What does the future of science look like in an AI-driven world? Beyond high-profile moves — such as Anthropic opening a wet lab — a quieter, widespread transformation is already under way. In a guest post, Adrian Sanborn of Endura Therapeutics describes the practical, near-term ways AI is reshaping front-line research in his company.
Sanborn holds a PhD in computer science from Stanford and spent much of it working at the bench in Roger Kornberg’s biochemistry lab, giving him the rare perspective of seeing what a large language model (LLM) does to both a codebase and a wet lab.
Two adaptation strategies: foundries and navigators
Biotech has adapted to AI through two complementary approaches:
-
Foundries reduce the cost of doing experiments. These platforms industrialize measurement, generating data an order of magnitude faster via new technologies and automation. Examples include Xaira, NewLimit, Octant, Tahoe, Endura (next‑generation sequencing and multiplexing); Insitro, Eikon, Noetik (high‑throughput microscopy); and Lila and Periodic Labs (physical automation). AI makes the resulting data legible and predictive, but the differentiating asset is the experimental data itself.
-
Navigators spend the surplus of thinking. Here, AI models are embedded in everyday company machinery to drive better decisions and faster processes: they influence how work is done, which tools are built, and which questions are worth an experiment. You don’t necessarily need a proprietary model or massive dataset — just a willingness to change how your company operates.
Building a foundry is a strategic, capital‑intensive, multi‑year commitment to a particular technology and is easy to spot. Navigation, by contrast, is operational and invisible: there’s rarely a press release about a path a team decided not to take. The visible change is concentrated at a few dozen companies; the quiet operational change is happening at nearly all of them.
Navigation runs fastest at early‑stage startups
Early startups have no legacy to shed — no long software contracts, standardized processes, or calcified org structures — and are under pressure to move fast with little. When a better way to work appears, it simply becomes the new normal.
The effects are everywhere: experiments iterate faster when analysis takes an hour instead of a week; software that would once have been licensed for six figures can be built in a day; disease programs can be selected from 500 candidates rather than the usual five.
Code now keeps pace with the science
Experimental science has a structural tension absent from most engineering: in research, changing requirements are the objective because each experiment is designed to learn something new. If an approach hasn’t evolved in six months, it often means no discovery is occurring.
A new experiment’s protocol will evolve many times in its first year, and each change ripples into analysis. Every measurement needs processing, normalization, and interpretation by code that tracks the experiment closely. Historically, this analysis was done by a separate computational person, creating a seam between the researcher who understands the measurement and the engineer who understands the code. That friction produced a tendency to propose fewer experimental changes to avoid reworking analysis, leaving options unexplored.
Now that writing code is fast, adapting analysis to a modified protocol can take an afternoon instead of 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 are easily fixed.
That shift also applies to interpretation: interactive visualization dashboards can now be built in minutes rather than days. Endura’s internal dashboard was built in a few hours.
The most visible consequence is access. Previously, results often waited in a computational queue; now the scientist who ran the experiment and holds the context can present their own results. Data is no longer gatekept behind someone else’s Python notebooks.
Software can now express your opinion
Every software interface embeds an opinion: it decides which comparisons are one click away and which require hunting. For two decades those opinions were set by a handful of vendors building lab software as a system of record, designed toward the lowest common denominator for many labs.
That is changing. An in‑house data portal can accommodate data quirks no commercial product would anticipate and be exactly as complex as the team needs, growing with their questions. Browsing and exploration become effortless, which changes behavior: patterns that were trapped in slide decks begin to surface.
Implementation can take a day, but deciding what the portal should do can take weeks. Those design discussions are critical because forcing a team to pick what belongs on a single screen forces them to articulate which comparisons actually drive decisions. When you buy a platform you outsource not only engineering but also the question of how you accomplish your goals.
There are trade‑offs: an in‑house portal can be less polished and have no external support team, and larger organizations with validation requirements and contractual obligations will struggle more to follow. But software companies long ago learned that the best internal tools come from engineers embedded alongside users. 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: a single successful program typically requires about a decade and roughly a billion dollars, so the decision is diligenced carefully. Traditionally, a mixed group of internal and external experts gathers and debates evidence for a month or more, assuming there’s already a shortlist of about five diseases to argue over.
Endura’s case differed because of its therapeutic mechanism. The company is developing a CRISPR‑in‑a‑pill approach: an orally administered small molecule that creates a chemical scar on a specific genetic message to shut off a disease‑causing protein. Finding such drugs required 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 with five diseases, Endura had to triage the entire disease map. They built a two‑stage triage and deployed a fleet of LLM research agents. The first pass covered about 500 disease targets, producing roughly three‑page reports each and filtering on foundational questions: is the disease prevalent enough, is it already addressed by existing drugs, and would lowering the target actually relieve disease? The second pass narrowed ~100 targets and produced about 30‑page dossiers each, working through disease biology and the competitive landscape thoroughly. The prompts for the second pass were written to act like skeptical experts: name programs that failed, explain why each failed, and state what would need to be true for Endura to succeed where others didn’t.
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 outputs are still checked against primary sources and selected programs receive full human diligence. But a search this broad and this deep was not possible a year ago.
In research, the expensive mistakes are the unknown ones: a team commits to a direction and months later a negative experiment reveals the mistake. Often a specialist could have said in a sentence that a pathway was tried or a readout never predicted anything. Access to expert‑level depth at scale is a game changer because being told “no” early is so valuable.
This is only the shallow end
Everything described above at Endura — flexible, dynamic analysis; internal tools built in a day; a 500‑disease search — is only the beginner version of navigation. Each generation of language models removes constraints we once accepted. Soon we may skip building persistent analysis pipelines or dashboards: a scientist will ask the question she actually has, and the analysis and interface to answer it will be assembled on demand. Software will stop being a product and become something that appears around the question.
This access to expertise at scale opens work that nobody could attempt before. Drug repurposing is a clear example: safe, approved drugs may act on mechanisms nobody has been looking at, and the literature contains many useful but undiscovered connections. A model can digest the literature and, with appropriate prompting, connect the dots. A whole ecosystem of companies is forming around this bet; pharma and investors are running their own versions. The open question is whether this produces three new drugs or 300.
Sanborn’s conclusion is that the most accelerating AI in science today is often not a model trained on scientific data but the system that helps a scientist or executive decide what is worth doing each Monday morning.
Author and acknowledgments
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, much of which he spent at the bench in Roger Kornberg’s lab. He is @AdrianSanborn on X. Sanborn thanks Brandon Anderson, swyx, and Lauren Richardson for reviewing drafts and providing feedback.
(While this post was being finalized, an empirical paper appeared that observed many of the same insights in AI x Science.)
