I’ve been building in biotech for the last 4 years now (although not directly in therapeutics) and have tracked the AI for bio companies that were built over the course of the last decade. Given all of the recent interest, I thought I’d share how I make sense of where people are building and how pharma as an industry could evolve.

I categorize the majority of current AI life science efforts into 2 tracks:

1. New AI enabled scientific tools - virtual cells, predictive models for toxicity, generative molecule design, genome design, etc.

2. AI for scientific reasoning - the scientific tasks involved in drug discovery: target ideation, interpreting binding data, lead candidate nomination, interpreting animal toxicity data, clinical trial design, etc.

The first creates new scientific tools that help us predict, measure, or design biology more effectively and the second teaches computers how to reason like a great scientist in a pharma. My view is that the AI’s revolution in biotech won’t come from a single model breakthrough, but rather through making each step of the drug-discovery and development system more systematically intelligent: better ideas for targets, designing experiments, designing modalities, predict toxicity, identify responders, and make clinical development decisions.

New AI scientific tools for discovery:

This bucket is the one people most intuitively associate with “AI for bio.” The original cohort of companies here are Recursion, Insitro, Exscientia, etc. all of which were founded ~10 years ago now. Collectively, these companies have raised over $2B, however, looking at the number of drugs approved, asset acquisitions, etc. there’s not much yet.

This is due to a number of reasons, a few examples include: datasets generated not being very useful (models learning vague cellular representations that just don’t translate well to humans) or wrong problem selection (how impactful is faster small molecule design if for example you choose the wrong target to start with). The failures here highlight how intricate drug discovery and development is, AI can improve a local step in drug discovery without changing the overall probability of success if the binding constraint sits elsewhere.

However, with the most recent advancements in model architectures, our ability to read biology at a higher granularity (single cell sequencing for example), and our ability to perturb biology (novel editing tools), we’re seeing new tools emerging:

1. Virtual cell models:

   These models (built by very talented folks at places like Arc) aim to predict a cell’s response to various perturbations. Drugs are ultimately designed to perturb cells in the body in specific ways. A sufficiently powerful virtual cell model could theoretically allow scientists to predict if a potential drug is shifting a cell’s state in a desirable direction (from ‘diseased’ -> ‘healthy’) fully in silico and across cell types. This is a very ambitious endeavor, a model may work on the cell lines, perturbations, or conditions it has seen, but fail when asked to predict responses in unobserved contexts. A lot of the virtual cell work today is focused around predicting a cell’s post perturbation molecular state, often its expression profile, from an initial cell state, a perturbation, and biological context like cell type for example. That makes a lot of sense for a starting point in developing this technology, but a real drug response depends on much more than expression profiles. My guess is that it will be quite a while before broad, general purpose virtual cells materially improve drug development success, but the goal is important enough that it is worth working on regardless.

   While the longer term objective of virtual cells is to build models that can generalize across cell types, there has been some really interesting work done by the folks at NewLimit, a company focused on epigenetically reprogramming cells to make them behave younger for age related diseases. They’re training models to learn the relationship between cellular state and functional age or disease severity, then use perturbation models to identify interventions that push old or diseased cells toward younger, healthier states. This is effectively a very early virtual cell model, which is exciting! These preliminary models have directly informed the discovery of a few drugs that they are now gearing up to take through clinical trials to restore youthful function of livers in patients with alcohol liver disease (along with a few other therapeutic areas). The founder, Jacob Kimmel, has a wonderful blog where he documents their approach and all their progress in more detail.

2. Models for drug design:

   Another clear category of AI enabled scientific tools are models for drug design. This space has attracted a ton of capital and talent over a much longer period of time in comparison to many other areas of AI x bio. My views here are mixed. Drug design can be valuable when a target is clearly validated but lacks a good molecule, or when a historically hard-to-drug target becomes reachable. But the list of clearly biologically validated targets simply waiting for a better molecule to be designed is probably not gigantic. There are many incredibly talented chemists out there who can design these molecules. Additionally, better design cannot answer the upstream question of whether the biological hypothesis is right. If the target biology is wrong or the disease model does not translate to humans (which many don’t), even a perfectly designed molecule will still fail. For that reason, I think many drug design efforts might be useful but not sufficient on their own to materially improve drug success.

   I am more excited when generative design models are used to innovate on therapeutic modalities versus optimizing design within an existing one. Profluent is a good example: they are using protein language models to design new gene editing systems, which could potentially overcome constraints of naturally evolved editors around size, specificity, activity, etc. Using AI to create new biological tools that were hard to discover through mining nature alone feels like a better pursuit than incrementally improving the design of simpler modalities.

3. Toxicity prediction:

   Lastly, an interesting frontier in the world of new scientific tools are models trained on datasets to predict human toxicity in ways that are more useful than animal models (like what my good friend Brandon White is doing at axi.om). Toxicity remains one of the major reasons drugs fail in clinical development, accounting for roughly 30% of clinical failures. Part of the problem is that animal models are an imperfect surrogate for human biology. An analysis of 182 molecules with paired animal toxicology data and Phase I human data showed this clearly. In the chart below, the yellow square highlights cases where drugs that were toxic in humans showed no corresponding toxicity signal in animals:

   

   Building in silico models that are potentially better at predicting in human tox than animal models could lead to massive cost and time savings in discovery/development. The FDA has expressed interest in reducing the industry’s dependence on animal models generally (source), which I think is a great tailwind for in silico toxicity prediction broadly. This problem also seems well shaped for AI. Many important tox signals like hepatotoxicity, cardiotoxicity, genotoxicity, mitochondrial toxicity, etc. seem to have distinct molecular, cellular, or phenotypic signatures. Models that are trained on the right human derived datasets, could learn these signs of safety problems and could save pharma companies billions of dollars if they’re accurate, which I think is quite tractable, at least for certain molecule classes like small molecules.

   What I like a lot about this is the fact that a) these models can be trained on real human data and b) they’re geared towards tackling a real clinical development problem. A lot of the efforts on AI applied to biotech are focused pre-clinically, whereas 90% of R&D dollars lost, time lost, etc. is in clinical development.

This section wasn’t meant to be collectively exhaustive, there’s plenty of really interesting work done in other areas that I’d categorize as ‘novel scientific tools’, but I’m highlighting a few areas with examples where I’ve seen talent and capital cluster in.

Scientific execution and reasoning:

The second category is much more recent and pertains to teaching computers how intelligent scientists might reason about scientific problems from drug discovery through to development. Some interesting work done on this end by folks at Edison Scientific, Lila.ai, etc. If I had to guess, this is likely what folks like Anthropic are planning on building. With the rise of agents, lab robotics, and models that have stronger reasoning abilities, I think these applications have been made increasingly possible. Image below from Lila.ai :

There are a very large number of questions, functions, and tasks that any pharma company goes through. A few examples (out of many) include things like:

• Given human genetics, disease biology, expression data, prior literature, and clinical precedent, is target A worth pursuing for this disease X?

• Given potency, selectivity, tractability, early ADME, PK, off-target activity, is this hit series worth pursuing?

• Post in vivo mice studies, given potency, selectivity, PK/PD, in vivo efficacy, preliminary safety, formulation, etc. should this molecule be advanced into IND-enabling studies?

• Given mechanism, disease biology, prior clinical precedent, regulatory expectations, how should we design this clinical trial? (patient population, endpoint strategy, dose, duration, and go/no-go criteria, etc)

The challenge in pharma is that teams are integrating data from imperfect animal models, an incomplete understanding of human biology, and in some cases no clinical precedent (if it’s a new target). Even very strong teams get these decisions wrong and in large pharma companies teams are siloed, incentives are misaligned (scientists are not rewarded for killing programs), and sunk cost fallacies exist.

As a result, many programs are advanced that should have been killed earlier while other programs are shelved despite having real potential. Companies have been built successfully (Roivant) to exploit these inefficiencies and bring drugs to market that otherwise would’ve remained on the shelves. Given the inefficiencies of the industry I think there are probably still more opportunities like this.

Scientific reasoning is an area AI has a good chance of being impactful in. For a given set of fairly well defined tasks across discovery or development, models can become sufficiently capable at addressing these questions/tasks individually. And overtime, agents might be able to reason across the various questions involved in taking a drug from the earliest stages to clinical trial design execution.

In the future, agents will scour the biomedical literature + internal datasets to identify and validate promising novel targets, design assays to measure target modulation, coordinate with end-to-end CROs like WuXi to start executing hit to lead programs, while interpreting data from the CROs along the way. These agents can execute in parallel across on-going drug programs while also keeping in mind the context of the entire pharmaceutical company’s history to guide its decision making on the optimal experiments to run. Theoretically, this should lead to far less inefficiencies in the industry and should increase the number of novel medicines brought to market.

Pharma companies face a wide range of capital allocation and scientific decisions. Historically these decisions have been made by humans, which has inherent biases and limitations. If an increasing portion of the reasoning and decision making in a pharma company is done by agents that are supposedly more rational or have better judgement, pharma companies might start to look a lot more ‘quantitative’. Decisions will be made more algorithmically, absent of any cognitive biases, which should improve performance and ideally results in novel drugs and more medicines brought to market.

I think that the most important application in reasoning is target identification (what molecule do we want to design a drug for and what therapeutic area may this be relevant in) and systematizing the process of good target idea generation. Out of all the decisions a biotech has to make, deciding what target to develop your drug for is one of the most important decisions. Synthesizing as much of the increasing body of biomedical knowledge (+ internal datasets) and coming up with good drug target ideas repeatedly, I think would be the most significant impact of AI in biotech.

Closing thoughts

Advancements in algorithms are occurring much faster than we can develop drugs, which means every few years there will likely be a new batch of AI companies formed around the latest model capabilities and the datasets that have become newly usable. The first wave of these companies haven’t panned out clinically, at least relative to the original ambitions. The reality is that we are still learning where these models actually map onto the bottlenecks in drug discovery and development.

I’ve outlined the areas where I’m seeing a lot of talent and capital concentrate. My suspicion is that the two will increasingly reinforce each other. Better virtual cells, toxicity models, perturbation maps, and design systems should create richer scientific inputs; better reasoning systems should help decide which of those inputs matter, what experiments to run next, and which programs are actually worth advancing.

I’m particularly excited by companies that are grounded in human data and focused real discovery or development bottlenecks. Target identification feels like the most important discovery question: if parts of molecule design become increasingly commoditized, then the value shifts toward knowing which biology is actually worth perturbing. I’m also excited by models that improve translation later in development, especially human toxicity prediction, patient response prediction, or even making it cheaper/faster to run trials, given that this is where the vast majority of dollars are spent. I don’t think any single model, dataset, or breakthrough will be enough on its own to meaningfully bring more medicines to market. The impact will come from compounding improvements across the system: better target selection, better molecules, better tox and patient selection/response prediction, better trials, etc. If AI changes biotech, I suspect majority of its impact will be by making drug discovery and development more systematically intelligent over time.