In the most recent episode of The Genetics Podcast, Patrick Short sat down with Dr. Dave Hallett, Chief Scientific Officer at Recursion. Dave has spent nearly three decades in drug development, from Merck through big pharma to the CRO Evotec before joining Recursion. He trained as a medicinal chemist and came up through neuroscience. Their conversation separates where AI really changes drug discovery from where the promises run ahead of the biology.
Dave is specific about where AI earns its place. It accelerates multi-parameter optimization, generative chemistry and active-learning loops, and clinical-trial simulation. Protein-structure prediction from tools like AlphaFold and Boltz has reset how quickly a target's shape can be resolved.
Computer vision now reads microscopy images at a scale no lab could review by hand. Recursion reports roughly an 80% reduction in the number of compounds made, and 50 to 70% time savings on the path to a development candidate.
The limits are just as concrete. Simulating whole-body human pathophysiology, idiosyncratic toxicity, and long-term organ effects stays beyond what the models can credibly predict. AI cannot shorten the parts of the process governed by physiology and regulation.
"Unfortunately, AI cannot change the laws of physics. If the FDA says, 'This is the endpoint you're looking at, and that endpoint is going to take four years to measure,' AI will not make that endpoint one year."
Models improve daily. The data underneath them does not improve on its own, and that is where programs go wrong.
Dave's guidance is "buyer beware": verify what a model tells you and ask where its training data came from. He points to a decade-old pharma analysis that found 70 to 80% of key biological results could not be reproduced across labs. Biology is noisy, so a single clean-looking output is really just a starting point.
Provenance is the practical test. Dave pushes teams to ask how a dataset was generated and whether it reflects real disease biology before trusting anything built on top of it. That scrutiny makes data curation part of the core science rather than a support function.
Recursion's core method is the perturbational map. Scientists systematically perturb cells with CRISPR and compounds, image the results, then embed them in high-dimensional space so that phenotypically similar genes and compounds cluster together. Across the platform, that produces roughly three trillion searchable gene and compound relationships.
The scale of a single map is hard to picture. For its neuroscience work, the team manufactured on the order of a trillion iPSC-derived neurons, knocked out 17,000 genes one at a time, and captured more than 30 million images.
Because a map is built once and then queried many times, it can surface genuinely novel neurodegeneration targets that a hypothesis-led search might never reach. A recent microglia-map milestone with Roche and Genentech shows the approach extending into partnered discovery. What keeps the output useful, in Dave's view, is the grounding.
Dave names the most common way drug programs collapse: picking a biological target that was never causally related to the disease. In his framing it is the single largest source of failure, and no downstream optimization can rescue a program built on the wrong target.
Recursion's answer is a four-step validation sequence: orthogonal confirmation of the finding, evidence of disease relevance, an honest tractability assessment, and third-party or partner validation. The through-line is repeatability. A result that shows up once is a hypothesis; a result that holds across independent methods is something you can build a program on.
The discipline extends into the clinic. Dave describes using AI to pressure-test inclusion and exclusion criteria that were often set decades ago, and to locate where eligible patients actually are. The cost of getting this wrong is measurable: about one in five trial sites fails to recruit a single patient, at roughly $40,000 per day to run a site. Applied to Recursion's PI3K program, simulation expanded the reachable patient population by about 20% and drove recruitment gains of 30 to 40%.
Dave's advice to scientists is to be "bilingual" across biology, chemistry, and computation while still building genuine depth in a domain. He frames the posture as being an optimistic skeptic: trust the tools, then verify what they tell you. That habit, more than any single model, is what turns computation into medicine.
Listen to the full episode below.