The teams that did the most interesting work were almost never the ones with the deepest shared expertise. They were the ones where people came in with genuinely different mental models, different instincts about what mattered, and different vocabularies for describing the same thing. That combination - uncomfortable as it can sometimes feel - consistently produced better questions, broader solution spaces, and results that none of the individuals involved could have reached working within their own discipline alone.
I want to say something more specific than “diversity is good,” because the vague version of that argument is easy to agree with and easy to ignore.
The pattern becomes more convincing when it holds across settings that are structurally quite different.
A competitive hackathon with a tight deadline and a hard problem - predicting which biological pathways novel compounds would affect, using high-content imaging data. The team included people from computational chemistry, software engineering, and biology. The solution we built - using transformer encoders on molecular representations alongside image embeddings - came from a conversation between someone who thought about molecules and someone who thought about language models, mediated by someone who understood what the biology actually needed to tell us. No single person had all three pieces.
A similar format, a different problem. Predicting how mutations affect protein function using deep mutational scanning data. The team spanned protein biochemistry, deep learning, and statistics. The biochemist knew which parts of the sequence space were biologically meaningful. The statistician kept asking whether what the model was learning was actually what we thought it was learning. The ML work sat in the middle, depending on both. The result was a model that generalised well precisely because it was built with those constraints in mind from the start.
This one was different in character - no competition, no winning team, just a group of researchers brought together to work on a hard problem over a week. Without a competitive frame, the collaboration became more openly exploratory. People were more willing to say they did not understand something, more willing to follow a line of thinking into territory they were not expert in. The output was a published report, but what I remember most is the quality of the conversations - particularly the ones that happened when someone from outside the domain asked a question that the domain experts had stopped asking because they assumed they knew the answer.
Three different settings, three different problems, three different team compositions. The same underlying dynamic: the most useful contributions often came from people who were not the deepest experts in the relevant field.
Every field has terminology that carries compressed meaning - meaning that can quietly distort how you think about a problem. When someone asks you to explain something without the jargon, you often find the concept is less solid than you thought, or that the jargon was obscuring an assumption worth questioning. This is mildly uncomfortable and consistently useful.
A biologist looking at a classification problem will reach for one set of tools. A software engineer will reach for another. A statistician will question whether it is actually a classification problem at all. Three different framings of the same problem is more valuable than three people with identical training each arriving at the same framing independently - even if reconciling those framings takes more time upfront.
The questions that come from genuine confusion - from someone who does not understand why you are approaching something a particular way - are frequently the ones that matter most. Not because naivety is a virtue, but because accumulated expertise builds accumulated assumptions, and those assumptions are often invisible from inside the field.
“The most useful question I heard in three days of a hackathon came from a software engineer who had never worked in biology. He asked why we were treating the problem as single-stage rather than decomposing it. It was not a biological insight. But none of the biologists had thought to ask it.”
In every one of these settings, a significant part of what made things work was not just the diversity of expertise, but the presence of people who could move between domains - who had enough of a foundation in more than one area to translate, to hold context across multiple conversations, to know when two people were actually agreeing without realising it and when they were genuinely at odds.
These people are often hard to place in a traditional academic hierarchy. They do not have the deepest publication record in any single area. They are not the most cited person in the room. But they are frequently the reason the room functions. They carry the thread of a conversation across domain lines. They notice when an assumption from one field is about to cause a problem in another.
It is a kind of intellectual generosity: a genuine interest in what other people know, and a willingness to do the work of understanding it rather than waiting for it to be translated into your own terms. Combined with enough breadth to actually follow the conversation, that quality is consistently valuable in ways that are hard to quantify and easy to overlook in hiring.
Shared expertise feels safer. A team of specialists who all speak the same language is easier to manage, easier to evaluate, and easier to explain to a funder. The risk is legible. The counterfactual - what might have been discovered by a broader team - is not.
The incentive structures in academia push against it. Grants, publications, and career progression all reward depth. A researcher who spends time building bridges between disciplines is doing something genuinely valuable that the system has limited ways of recognising. That does not make it wrong - but it does make it harder to sustain.
Hiring is also hard across domain lines. It is difficult to evaluate someone whose expertise is partly outside your own, which means interdisciplinary candidates often get filtered out early in processes designed to find domain specialists. The people who are hardest to evaluate are sometimes exactly the people a team needs most.
That is what I think interdisciplinary teams are actually for. Not just to produce better results, though they often do. But to generate the kind of thinking that only happens when people who see the world differently are given a problem and asked to work on it together. That thinking is harder to measure than a citation count or a winning model. It is also, I suspect, where most of the important ideas come from.
]]>This post tries to give an honest account of where things actually stand - what has been demonstrated, what is still unproven, and what the current wave of industry partnerships suggests about where this is heading.
The most significant single development was AlphaFold. In 2021, DeepMind’s AlphaFold 2 predicted protein structures with accuracy comparable to experimental methods - effectively solving a problem that the structural biology community had been working on for 50 years. By 2022, the AlphaFold Protein Structure Database had released predicted structures for over 200 million proteins, covering nearly every known organism. For drug discovery, where knowing the 3D shape of a protein target is often the starting point for designing a molecule that binds to it, this removed a significant bottleneck. AlphaFold 3, released in 2024, extended the approach to DNA, RNA, and small molecules, making it directly relevant to structure-based drug design.
ESM-2, Meta AI’s protein language model trained on 250 million sequences, showed something different but equally important: that a model trained purely on sequence data - without any 3D structure information - could learn evolutionary relationships and predict functional properties. This matters because sequence data is abundant and cheap to generate in ways that structural data is not. ESM-2 and the ESMFold structure predictor it enables have become practical tools for protein engineering - I used them directly in my own research at Roche, working on antibody fitness prediction.
Newer models have expanded the toolkit further. RFdiffusion from the Baker Lab at the University of Washington uses diffusion models to design entirely new protein sequences with specified properties - binders, enzymes, and vaccine candidates designed from scratch and experimentally validated. Geneformer, from the Broad Institute, was pre-trained on 30 million single-cell transcriptomes and can predict the effects of gene perturbations in silico. These are not demonstration projects. Researchers are actively using them.
NVIDIA's BioNeMo platform wraps many of these models - ESMFold, DiffDock, MolMIM - behind accessible APIs, meaning researchers can use them without managing the underlying infrastructure. Over 100 companies were using the platform as of 2024, including Amgen, Genentech, AstraZeneca, GSK, and Novo Nordisk. This is the infrastructure layer that makes the models practically accessible at scale.
The most important piece of clinical evidence to date comes from Insilico Medicine. The company used its end-to-end AI platform - combining a target identification engine called PandaOmics with a generative chemistry engine called Chemistry42 - to discover a novel drug target for idiopathic pulmonary fibrosis (IPF) and design a small molecule inhibitor to hit it.
IPF is a progressive and irreversible lung disease affecting approximately 5 million people worldwide. Patients typically die within 2–5 years of diagnosis, and current treatments slow progression but do not halt it. There is a large unmet need.
The drug candidate, INS018_055, went from target identification to Phase I clinical trials in under 30 months - roughly half the traditional timeline. The Phase I results were positive. In 2023, it entered Phase II trials simultaneously in the US and China. In November 2024, Insilico reported positive Phase IIa topline results: the drug was safe and well-tolerated, and showed encouraging dose-dependent improvement in forced vital capacity - a measure of lung function - over 12 weeks. The full results were published in Nature Biotechnology in 2024.
INS018_055 is the first drug discovered and designed end-to-end by generative AI to demonstrate efficacy signals in a randomised controlled trial. That is a meaningful milestone - not just for Insilico but for the broader question of whether AI-driven drug discovery actually produces clinically useful compounds. The answer, at least in this case, is a qualified yes - though Phase IIb and pivotal trials are still needed before anything definitive can be claimed about efficacy at scale.
It is worth being clear about what this does and does not show. It shows that AI tools can materially compress timelines and that AI-generated molecules can be safe and potentially effective in humans. It does not show that AI-designed drugs are systematically better than traditionally discovered ones, or that the approach will work across all disease areas. One data point is one data point.
The industry is not doing one thing. It is doing several things at once, and it is worth distinguishing between them.
The first model involves large pharma companies forming partnerships with specialist AI drug discovery firms, where the AI company contributes platform and capability and the pharma company contributes biological expertise, data, and clinical infrastructure.
Recursion Pharmaceuticals is probably the most prominent example. The company has built a large-scale biological operating system that translates cellular images into research findings on cellular state - and has active collaborations with Roche’s Genentech and Bayer, among others. Recursion also joined NVIDIA’s BioNeMo platform as the first third-party addition in 2024, making its Phenom-Beta imaging model available to other companies.
Cradle Bio focuses on protein engineering using generative AI, and has been working with pharma partners on antibody and enzyme design. Twist Bioscience sits at the synthesis end - providing the DNA and protein synthesis capabilities that make experimental validation of AI-designed molecules faster and cheaper. These companies are not replacing pharma’s discovery function; they are compressing specific, expensive steps within it.
The second model is different in character. Rather than outsourcing AI capability to specialist firms, some large pharma companies are building it in-house by partnering directly with NVIDIA to install serious compute infrastructure and access BioNeMo’s model ecosystem.
In late 2023, Roche’s Genentech signed a multiyear collaboration with NVIDIA, with the explicit goal of enhancing its existing AI research using BioNeMo and running its “lab in a loop” framework - a continuous cycle where experimental data feeds back into AI model training and improves molecular designs.
In early 2024, Amgen announced it would deploy an NVIDIA DGX SuperPOD - a full-stack AI supercomputer with 248 H100 Tensor Core GPUs - at its deCODE Genetics subsidiary in Reykjavik, Iceland. The system, named Freyja, is specifically designed to train AI models on deCODE’s genomic dataset, one of the largest and most comprehensive human genetic resources in the world. The goal is to build precision medicine models for discovering drug targets and disease biomarkers.
The most significant announcement in this category came in January 2026. NVIDIA and Eli Lilly announced a joint investment of up to $1 billion over five years to build a co-innovation lab in San Francisco, where Lilly’s biologists and NVIDIA’s AI engineers will work side by side. The lab will be built on BioNeMo and NVIDIA’s Vera Rubin architecture, and will focus on molecule design and simulation, clinical development optimisation, and manufacturing. It has been described as the largest disclosed AI collaboration in the pharmaceutical industry to date.
The scale of the Lilly deal is notable. It is not a research collaboration or a platform access agreement. It is a physical infrastructure investment with engineers co-located - closer in character to a joint venture than a standard pharma-tech partnership. That suggests Lilly, at least, has concluded that AI capability needs to be embedded in the core of drug discovery rather than bolted on.
None of this means the hard problems are solved. A few things are worth being clear-eyed about.
The Insilico result is encouraging but preliminary. Phase IIa is designed to assess safety and get early efficacy signals, not to prove that a drug works at scale. INS018_055 still needs to succeed in a larger pivotal trial before it can be prescribed to patients. The history of drugs that looked promising in Phase II and failed in Phase III is long.
AI models are very good at optimising for properties they have been trained on. They are less reliable when it comes to properties that are harder to measure - off-target effects, long-term toxicity, behaviour in diverse patient populations. The molecular spaces these models explore are vast, but our ability to evaluate what we find there is still limited by how much wet-lab experimental data exists to train on.
There is also a question about whether the benefits will be evenly distributed. The companies best positioned to use these tools are the ones with the largest proprietary datasets, the most compute, and the most ML expertise. That is a small number of very large organisations and very well-funded startups. Whether this leads to better medicines for neglected diseases and underserved populations, or primarily accelerates discovery in commercially attractive therapeutic areas, depends on choices the industry has not yet made.
For working scientists, the practical implication is not that AI will replace drug discovery but that it is changing which parts of it are the bottlenecks. Target identification, structural prediction, lead generation, and molecular optimisation are all areas where AI tools are now genuinely useful - not as replacements for biological intuition but as ways to explore larger spaces faster and to filter candidates more efficiently before committing to expensive experimental work.
The challenge is that accessing and applying these tools still requires an unusual combination of skills. Most bench scientists do not have the computational background to run BioNeMo or fine-tune a protein language model. Most ML engineers do not have the biological training to know what questions to ask or how to interpret the output. Closing that gap - which is partly a tooling problem, partly a training problem, and partly a communication problem - is what I think about most, and what drives most of what I write and build.
]]>