Atlas Discovery launches AI models to predict patient response in clinical trials

Atlas Discovery launched out of stealth with foundation models that predict individual patient drug responses in clinical trials, targeting the costly gap between preclinical results and human efficacy. Y Combinator general partner Ankit Gupta, who co-founded Reverie Labs, said the models achieved an AUROC of 0.76 in a retrospective analysis of a phase 3 ustekinumab trial for ulcerative colitis, potentially reducing patient enrollment by roughly 450 while maintaining statistical power. The company is seeking research partnerships with hospitals and biopharma teams.

Atlas Discovery has launched out of stealth with foundation models designed to predict how individual patients will respond to drugs in clinical trials, according to an article published on X https://x.com/i/article/2080707729870430209 by Y Combinator general partner Ankit Gupta. The founders - Shaamil Karim, Sanjukta, and Christian - are targeting clinical translation, the costly gap between promising preclinical results and evidence that a drug works in humans. Gupta, who previously co-founded drug discovery startup Reverie Labs, said Atlas plans to use its predictions for drug repurposing, rescuing failed drug candidates, selecting patients for trials, and evaluating assets for in-licensing. Atlas trains its models on a combination of abundant preclinical data and scarcer clinical data. The company says this approach lets the models learn broad biological patterns before adapting them to predict human drug response. In retrospective analyses of cancer and autoimmune clinical trials, Atlas reported that its models distinguished responders from non-responders with high accuracy. In one case study, Atlas analyzed baseline biopsies from a phase 3 trial of ustekinumab in ulcerative colitis. The company said its model predicted treatment response before dosing with an AUROC of 0.76. Its retrospective analysis suggested the trial could have retained the same statistical power with roughly 450 fewer patients. The result does not establish how the model would perform prospectively in a live trial, but it illustrates the economic and operational case Atlas is making to drug developers. Gupta also pointed to a public preprint in which the team's discrete-latent model achieved what he described as state-of-the-art results across the tested distributional metrics for single-cell perturbation prediction. The more consequential claim is that the same modeling approach transferred from preclinical biology to human drug response, a persistent challenge for AI drug discovery companies whose predictions often depend on cells or animal models. Karim studied computer science at Dartmouth and worked on applied AI at Etched and RunPod, according to Gupta. Sanjukta conducted research at Los Alamos and left a PhD program, while Christian studied mathematics and biology at Dartmouth and researched theoretical biology at Johns Hopkins. Gupta said the team has presented its research at ICLR, Cold Spring Harbor Laboratory, and ICML. Atlas is seeking research and data partnerships with hospitals and health systems, as well as relationships with biopharma teams running clinical trials or developing drugs. More information is available on the Atlas Discovery website https://atlasdiscovery.bio/ and in its clinical trial response prediction case study https://atlasdiscovery.bio/clinical-trial-response-prediction . The company did not disclose funding, customers, prospective validation results, or commercial terms in the announcement.