Where will AI actually move the needle in drug discovery?

Artificial intelligence is now a part of everyday language in drug discovery. New companies launch every week with AI at the core, large pharma partnerships are announced monthly, and capital has flowed into some major players at a scale the sector has never seen before.

But if we step back from the headlines, the central problem in our industry has not yet changed: most drug candidates still fail in clinical trials. Depending on the dataset you look at, around 85–90% of programmes that enter the clinic do not reach approval. When commercial considerations are set aside, the single biggest driver of that failure is lack of efficacy– because the biology explored in the lab fails to translate as we hoped into efficacy in patients. Indeed, new medicines appear to be taking longer and costing more to develop per successful approval, despite continuing scientific and technological advances (a phenomenon enshrined in “Eroom’s Law”). A lot of hope, as well as money, has been invested in AI to try to buck this trend.

Improving clinical outcomes is the key goal we are all working towards (or put another way, can we improve the probability that a drug will work in patients?). However, much of the attention in AI in drug discovery is taken up with compound optimisation rather than biological discovery. Will the former significantly improve clinical outcomes? Meanwhile, recent huge investments have been lavished on trying to predict biological outcomes through AI methods. Is this concentration of resources wise, and how can the majority of companies, without access to this kind of capital, contribute?

Considering these questions can help us understand the successes and limitations of AI in drug discovery today and where bigger impacts could come.

Meaningful efficiency gains

AI has already had a real impact on the chemical aspects of drug discovery. Generative models can now propose synthetically tractable molecules with predicted physicochemical properties. Binding affinity, solubility and ADME profiles can be estimated before a compound is ever made, and structure prediction tools such as AlphaFold have compressed months of tough structural biology work into seconds. Beyond small molecules, models of similar power are aiding antibody and oligonucleotide design.

These advances matter. Fewer compounds need to be synthesised, lead optimisation cycles are therefore shorter, and teams can progress assets to the clinic more quickly. Companies that maximise these tools can run more programmes, meaning more shots on goal.

It is worth being clear-eyed about the scope of these gains, though. Saving on time to optimise a lead, for example, by reducing the number of compounds needed to be synthesised and profiled, may shave off a year or two in the discovery phase. But the cost of discovery, already declining as assays become more miniaturised and synthetic chemistry increasingly offshored, is considerably less than the cost of development, in particular running clinical trials.

Most importantly, chemistry AI does not, on its own, solve the translation problem. A perfectly optimised molecule will still fail in Phase II if the underlying biological hypothesis is wrong.

Furthermore, there is no evidence yet that companies that heavily rely on AI drug discovery tools are coming up with more novel biology. In fact, the phenomenon of “herding”, whereby many companies follow one fashionable drug target, has been significantly increasing in recent years. Being able to increase shots on goal could simply fuel herding further.

What is arguably most valuable in the quest for better clinical outcomes is a deeper understanding of disease biology. But achieving this goal remains stubbornly resistant – particularly in complex, multifactorial diseases.

The right ambitio

The ambition to build a computational “virtual cell” model capable of predicting how any cell responds to any perturbation needs no defence – the implications for target identification alone would be profound. This depth would allow researchers to forecast therapeutic effects before a compound is ever tested in the lab, to turbocharge a route to the clinic through careful modelling of dose-dependent toxicity, and then to select the right patients for whom their molecule will be efficacious.

It is clear that the field is taking this seriously. Recently, Xaira Therapeutics announced X-Cell, a 4.9-billion-parameter virtual cell model trained on 25.6 million genetically perturbed single-cell transcriptomes across seven diverse cellular contexts – the largest genome-wide perturbation dataset ever reported, and a signal of intent off the back of a $1b Series A funding round announced in 2024. Tahoe Bio has open-sourced Tahoe-100M, a single-cell atlas fifty times larger than all previously available public chemical perturbation data combined. These are not incremental steps. The data challenge – historically the central bottleneck in building predictive models of cellular biology – is being addressed at genuine scale.

The critical questions relevant for near-term strategy are one of translation and one of timeline. To move the clinical needle, even exceptionally data-rich models must generate targets and hypotheses that hold up through a chain of translation. Not just in the cell lines or disease models in which they were trained, but in the heterogeneous biology of real, diverse patients. Expanding from controlled cellular contexts to the multiplicative complexity of primary cells, post-translational modifications, splice variants, cell-cell communication and multiple patient populations remains a daunting technical challenge for the field.

How much more data and compute will be enough to produce better clinical insights? The scale of data required for truly generalisable models of biology is an unknown. And omics data has always had the challenge of data quality issues to contend with as well.

For those of us without billion dollar seed rounds and vast computing resources, the most tractable path may be not to attempt everything everywhere all at once, but to focus on domains  – specific regulatory layers, defined data-rich disease mechanisms, tightly characterised cellular systems – where current, cutting-edge methods (including AI) combine with deep expertise to extract meaningful insight. Coupled with closed loop iterative learning (lab-in-the-loop) this should translate into well-designed clinical hypotheses with a clear route to market.

The democratisation cycle

There is a structural dynamic in play that changes the competitive calculus for smaller companies considerably.

When large, well-resourced organisations invest heavily in AI drug discovery – the kind of investment that generates hundred-million-cell atlases, billion-parameter models, or decades of proprietary chemistry data – the tools and datasets built on that investment tend, over time, to become accessible to the broader community. The cycle of bold investment, model development and democratisation means that the gap in computational capability between large and small is not fixed. It narrows.

The question for focused companies is how to exploit that investment intelligently: to use the ocean that others have boiled to make a very precise cup of tea.

We see this playing out directly. For example, Eli Lilly’s TuneLab programme gives selected biotechs access to AI and machine learning capabilities built on over $1 billion of Lilly’s internal research investment, covering predictive ADME and chemical profiling across hundreds of thousands of unique molecules. Using federated learning, participating companies can run Lilly’s models on their proprietary data, and contribute data back to the platform in return. Alongside open initiatives like AlphaFold3, Tahoe-100M and the UK’s OpenTargets and OpenBind consortia, this represents a new model of open innovation to boost the entire drug discovery ecosystem: world-class infrastructure is being made available to focused teams who bring innovative thinking, deep biological expertise and novel chemistry, and who stand to make an oversized contribution to our progress in fighting disease.

This is borne out by recent shifts in the source of therapeutic innovation. Big pharma portfolios have shifted from assets primarily developed internally to assets brought in through acquisitions or licensing, with 70% of revenues from new molecular entities since 2018 coming from externally sourced innovation. This reflects the depth of scientific innovation that occurs within small biotechs. Smaller, focused techbio companies that can exploit the latest gigascale datasets and AI models in a task-specific manner stand to reinforce this trend.

From more to better 

Ultimately, what matters most is not how many potential therapeutics reach the clinic, but how many emerge from it and meaningfully improve patients’ lives. That demands new approaches to uncovering mechanisms with strong causal links to disease, understanding how perturbing a given biological node reverberates through the wider network, and locking down robust biomarkers early that can confirm target engagement and track disease modulation in trials.

We are on the cusp of a revolution in drug discovery, powered by unprecedented investment in data and computational power. At the forefront are a handful of companies with the capital and compute to push the boundaries, even as the path remains non‑linear and clinical validation remains the ultimate test. For the rest of us, we have an equally vital opportunity: to use the increasingly available datasets to contextualise and explore biological hypotheses, to build or adapt AI models that yield focused insights into disease, and to exploit cutting‑edge optimisation models to move the right compounds forward faster.

In this new world, AI‑enabled, focused biotechs can continue to punch above their weight among the billion‑dollar titans of tech and pharma, generating innovative, translatable solutions to complex diseases grounded in deep biological understanding. Progress will not come from scale alone, but from concentrated innovation that bucks Eroom’s Law through capital efficiency and a singular expert focus. As broad, horizontal efforts attempt to map everything, there is room for more focused approaches that can both contribute to and benefit from the shared infrastructure they create.

The opportunity now is not to model everything everywhere all at once, but to extract high‑resolution insights from the right questions and data and convert these into hypotheses worth testing in the clinic. If we get that right, AI will do more than accelerate drug discovery; it will increase the probability that the drugs we discover actually work.

Peter Hamley