Daphne Koller

AI Predicts Human Trial Outcomes, Boosting Drug Discovery Success

Much of the work on AI in biology aims to replace lengthy, expensive lab experiments with in silico predictions. AlphaFold provided a computational alternative to protein structure determination, and virtual cells aim to predict cellular responses to perturbations. By predicting results of lab experiments, these tools can narrow hypotheses and focus on experiments more likely to succeed. An even more profound transformation will be achieved when we can predict the outcomes not of lab experiments, but of experiments in a human – clinical trials – which are incredibly long, expensive, risky, and fail over 90% of the time. Even modest improvements in predicting results of interventions in humans would materially impact the dismal success rate of drug discovery. During insitro 's company update at the J.P. Morgan Healthcare Conference on Monday, I'll present our AI causal biology platform that aims to predict how human phenotypes respond to interventions at specific genes (inhibiting or activating a target). Our platform integrates genotype–phenotype associations in large human cohorts (experiments of nature) with large-scale genetic perturbation experiments in disease-relevant cellular systems. Human genetics reveals causality, yet selection, limited samples, and experimental constraints restrict mechanistic insight and effect estimation. Cellular perturbation data overcomes limits to deconvolute loci, resolve variant-to-gene ambiguity, prioritize among drug candidates, and reveal hidden genes influencing causal networks. But only if the cellular systems themselves are biologically relevant. At insitro, we’ve built a corpus of disease-relevant cellular systems – many far closer to human biology than immortalized lines, including iPSC-derived differentiated cell types. The right cell systems are key: our novel, now validated, ALS targets manifested only in our iPSC-derived motor neurons, and not in immortalized neuronal systems. We collect high-content, multimodal data (imaging, omics, and beyond) that our AI methods transform into disease-relevant predictions; these too were essential to our ALS discovery. We've industrialized these capabilities through automation, pooled single-cell screens with genome-wide perturbations (as in our recent publication), enabling rapid generation of multimodal datasets at scale. This approach has allowed us to identify targets – in MASH, ALS, dry AMD, and more – with potentially higher probability of clinical success. Many have been validated in functional readouts in translatable model systems, both in vitro and in vivo. We are now advancing these through drug discovery via our AI-enabled therapeutics platform, with our first program slated to enter the clinic this year. Predicting outcomes of in vitro experiments is undoubtedly valuable. Accurately predicting outcomes in humans is potentially transformative for drug development. I look forward to sharing more about insitro’s platforms and pipeline next week.