In this guest post, Adrian Sanborn talks about the less flashy but more immediate ways he sees AI transforming front-line scientific research in his own company, Endura Therapeutics.

Adrian did a CS PhD at Stanford and spent much of it running experiments at the bench, which makes him one of the rare people who can tell you what an LLM is doing to a codebase and to a wet lab. Enjoy!

Language models have transformed how software gets built. Writing code, wrangling data, and architecting systems now move at a speed unthinkable three years ago.

When the product is software and the result is verifiable, cheaper coding turns directly into more software and more builders.2 But in science, everything is ultimately gated by physical experiments that take days or weeks to verify anything. Knowledge work around the experiment has become dramatically faster while AI has done little for the throughput of the experiment itself. Thinking got cheap and doing did not.

We think the biotech industry has adapted in two ways:

• Foundries shrink the cost of doing. They industrialize the measurement, using new technology to generate data an order of magnitude faster than before. Xaira, NewLimit, Octant, Tahoe, and Endura accomplish this with next-generation sequencing and multiplexing; Insitro, Eikon, and Noetik with high-throughput microscopy; Lila and Periodic Labs with physical automation, just to name a few. AI makes that data legible and predictive, but the differentiating asset is the experimental data itself.

• Navigators spend the surplus of thinking. The AI models sit in the ordinary machinery of the company, driving better decisions and faster processes. They increasingly govern how the work is conducted, what tools get built, and which questions are worth an experiment. A proprietary model or a massive dataset are not required, only a willingness to evolve how the company works.

Building a foundry is a genuine strategic commitment that takes capital, years, and a bet on a particular technology. Foundries are easy to see: the technology captures the imagination, the connection to AI is immediate, and there is always some new model or dataset to announce. Navigators are invisible by comparison because the gains are operational and nobody issues a press release about a path they decided against. But navigation is available to every company. The prominent change is happening at a few dozen companies, while the inconspicuous one is happening at all of them.

Navigation runs fastest at early-stage startups, which have no legacy to shed: no history of software contracts, standardized processes, calcified org structure, or compliance regime. They are also under pressure to move fast with very little. When a better way to work appears, it simply becomes the new normal.

The impact shows up everywhere. Experiments iterate faster when analysis takes an hour instead of a week. A category of software that would have been licensed for six figures becomes a one-day build. Disease programs get chosen from 500 candidates where a team could ordinarily evaluate five. Here’s what it looks like from the inside.

Code now keeps pace with the science

There is a structural tension in experimental science that software engineering has no real equivalent for. In engineering, requirements that change every few weeks are a symptom of poor planning. In research, they are the objective. The purpose of an experiment is to learn something, and that learning changes what the next experiment should be.3 If an approach has not evolved in six months, it means nothing is being discovered.

A new experiment’s protocol will evolve a dozen times in the first year, and every one of those changes propagates into the analysis. Each measurement has to be processed, normalized, and interpreted with code that tracks the experiment closely. Historically this analysis was done by a second person, creating a seam between the person who understands what the experiment is measuring and the person who understands what the code is doing. From this friction arises the tendency to propose fewer experimental changes to avoid analysis rework, which compounds into options left unexplored.

Now that writing code is fast, adapting the analysis to a modified protocol is an afternoon’s work rather than a project. The experiment is no longer constrained by the burden of changing the analysis pipeline. Experiments can be agile when flexibility is cheap and problems can be easily fixed; in other words, science gets to “move fast and break things.”

The same shift also applies one layer up, to interpretation. An interactive visualization dashboard can now be built in minutes, down from days.

The most visible consequence is access. Previously, when every experiment was analyzed by the computational person, results waited in a queue. Now the scientist who ran the experiment and has the context presents their own results. The data is no longer gatekept behind someone else’s Python notebooks.

Software can now express your opinion

Every software interface has an opinion. A data system decides which comparisons are one click away and which require hunting. Every lab needs somewhere to store and display its data, and the opinion embedded in that system ends up shaping what that lab notices.

For two decades that opinion was formed by someone else: a handful of vendors who build lab software that acts as the system of record. These vendors build one system for a thousand labs and necessarily design toward the lowest common denominator. Everyone accepted the approximation, because developing your own was more work than any lab could justify.

This is no longer true. A data portal built in-house accommodates the quirks of the data that no commercial product would have anticipated, and is exactly as complex as the team needs, growing as their questions do. Browsing and exploring become simple and effortless, which changes behavior. Consider how little time anyone would spend on social media if seeing the next post required switching tabs and copy-pasting. Patterns that have been sitting in separate slide decks start to surface.

Implementation takes just one day, but deciding what the portal should do can take weeks. Those design discussions turn out to be critical, because deciding what belongs on a single screen forces a team to articulate which comparisons actually drive its decisions. When you buy a software platform you outsource not only the engineering but the question of how you accomplish your goals.

There are tradeoffs: an in-house portal is less polished and there is no support team to call. Larger organizations, with layers of validation requirements and contractual obligations, will still struggle to follow in these footsteps. But software companies have long understood that the best internal tools come from engineers embedded alongside the people who use them. Now every research team can be its own forward-deployed engineer.

The old rule was “never build what you can buy”. The new rule is build the tools that shape how you think.

Expert-level depth now scales

Choosing which diseases to pursue is the most consequential decision a drug company makes. Everything is downstream of this decision and built to accommodate the specifics of the disease biology and its market. The choice is effectively irreversible, with a single successful program requiring about a decade and a billion dollars, so the decision gets diligenced carefully. A typical process convenes a group of internal and outside experts who gather, synthesize, and debate the scientific and market evidence for a month or more.

That process assumes you already know which five diseases you’re arguing about. We didn’t have this shortlist at my company, Endura, because of the unique mechanism of our medicines. We’re developing CRISPR in a pill: a drug, taken in the convenience of your home, that forms a chemical scar on one specific genetic message and shuts off production of a disease-causing protein.4 Finding these drugs required developing a new DNA sequencing method that reads those scars across every gene at once, so a single experiment returns candidate drugs across hundreds of diseases. Instead of starting from five diseases, we had to triage the entire map of disease.5

We built a two-stage triage and pointed a fleet of LLM research agents at it. The first pass covered about 500 disease targets, generating the equivalent of a three-page report on each and filtering on foundational questions: is the disease prevalent enough to justify our efforts, is the problem already addressed by existing drugs, and would the target-lowering effect of our drug actually relieve the disease. The second pass, on the roughly 100 remaining disease targets, produced the equivalent of thirty pages each, working through disease biology and the competitive landscape thoroughly. We wrote the second-pass prompts to behave like a skeptical expert rather than a summarizer: name the programs that failed, why each failed, and what would have to be true for us to succeed where they didn’t. This level of detail is necessary because, as in any market, the clearly good targets are crowded. Arriving at a defensible position means finding the specific disease and the specific reason our drug will do something that existing approaches cannot.

At the old rate, the first stage would have required about one person-year of reading, and the second closer to a century of expert time. The second pass still gets checked against primary sources and selected programs receive the full human diligence it always would have.6 But a search this broad, at this depth, simply was not possible a year ago.

In research, the expensive mistakes are the unknown ones. A team commits to a direction and finds out it was wrong months later when the experiment comes back negative. Often a specialist could have said so in a sentence: that pathway has been tried, this readout has never predicted anything, that company faltered on that patient population. Access to this kind of expert-level depth at scale is a game changer exactly because being told “no” early is so valuable in research.

This is the shallow end

Everything above happened at Endura — flexible and dynamic analysis, internal tools built in a day, a search across 500 diseases — and it is just the beginner version of navigation. Each subsequent generation of language models removes constraints we had taken for granted. Soon we could entirely skip building an analysis pipeline or data dashboard. Instead, a scientist will ask the question she actually has and the analysis and interface to answer it will be assembled from scratch. Software stops being a work product and becomes something that appears around the question.7

Access to expertise at this scale opens work nobody could attempt before. One clear example is drug repurposing, where a drug already proven safe in humans turns out to act on a disease mechanism nobody was looking at. The published literature is enormous, and some number of useful conclusions are sitting in it right now, unfound because no single person has read the right combination of papers. But a model can digest it all and, with the right prompting, connect the dots. A whole ecosystem of companies is now forming around that bet, and pharma and investors are running their own versions.8 What remains to be seen is whether this produces three new drugs or 300.

The navigators are testimony that the most accelerating AI in science right now is not a model trained on scientific data at all. It is the one that helps a scientist or executive figure out what is worth doing every Monday morning.

Adrian Sanborn is CEO and co-founder of Endura Therapeutics. He was a founding member of Atomic AI, where he led the biology side of the technology platform and defined the company’s therapeutic strategy. He holds a PhD in computer science from Stanford, most of which he spent at the bench in Roger Kornberg’s biochemistry lab. He is @AdrianSanborn on X.

Many thanks to Brandon Anderson, swyx, and Lauren Richardson for reviewing drafts of this post and providing critical feedback.

One struggle we’ve had: approximately NONE of us love the “AI for Science” moniker. We are excited to launch our Science section with Melissa Du, who proposed a useful taxonomy framework for thinking about how money and talent are funneling into 2-3 main approaches… and how they combine in a coherent plan for progress. By coincidence, she introduces many of our upcoming guests on the Science pod!

We’ve witnessed the meteoric rise of LLMs over the past 5 years. Through scale alone, the models have grown from naive stochastic parrots into entities we credit with agency and emotional depth. 66% of physicians use AI in the clinic; 47% of software developers rely on AI coding assistants daily (surprised this number isn’t higher…). 79% of law firms report AI adoption in document review. AI has nearly mastered language and humanity’s digitized knowledge.

When Dario Amodei published Machines of Loving Grace in 2024, he promised that AI would eliminate all bodily and mental ailments, resolve economic inequality, and create material abundance. But these aren’t exclusively language problems. Curing cancer, designing new materials, and solving energy will require AI that can interface with and predict the physical world, not just reason about text. Modern day AI for science discourse largely conflates progress in language models with progress in scientific modeling more broadly. And while the former will certainly accelerate our capacity to understand the natural world, the field of scientific modeling has had its own tribulations and successes predating the launch of ChatGPT.

Dario himself gestures at this distinction in his manifesto. He writes that AI can accelerate the eradication of disease by “making connections between the vast amount of biological knowledge humanity possesses” and “developing better simulations that are more accurate in predicting what will happen in humans.” He’s describing two different types of AI models.

The first reviews literature, draws deep connections across studies, generates hypotheses, designs experiments, and updates its priors. It excels at reasoning, digesting large corpuses of knowledge, and keeping many ideas in working memory—precisely the areas where LLMs excel. With all due respect to the academics (I identify as one myself), we’ll call these models the scientists.

The second learns dynamics directly from data. It predicts outcomes within specific physical domains and learns structure that language alone cannot represent. We’ll call these models our simulators.

Scientists (LLMs) and simulators (domain models) are different programs within ML research that require distinct talent pools and data infrastructure, and they must work together to produce coherent models of the world. The full stack for AI-driven scientific discovery looks like this:

The capacity for LLMs to accelerate (and eventually replace) scientists cannot be overstated. OpenAI recently published a blog post demonstrating how ChatGPT could be plugged into a wet lab loop to accelerate a molecular cloning protocol. Anthropic has rolled out a suite of features for biologists that integrate Claude with existing platforms like Pubmed (paper repository), Benchling (experiment tracking), and 10x Genomics (single cell and spatial analysis).

These types of integrations are examples of giving LLMs tools for accomplishing real-world tasks, exactly the thesis undergirding the rise of agents. In these cases, OpenAI and Anthropic have provided their chatbots with hands to run wet lab experiments and access existing data platforms, but the past two years have seen the rise of magic blackboxes for search (Exa, Perplexity), document processing (Reducto), browser use (BrowserBase), etc, all of which are arguably more targeted towards agent than human use.

The blackbox tools that are necessary for LLMs to be effective in scientific discovery go beyond software. We’ll need abstractions for scientists to measure physical systems–automated laboratories, programmable in situ monitoring systems, and the like (data infrastructure). We’ll also need blackboxes for understanding and predicting the behavior of physical systems, which offer a fidelity to the world that language alone cannot capture (simulators).

Physics is simple, biology is complex

Why are simulators valuable? The core distinction between scientists and simulators, per our definition, is the reliance on text and reasoning as opposed to the reliance on domain-specific foundation models. Reasoning is sufficient when a domain has enough theoretical structure to support chain-of-thought derivation, but when theory is lacking, we require models that can learn directly from the data. The question of when this transition happens points to a deeper tension at the heart of scientific modeling: when can you derive predictions from theory and first principles, and when do you have to build models that pattern-match from empirics?

Silicon Valley loves “first-principles thinking,” and, as it happens, so do academics. Everything is derivable from first principles. Chemistry emerges from physics, biology from chemistry, cognition from biology. If you encode the fundamental laws and apply enough computation, everything can be simulated.

This is true! But unfortunately not as helpful for simulation as we’d like. Everything is atoms, but modeling atoms quickly becomes computationally intractable:

• The equations of quantum physics define how electrons interact, but our computers can only solve them on the order of tens of atoms.

• The next layer up is density functional theory (DFT), an approximation for electron clouds that scales to hundreds of atoms.

• Then, we replace our electron interactions altogether with force fields — the basis of classical molecular dynamics — which takes us to millions of atoms…

Every time we’ve hit a wall of complexity, we’ve developed new ways to measure systems and new abstractions and rules that enable us to reason about them. Rule-based simulation has given us countless early successes. Numerical weather prediction extended reliable forecasts from one day to seven by relying solely on fluid dynamics. Every transistor on every chip is simulated from Maxwell’s equations before fabrication.

But theory-driven simulation only works when the system permits elegant compression. Physics discovered that you don’t need to track 10²⁵ individual molecules to predict how a gas behaves; temperature and pressure are sufficient for prediction. A handful of forces and symmetries are sufficient to parametrize the properties of gasses, the motion of planets, the behavior of circuits.

The same is not true for biology. Or at least, we’ve yet to find an equivalent shortlist of parameters. The genome is three billion base pairs with complex regulatory logic. Expression patterns change depending on the cell cycle, tissue types, and the local chemical environment. A typical human cell contains roughly 10 billion protein molecules engaged in somewhere between 130,000 and 650,000 distinct types of protein-protein interactions. And unlike the molecules in a gas, the specific identities of the molecules matter. The specific transcription factors, regions of DNA, molecules, and interactions between them could determine whether a cell becomes cancerous. The microscopic details can't be averaged away.

from: Steve Jurvetson. Seriously… watch and realize this is happening in your body!

Biology has accepted certain theories—genetics, the Central Dogma, evolution—but these are descriptive, and lack the computational precision to power predictive engines. Thus, biology is traditionally taught as a dull exercise in memorization, the aggregation of never-ending facts, with very few unifying frameworks on which to perform inference or computation.

Chomsky, a linguist who spent his career developing the theory behind language, embodied the failure mode of theoreticians with respect to the field of linguistics. He wrote, in criticism of the technology behind ChatGPT, that the human mind is NOT “a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data.” Maybe. But maybe the universe doesn’t owe us interpretable laws (and we are glorified pattern-matching machines after all).

It’s underappreciated how far the humble neural network has come. After early theoretical critiques in the 1960s, neural approaches fell out of favor for nearly two decades; anyone working on machine learning was largely seen as a joke, dismissed for chasing a scientific dead end.

I had a stormy graduate career, where every week we would have a shouting match. I kept doing deals where I would say, ‘Okay let me do neural nets for another six months and I will prove to you they work.’ At the end of the six months, I would say, ‘Yeah, but I am almost there, give me another six months.’

— Geoffrey Hinton, “Godfather of AI”

The turning point came in 2012, when AlexNet, an image classification model, won an image classification benchmark known as ImageNet and became the watershed moment that convinced the broader ML community of the capacity for neural networks to scale. LLMs followed on as an even more prodigious success, exemplifying the necessity of data-driven learned simulators and launching the transformer architecture to well-deserved acclaim.

Attention is not all you need

The scaling laws transformed modeling in scientific domains as well–learned simulators are outperforming traditional physics-based methods in both accuracy and speed:

• Weather forecasting relied on Numerical Weather Prediction for over 50 years—physics-based models simulating atmospheric dynamics. The European Centre for Medium-Range Weather Forecasts (ECMWF) refined this into the gold standard, and their models served as the backbone for weather.com, national weather services, and Google Weather.

  • In 2023, Google DeepMind’s GraphCast, a graph neural network, exceeded ECMWF’s accuracy while making 10-day forecasts in under a minute on a single TPU, compared to hours on a supercomputer for traditional methods.

• Protein structure prediction tells an analogous story. AlphaFold2, created in 2021, directly maps amino acid sequence to structure, incorporating insights from evolutionary history encoded in multiple sequence alignments. It has since predicted structures for over 200 million proteins, covering essentially all known sequences.

• Materials discovery was revolutionized when GNoME, DeepMind’s Graph Networks for Materials Exploration, discovered 2.2 million new crystal structures, a feat that previously required approximately 800 years of traditional experimental discovery.

To be clear, none of these successes came from raw observation of data. GraphCast was trained on the output of fifty years of physics-based weather modeling. AlphaFold’s alignments encode evolutionary constraints as a biological prior. GNoME’s active learning loop uses density functional theory as the ground-truth oracle. In each case, ML learned to approximate or accelerate existing scientific knowledge; the theory came first. Moreover, progress along these alternative ML approaches isn’t necessarily bundled with the advances in ML that drove ChatGPT–the transformer architecture and its legacy. GraphCast is a graph neural network. Research has shown hybrid architectures including state space models to be most effective for large-scale DNA modeling. Different domains may have different inductive biases and data structures relative to language.

TLDR: ChatGPT releases and Epoch evaluations are a spectacular horse race, but certainly not the only one to invest in. There’s a massive long tail of ML problems that the frontier labs (outside of DeepMind) have yet to invest in.

Slaves to the physical world

Biology is arguably where the simulator is both the most necessary and least developed. The accessibility of data in other domains is a largely solved problem. Weather had ERA5 reanalysis data—decades of global atmospheric observations, assimilated and quality-controlled, publicly available. For materials science, training data comes largely from DFT calculations, which are expensive but automatable.

But biology wet-lab data is slow, noisy, expensive, often proprietary, and has been historically impossible to translate to real world validity. Cell lines don’t reliably predict what will happen in humans and animal models fail constantly; over 90% of drugs that work in mice fail in human trials. Single experiments can cost millions of dollars over the course of months. Sequencing has become cheap, but sequencing is only one modality. Predicting gene expression from sequence is hard. Predicting protein function from structure is hard. Predicting drug efficacy from molecular interactions is very hard. Arguably, we don’t even know what the right data to collect looks like.

Noetik has distinguished themselves with a multimodal approach of training cancer world models on multiplex protein staining, spatial gene expression, DNA sequencing, and structural markers. Biohub has been racing to build diverse measurement tools across scales, from individual proteins to whole organisms. Generating data across a plurality of modalities for a plurality of models is the strategy of having no strategy (and it applies to the entire field of biology).

It’s even possible that holistic theories of biology will continue to evade us. If so, progress will look less like physics and more like engineering–narrow focuses on particular diseases, particular organs, particular modalities. The work is unglamorous and the timelines are long. We remain slaves to the physical world.

AGI AI for science timelines

So why not just wait for the LLMs to figure it out? There’s credible evidence that the big labs have invested direct effort into building AI scientists for ML research, a potential route towards recursive self improvement. But even if LLMs build or significantly accelerate the creation of accurate simulators, the scientist and simulator systems can still be distinguished on their technical basis, data requirements, and deployment timelines, which have tangible impacts on investment and policy. GPT-7 may very well have the cognitive capabilities to design digital twins that simulate human biology, but it will have been enabled by many other players already advancing the algorithms behind effective simulators and building automated data infrastructure. In the same way that ML for world models, voice, and image generation were pushed forward by ElevenLabs, Midjourney, and WorldLabs, among others, we should expect ML for science to be pushed forward by a plurality of efforts.

The scientist (LLMs for reasoning and synthesis) is being built by the frontier labs.

The simulator (domain-specific ML models) requires specialized architectures and domain expertise. DeepMind has done impressive work here, but it’s not their core business.

Data infrastructure (automated labs, high-throughput assays, simulation pipelines) requires capital-intensive physical facilities and years of iteration.

There are already a few companies that are betting on this thesis for AI-driven scientific discovery, but the funding landscape remains arid. Lila Sciences, backed by Flagship Pioneering, is building “AI Science Factories”—automated labs where AI designs experiments, robots execute them, and results feed directly back into model training. Periodic Labs, founded by former OpenAI and DeepMind researchers, combines AI models with automated synthesis to create new materials.

Unless Anthropic starts building protein folding models, we shouldn’t expect them to solve diseases. Unless OpenAI starts building geospatial models, we shouldn’t expect them to become frontier weather forecasters. The big labs are focused on intelligence—reasoning, long context, tool use. Domain-specific simulation and data collection are massive undertakings that lie outside their core competencies and business models.

The discourse is focused on AGI timelines and the scientist’s capacity to reason. The work that will cure diseases and discover materials is more specific, more pluralistic, and more bottlenecked by the physical world.

Melissa Du is a Research Engineer at Radical Numerics and is on X and on Substack. Give her a follow!