brought a century-old theory to life.

The impact of Irving Langmuir’s Nobel Prize-winning work

When I began tracking the citation counts of Nobel Prize-winning papers over time, I noticed a striking pattern: several older papers, especially those published before 1950, remained lightly cited for decades before experiencing a sudden surge in attention. These papers, sometimes called the “Sleeping Beauties” of science, are works that remain in obscurity before waking to widespread influence.

Irving Langmuir’s prize-winning papers are among the most extreme cases. In terms of delayed recognition, they stand out even among Nobel Laureates. The chart below shows Langmuir’s papers (highlighted in red) in the context of all Nobel-winning works between 1902 and 2016 (in gray). One in particular, “The Adsorption of Gases on Plane Surfaces of Glass, Mica and Platinum,” is now cited more than 21,000 times.

The case of the Nobel Prize “Sleeping Beauties” is especially revealing. These are papers that were recognized as groundbreaking at the time, yet they lingered with little attention for decades before exploding in influence. I was curious to see what caused the surge in citation counts for Langmuir’s papers.

To understand this surge in relevance, I used the OpenAlex research database to examine the fields and topics of the papers citing Langmuir’s most famous work, which is his 1918 paper on adsorption. That single paper has been cited by 21,370 papers to date. OpenAlex metadata assigns a “primary topic” to each paper, drawn from a classification system spanning over 4,500 topics1. This made it possible to trace where Langmuir’s influence is now most visible.

The results are striking. Below are two charts that show the top 20 subfields and top 20 topics associated with the papers that cite Langmuir’s 1918 paper. As the charts below show, over 43% of citing papers fall into the subfield of Water Science and Technology. Within that group, a staggering 95% are specifically focused on Adsorption and Biosorption for Pollutant Removal. Langmuir’s surface chemistry model has become a cornerstone of water decontamination research. The primary topics are color coded according to the subfield they fall into.

So, what makes the fundamental study of molecular interactions at surfaces so critical for removing pollutants from water? Why did it take 80 years for applications of Langmuir’s work to surge? And how did a man best known for inventing the gas-filled incandescent lightbulb and discovering atomic hydrogen end up winning the Nobel Prize for something entirely different?

His research intensity was akin to “the Knights of the Round Table in their quest for the sacred vessel”

Irving Langmuir was born in Brooklyn, New York, on January 31, 1881. He earned a bachelor’s degree in metallurgical engineering from Columbia University’s School of Mines in 1903, then traveled to Germany, then the epicenter of physical chemistry, to pursue a PhD at the University of Göttingen. There, he studied under Walther Nernst, of the Nernst equation from thermodynamics and a future Nobel Laureate himself.

Under Nernst’s guidance, Langmuir explored the behavior of gases around heated filaments. His early experiments focused on the formation of nitric oxide near a glowing Nernst filament—a ceramic rod used in early electric lamps. Nernst had hypothesized that the glow resulted from a chemical equilibrium between oxygen and nitrogen in the surrounding air. Although that idea proved incorrect, it set Langmuir on the path to eventually develop a much-improved and longer-lasting light bulb.

His thesis eventually turned to the dissociation of carbon dioxide near a platinum filament. This work proved unexpectedly significant. Langmuir discovered that in the thin boundary layer of gas (just a few tenths of a millimeter thick) thermal conduction, not convection, dominated heat transfer. This insight challenged prevailing assumptions and had immediate implications for vacuum technology and filament design.

Langmuir’s doctoral work laid a foundation for his later contributions to surface chemistry and plasma physics, where chemical interactions at the boundaries are important. Just as importantly, it revealed his scientific style: starting from first principles, designing simple but revealing experiments, and building theory from observation.

He completed his PhD in 1906 and returned to the U.S. to teach chemistry at the Stevens Institute of Technology. His three years there were the most frustrating of his career. As one of only three faculty members in the new department, Langmuir was overwhelmed with teaching, discouraged by uninspired students, and had little time or support for research.

In 1909, through mutual contacts, he was invited to spend the summer at the General Electric Research Laboratory. What began as a short visit became a lifelong appointment. As recounted in Langmuir: The Man and the Scientist, he pursued his research at GE “with the intensity ascribed to the Knights of the Round Table in their quest for the sacred vessel.”

The General Electric Research Laboratory: the academic freedom Langmuir yearned

The GE research laboratory was born out of GE’s strong desire to outcompete its competitors in developing a better light bulb. Light bulb technology was at its infancy: they flickered, burned out quickly, caused blackening of the bulb’s interior, and consumed large amounts of current. Despite these flaws, demand for electric lighting was rising rapidly, and GE recognized that whoever cracked the code on a durable, efficient, affordable light bulb would dominate the market.

Concerned that incremental engineering alone wouldn’t yield major breakthroughs, GE’s leadership envisioned something radical: a corporate lab focused not on production, but on scientific discovery. As one executive put it, they wanted a lab that would “search out scientific principles and act as a source of technical information for the company.” With that mandate, the GE Research Laboratory was founded in 1901 in a barn in Schenectady, New York. Free from commercial pressures, it became a place where physicists and chemists could pursue fundamental problems, guided not by directives, but by curiosity. The unofficial motto, under the lab’s director Willis Whitney, became a simple question: Are you still having fun?

This was the environment Langmuir entered in the summer of 1909. Like all new recruits, he spent his first days wandering the lab, looking for something that piqued his interest. He returned to a familiar challenge: lamps and filaments. GE had already developed a tungsten-filament bulb that improved on previous designs, but a persistent issue remained: residual gases inside the bulb caused blackening of the glass. Most believed the solution was to improve the vacuum seal. But Langmuir, skeptical of that assumption, approached the problem differently.

He found that GE’s vacuum was already far superior to any he had used during his PhD work, and instead turned to fundamental questions: How do different gases behave when exposed to a heated tungsten filament? What reactions occur inside the bulb at high temperatures? For three years, he ran meticulous experiments, chasing many dead ends but learning something new each time.

Eventually, Langmuir discovered the key: at high temperatures, the inner surface of light bulb glass itself released water vapor, which then reacted with the tungsten filament. This caused tungsten atoms to evaporate, travel through the bulb, and deposit onto the glass, resulting in the blackening. The problem wasn’t the vacuum; it was the temperature-dependent chemistry of the materials inside the bulb.

In parallel, Langmuir tested how different gases behaved under heat. He found that hydrogen dissociated into atoms when heated (discovering atomic hydrogen), while nitrogen remained stable. Introducing nitrogen into the bulb dramatically reduced tungsten evaporation and therefore blackening. These experiments led to the invention of the gas-filled incandescent lamp, a major leap forward in lighting technology.

But the implications of that work reached far beyond lamps. Along the way, Langmuir discovered atomic hydrogen, laying the groundwork for hydrogen welding. He also began investigating how these reactive hydrogen atoms adhered to glass surfaces, and noticed that they formed a single atomic layer. This insight sparked his interest in adsorption and led to his work in surface chemistry, for which he would eventually win the Nobel Prize.

Langmuir credited the freedom he had at the GE lab for making these discoveries possible. After three years on the tungsten problem, the lab’s director, Willis Whitney, asked his usual question: “Are you having fun?” Langmuir replied, “I’m having a lot of fun, but I don’t know what good this is to the General Electric Company.” Whitney simply responded, “That’s not your worry, it’s mine.”

Unlike academia or other industrial labs of the time, Langmuir wasn’t assigned a research program. Whitney provided the equipment, the staff, and, most importantly, the freedom to follow his own questions. GE, in return, reaped enormous benefits, not just in patents and products, but in foundational science. Langmuir came to believe that industry had a responsibility to support basic scientific research, “regardless of any payoff in profits, because of the debt which all industry owed to science for its very existence.”

Langmuir’s curiosity about hydrogen atoms adhering to the inside of the glass bulb soon expanded into a broader question: how do molecules behave at surfaces? That question would take him far beyond light bulbs and into the foundation of the new scientific field of surface chemistry.

Adsorption: the flatland of the chemical world

Adsorption is the accumulation of atoms or molecules from a gas or liquid onto the surface of a solid or liquid. It is a surface-specific phenomenon, unlike absorption, in which substances penetrate into the bulk of a material. In adsorption, the adsorbate clings to the adsorbent’s surface, forming a layer that is typically only one molecule thick.

Interest in surface phenomena dates back to the 18th century. Benjamin Franklin experimented with how oil spread across water to study wakes left by ships. Michael Faraday explored the distinct physical forces at play at surfaces, separate from those inside the bulk of materials. In the late 19th century, Lord Rayleigh, building on Franklin’s work, used oil films to demonstrate that molecules have a definite size and shape. Willard Gibbs then introduced a thermodynamic framework, the Gibbs adsorption isotherm, to describe how concentration at a surface changes with chemical potential.

Still, by the early 20th century, adsorption remained poorly understood and largely based on empirical evidence. Most scientists believed it to be a diffuse, gradual, multilayer process, akin to the way Earth retains its atmosphere. Only Rayleigh had proposed that adsorption might involve a single molecular layer, but his idea lacked theoretical support.

Langmuir’s experiments with hydrogen in light bulbs changed that. He observed that hydrogen atoms adhered to the inner glass surface in a single layer. Influenced by emerging ideas in X-ray crystallography and molecular bonding, he wasn’t surprised by the regularity. Across a variety of gases and surfaces, he consistently found the same: atoms packed tightly into a stable monolayer that required significant energy to dislodge.

This led him to propose a new idea: adsorption is a chemical process, not merely a physical one. Instead of weak physical forces or fields extending from the surface, Langmuir argued that adsorbed atoms form actual chemical bonds with the surface itself. This view contrasted sharply with the prevailing physics-based models of his time. As he later wrote:

“The older views on adsorption are now often referred to as constituting a physical theory of adsorption, whereas the newer ideas are described as a chemical theory. I think this distinction involves not so much the nature of the forces involved as the traditional attitude of mind of physicists and chemists. The physicist, probably ever since the time of Newton, has been inclined to consider forces which vary as some power of the distance between the bodies acted upon by the forces. He has also had many occasions to consider fields of force which extend throughout space, such as gravitational, electric, and magnetic fields. The chemist, on the other hand, has nearly always considered that chemical action between molecules or atoms takes place only when they are in contact. It is a natural tendency for the physicist in considering the problem of adsorption to think of the behavior of the adsorbed atoms or molecules as being in a field of force produced by a solid or liquid surface. It would be more in accord with the tradition of the chemist, however, to think of an atom on the surface as being attached to one or more atoms of the underlying surface.”

Langmuir pictured adsorption through chemical bonds and reasoned that because bonds between two different materials, i.e. at the surface, will be different from the bonds of the second layer, which are between the atoms of the same substance and not at the interface, adsorption is monomolecular.

This insight launched his exploration of monolayers, including films on water surfaces. Through careful experimentation, he showed that a surface can hold only a fixed quantity of atoms before reaching saturation. From this, he derived the Langmuir Isotherm, a mathematical relationship between the pressure of a gas and the quantity adsorbed on a surface.

Langmuir extended his studies across many surfaces and compounds, even exploring the behavior of proteins and enzymes. His deep understanding of molecular interactions at surfaces led to the “Principle of Independent Surface Action,” allowing him to predict chemical behavior based solely on molecular structure.

Much like the fictional inhabitants of Flatland, who lived in a world of two dimensions, Langmuir immersed himself in the chemistry of surfaces: thin, invisible planes where entirely new rules applied. His pioneering work opened up a new field of surface chemistry and in 1932, he was awarded the Nobel Prize in Chemistry for his contributions.

The impact of Irving Langmuir’s Nobel Prize-winning work

While Langmuir essentially began the field of surface chemistry, this part of his work, although foundational, only had a modest impact in the few decades following its publication. Surface science was still an esoteric and messy field. That began to change with the rise of the environmental sciences and the urgent need for better tools to address pollution and resource management. Langmuir’s isotherm, once a theoretical model grounded in first principles, found renewed relevance as a practical tool.

An isotherm is a mathematical relationship describing how two variables, such as pressure and surface coverage, relate at constant temperature. Langmuir’s version, now known as the Langmuir isotherm, models how molecules adsorb to a solid surface in a single layer. It assumes a finite number of identical adsorption sites with no interaction between molecules. The equation is simple but powerful: θ = (K P) / (1 + K P), where θ is the fraction of occupied sites, P is the pressure, and K is the equilibrium constant. This was the first quantitative theory of surface adsorption, and it remains the most widely used.

The Langmuir isotherm is critically important for applications of molecular separation and storage technologies because it provides a quantitative framework for understanding how molecules selectively bind to surfaces. This principle is central to processes like gas purification, water filtration, and chemical separation. In systems such as activated carbon filters, the Langmuir model helps determine how much of a contaminant can be captured at a given concentration. For gas storage, such as hydrogen or carbon capture, the isotherm informs how efficiently porous materials can hold gases under specific conditions. Its simplicity and theoretical grounding make it ideal for optimizing material design, surface area, and process conditions.

Today, modern materials with intricate porous nanostructures, such as metal-organic frameworks and graphene-based composites, dominate storage and separation systems since they have significantly more efficient adsorption properties. Yet Langmuir’s model, developed for flat surfaces, remains the standard. It is often preferred over more empirical alternatives like the Freundlich isotherm, which lacks a clear molecular interpretation and cannot easily predict adsorption behavior.

The Langmuir isotherm has guided the development of zeolites, crystalline aluminosilicates used in water purification and petrochemical catalysis; magnetite-graphene composites, which remove heavy metals like arsenic and lead from drinking water; and even pharmaceutical formulations, where it helps determine dose-response dynamics of drug binding.

Langmuir’s adsorption model is simple, grounded in theory, capable of predicting adsorption properties of materials, and applicable to a wide variety of materials and surfaces. Langmuir certainly had an eye for applications throughout his research career, and he is most well-known for some of the applications his research yielded, such as the gas-filled light bulb and hydrogen welding. In a 1939 film, he demonstrated how surface chemistry could be used to understand proteins and biological membranes. Yet even he might not have imagined that his work would become central to delivering clean water around the world.

The man who once hauled rocks by canoe to protect a single lake now, a century later, provides a conceptual foundation for purifying water on a global scale.

A recent post by Brian Potter on Construction Physics took a deep-dive into the people behind the prizes in chemistry, physics, and medicine. This excellent piece examined questions like: Where were Nobel laureates born? Where were they educated? Where did they conduct their groundbreaking work? And when did they receive recognition?

I loved the focus on people. Understanding the individuals behind scientific progress is deeply underrated. We might learn far more about how to foster breakthrough discoveries by studying the lives and trajectories of those who produce them. That said, Brian’s post also got me thinking about the nature of the work itself. Are there patterns in how impactful a discovery is before it wins the Nobel? Does the prize boost its influence? Could we predict Nobel-worthy work and accelerate its impact?

One measure of scientific impact is citation counts, which is the number of times a paper is referenced by other papers. Citation counts are appealing because they provide a simple, quantitative measure of influence within the scientific community. But they also come with serious limitations. This paper argues that the obsession with citations has pushed science toward safe, incremental advances rather than bold, exploratory work. As the authors write, “our main concern with citations is that they reduce scientific progress to a set of numbers that capture just one important dimension of scientific productivity,” and “citations are an imperfect and noisy way to measure scientific influence, and yet, citations are a useful measure of scientific influence.”

While they go on to suggest alternative metrics, such as quantifying novelty, citation counts remain our best available tool for now. With those caveats in mind, let’s explore what they can reveal about the impact and trajectory of Nobel-winning work.

Gathering data on citation counts

Brian’s dataset on GitHub includes information on 545 Nobel laureates in physics, chemistry, and medicine from 1902 to 2016. For each laureate, it lists the type of prize, the award year, and the publications associated with the prize winning work, along with their digital object identifiers (DOIs) and original publication dates. (See Brian’s post for more details on how the data were compiled.)

This was a great foundation for adding citation data. I used the OpenAlex database to retrieve year-by-year citation counts for each publication. Some entries were missing DOIs, so I manually filled in gaps when possible. I then automated the process using the OpenAlex API. I was able to find citation data for 696 of the 749 unique publications. The ones with missing information tended to be older publications or laureates' entries with incomplete metadata. (You can find my code and dataset on GitHub.)

One important caveat: how do we know which papers truly contributed to the Nobel-winning discovery? Often, breakthroughs are built on a foundation of earlier work, and the “key” paper may not be obvious. For this post, I’ve relied on the papers included in Brian’s dataset, which itself draws from the methodology used in this paper.

Total citation counts for Nobel Prize work

A simple place to start is by asking: how many total citations do Nobel Prize-winning papers receive? Naturally, these numbers depend both on how long a paper has been published and on broader trends like the growing number of papers published each year.

The scatter plot below shows each Nobel Prize-winning publication as a point, plotted by its year of publication and total citation count. (The y-axis is on a logarithmic scale.) One pattern is immediately clear: newer papers tend to receive far more citations than older ones, even though they have been around for less time.1

Citation counts over time

While total citation counts offer a snapshot of overall impact, looking at how that impact grows over time can be more revealing, especially when thinking about how to accelerate scientific progress.

The chart below shows the cumulative citation count for each prize-winning publication, starting from its year of publication. Each trace begins at zero and tracks how citations accumulate over time. The lines are color-coded by publication year: darker blues for older papers, and greenish hues for more recent ones. (The y-axis is on a logarithmic scale.) By aligning all traces to year zero at the time of publication, we can compare how different papers accumulate citations over time.

Newer papers tend to show a rapid initial rise in citations, often followed by a plateau. Older papers, by contrast, accumulate citations more slowly. But intriguingly, some older papers experience a second surge in attention roughly 80 years after publication. These are raw citation counts and don’t account for the growing volume of scientific literature over time, which likely skews results in favor of newer papers.

To better compare trajectories, I normalized each paper’s cumulative citation count by its total citations to date. The chart below shows these normalized curves, with all papers aligned to their publication year (set to year 0). I’ve also marked the year each paper won the Nobel Prize with a dot.

The plot reveals a wide range of citation trajectories. Some papers make an immediate impact, while others remain quiet for decades before suddenly gaining attention. On average, newer papers take off more quickly, whereas older papers often lie dormant for many years before their influence emerges.

The second awakening

Some older papers show a striking pattern: a second surge in citations many decades after publication. I find this fascinating. Were these papers simply ahead of their time, like theoretical work that required experimental advances to be appreciated? Or do these later bursts reflect new applications of old ideas?

Understanding the causes of these second awakenings might offer clues for accelerating scientific impact. After all, these weren’t obscure papers. They were part of Nobel-winning work, already recognized as major contributions. And yet, many remained relatively quiet for decades, sometimes more than 80 years, before gaining broader influence.

The chart below highlights only those papers with more than 1,000 total citations. While the 1,000-citation cutoff is arbitrary, it helps illustrate that this isn’t just about a paper going from 2 to 20 citations. Many of these late bloomers have become highly influential.

In the earlier plot, all citation curves were aligned to year zero at the time of publication. But I was curious if the second citation surge tends to happen around the same real-world time for these older papers? For example, was there a common trigger, like something in the 1990s, that revived interest across multiple works?

The chart below re-aligns the data to actual publication years and includes only papers published before 1950, for visual clarity. Many of these curves resemble classic exponential growth, while others show distinct inflection points, which could be possible signs of external events or developments that reignited attention.

One striking example is Irving Langmuir, whom I’ll explore more deeply in a future post. He won the Nobel Prize in Chemistry in 1932 for his work on surface chemistry. In the chart below, I’ve highlighted his four prize-associated papers in crimson, with all others shown in gray.

Three of those four papers received less than 10% of their total citations before the early 2000s, despite being recognized with a Nobel just 15 years after publication. One in particular, his 1918 paper titled “The adsorption of gases on plane surfaces of glass, mica and platinum,” now has over 20,000 citations. What was it originally recognized for? What sparked its resurgence nearly a century later? And what had to change in the field, or for science more broadly, for that revival to happen?

Sleeping Beauties of Science

In metascience, a commonly used term for these late-blooming papers is “Sleeping Beauties.” Coined by Anthony van Raan, the term refers to scientific publications that remain unnoticed or under-cited for decades before suddenly receiving widespread attention. Since then, several groups have attempted to quantify Sleeping Beauties using citation data. For example, one large-scale study analyzed over 22 million papers across the natural and social sciences to identify the most prominent examples of this phenomenon. Below is a table from this paper listing the top Sleeping Beauties.

Among the papers identified in the large-scale study were several well-known examples: the Einstein, Podolsky, and Rosen (EPR) paper, “Can quantum-mechanical description of physical reality be considered complete?”; Langmuir’s “The constitution and fundamental properties of solids and liquids. Part I. Solids”; and Wenzel’s “Resistance of solid surfaces to wetting by water.”

While the EPR paper was a Sleeping Beauty in terms of citation delay, it, and Einstein himself, were highly recognized at the time. Similarly, Langmuir won the Nobel Prize in Chemistry in 1932, but the broader impact of his prize-winning work didn’t take off until the early 2000s. Of the three, only Wenzel’s paper fits the strictest definition of a Sleeping Beauty: it received little to no recognition or impact until decades later.

So while citation data might suggest that some papers had no early influence, that doesn’t always mean they were overlooked. As this excellent essay in Works in Progress argues, there are many reasons a paper might appear to “sleep”, like from being ahead of its time to needing the right application context. Based on the examples above, even highly celebrated, Nobel-recognized work can remain dormant for much of its life.

Conclusions

In this post, I explored the impact of Nobel Prize-winning work through the lens of citation counts. While citation counts have clear limitations, they offer a useful, if one-dimensional, measure of a publication’s influence within its field.

Notably, newer papers tend to have far more citations than older ones, despite having been published more recently. This may be due to the sheer increase in the number of papers published today, as well as the tendency of authors to cite only the most immediate prior work. For example, if I develop a new microscope, I would cite recent advances in imaging and not Newton’s optics. These patterns create a compounding effect that favors recent publications.

Even among Nobel-winning papers, citation trajectories vary widely. Some take off quickly and plateau; others remain dormant for decades before exploding in impact. I find this second group especially fascinating. It’s tempting to assume these late bloomers were simply unrecognized at first. But the papers analyzed here were all identified as Nobel-worthy, yet they only reached their full influence many decades later.

So, how might we speed up these delayed awakenings? One approach is to learn from history. If we want to predict which papers could become transformative, we might start by identifying already-recognized work, like Nobel-winning research, and finding ways to bring it into new fields more deliberately, rather than waiting for serendipity. Maybe AI can help us do that.

Thanks for reading!

In my time as a researcher across different roles and scientific structures, I learn not only about neurons and optics, but also about the broader ecosystems that shape scientific work. Scientific progress is not just the product of individual brilliance or laboratory output. It is shaped by the environments we build: the labs, institutions, funding structures, and cultures that support researchers and their work.

Despite significant investments in science, we still know little about the conditions that reliably produce high-impact research. Even when insights exist, they rarely translate into structural changes within a research organization. Why did places like Bell Labs generate sustained breakthroughs? How do we identify and nurture extraordinary talent? What makes a research organization not just productive, but also dynamic and enduring?

Engineering Discovery explores these questions. This newsletter sits at the intersection of scientific research, metascience, and data-driven inquiry.

In recent decades, scientific discovery has become increasingly dependent on collaborative ecosystems. No single person is “curing cancer” or “solving the brain.” As science becomes more interdisciplinary and team-based, we need to rethink how we design our research structures so as to not just reward individual achievement, but also to create cultures, environments, and institutions that enable teamwork.

Drawing from my scientific background, where I build multimodal imaging systems to study brain health, I hope to bring an analytical and systems-engineering mindset to the scientific enterprise itself. Just as imaging combines modalities to reveal complexity in living systems, we can combine data, history, and lived experience to illuminate complexity in the research ecosystem.

Of course, data has its limits. Some of the most interesting insights live at the margins, in what might appear to be outliers. Through Engineering Discovery, I aim to blend quantitative analyses with personal narratives by zooming out to see patterns and zooming in to understand the people and environments behind the discoveries.

Some of what you’ll find here:

• Essays on the histories of scientific discoveries

• Analyses of scientific trends using bibliometric data

• Case studies on the people and environments behind scientific breakthroughs

• Reflections on institutional culture and its role in scientific outcomes

• Commentary on papers I find interesting, placed in the broader context of the field

Welcome again to Engineering Discovery! Here, I will explore scientific breakthroughs and how we might engineer better discovery processes to accelerate their societal impact. My first post will explore the impact of Nobel Prize-winning work over time by looking at citation patterns.

Please subscribe if these topics resonate with you, and reach out if you have feedback, suggestions, or just want to chat!

Smrithi