AI in Drug Discovery: What the Evidence Actually Shows
AI-designed molecules have entered clinical trials. AI-predicted protein structures have transformed structural biology. But the evidence for AI's impact on drug discovery success rates is more complicated than the headlines suggest.
The claims around AI in drug discovery have been extraordinary. Compressed timelines from target to candidate. Molecules designed de novo with predicted properties. Protein structures solved in hours that would have taken years by X-ray crystallography. Drug repurposing opportunities identified in databases that no human team could have searched systematically.
Some of these claims are real, and the progress in specific areas has been genuinely impressive. Some are premature, reflecting pipeline-stage optimism that has not yet met the test of clinical outcomes. The difficulty, for anyone trying to assess what AI actually contributes to drug discovery, is that the signal and the noise are currently mixed in a ratio that makes clear-headed evaluation difficult.
This matters because decisions are being made on the basis of this assessment: investment decisions, partnership decisions, regulatory strategy decisions, and build-versus-buy decisions at pharmaceutical and biotech companies. Getting the picture wrong in either direction, overestimating or underestimating AI's contribution, has real consequences for how research resources are allocated.
What AlphaFold actually changed
The release of AlphaFold 2 by DeepMind in 2021, followed by the open-source release of the model and its predictions for essentially the entire human proteome, was a genuine scientific breakthrough. The ability to predict protein structures from sequence with accuracy approaching experimental methods transformed structural biology. The AlphaFold Protein Structure Database now holds predicted structures for over 200 million proteins across nearly all known organisms.
For drug discovery, the implications are real but specific. Protein structure is not a drug target. It is an entry point for target validation and for structure-based drug design once a target has been identified and validated. AlphaFold accelerates the structural biology phase of discovery for targets where obtaining experimental structures is difficult or time-consuming. It does not accelerate target identification, target validation, or the clinical phases where most drug development failures occur.
The structural biology phase is important. It is not the bottleneck. The bottleneck in drug discovery has always been, and remains, understanding which biological targets, when modulated in which ways, in which patient populations, produce therapeutic benefit with acceptable safety. AlphaFold does not address that question directly, and claims that it has transformed the drug discovery success rate should be evaluated with that distinction in mind.
De novo molecule design: the gap between in silico and in vivo
Generative AI models for molecule design, systems that can propose novel chemical structures with predicted properties, have attracted enormous investment and generated genuine excitement in the medicinal chemistry community. The technology is real, and its outputs are being taken into synthesis and biological testing at a growing number of companies.
What is harder to assess is whether AI-generated molecules are more likely to succeed in development than molecules generated by traditional medicinal chemistry. The comparison is not straightforward. Most AI-discovered drug candidates are still in early clinical development, and the time horizon for evaluating clinical outcomes is measured in years and decades, not the months since these programmes were announced.
The first AI-designed drug to enter clinical trials, reported with considerable fanfare, went into Phase I in 2020. As of early 2024, the evidence for superior clinical outcomes from AI-designed molecules relative to conventionally designed ones remains essentially absent, not because the approach is wrong, but because it is too early to know. The history of drug development is full of approaches that were promising in preclinical stages and failed when they met the complexity of human biology in clinical trials.
Where AI adds well-documented value
The evidence is strongest for AI contributions in areas where the task can be evaluated computationally without waiting for clinical outcomes.
ADMET prediction, estimating a molecule's absorption, distribution, metabolism, excretion, and toxicity properties from its structure, is an area where machine learning models have demonstrably improved in predictive accuracy over the past decade. Compounds that fail in development due to poor ADMET properties are a well-characterised source of waste in pharmaceutical development. Models that can predict these properties more accurately at earlier stages allow chemists to deprioritise problematic compounds sooner, which is valuable even if it cannot be directly translated into a clinical success rate improvement without much longer time horizons.
Hit identification from virtual screening is another area with evidence-backed impact. AI-enhanced docking and virtual screening methods can screen large compound libraries computationally with better enrichment rates than traditional approaches, meaning more of the compounds progressed to biological testing are actually active. The efficiency gain is real and measurable at the hit identification stage.
Patient stratification and biomarker identification using clinical and genomic data are areas where AI applications are closer to clinical validation. Several FDA-approved companion diagnostics and biomarker-driven trial enrichment strategies have been developed using machine learning on large genomic datasets. The EMA's qualification procedure for novel methodologies provides a pathway for formally validating AI-derived biomarkers before pivotal trials begin. The translational path from computational analysis to clinical outcome is shorter in this domain than in the molecule design space.
The honest state of the evidence
AI has not yet materially changed the attrition rate in drug development. The probability that a compound entering Phase I will receive regulatory approval has not improved significantly in the past decade despite the substantial investment in AI tools across the pipeline. This is not a condemnation of the technology, it is a recognition that the problems driving attrition, unpredictable toxicity in humans, lack of clinical efficacy in broader populations than the studies suggest, and complexity of disease mechanisms that preclinical models do not capture, are not problems that better computational tools at the front end of discovery can directly address.
What is changing is the quality and speed of the decisions made at the early stages of discovery. Better structural information, more accurate ADMET predictions, more comprehensive virtual screening, and faster iteration on chemical series are genuine contributions. Whether they translate into better clinical outcomes over the timescales where drug development is evaluated will be the question of the next decade.
For research teams synthesising the literature in this area, the methodological challenge is evaluating a rapidly evolving body of evidence where many of the most-cited claims are from company press releases, preprints, or publications from groups with direct commercial interests. A more complete and critically structured evidence map, one that distinguishes between what has been demonstrated computationally and what has been validated clinically, is genuinely useful and genuinely difficult to assemble manually. It is the kind of problem we help research and commercial teams work through at NousLab. Get in touch if you are navigating this space.
