Descriptions:
Latent Space Science hosts RJ Haniki and Brandon Anderson sit down with Ron Alfa (co-founder and CEO) and Daniel Bear (VP of AI) of Noetik, a biotech company applying self-supervised transformer models to one of medicine’s most persistent problems: the 90–95% failure rate of cancer drug trials. Their thesis is that most failures stem not from poor pharmacology or target selection, but from inadequate patient stratification—giving the right drug to the wrong population.
Noetik’s platform trains multimodal models on H&E staining images—the standard pathology format present for virtually every patient in any clinical trial—alongside proprietary tumor sample data generated in-house from a custom processing pipeline built from scratch. At inference time, only the H&E image is required, making the system retroactively applicable to historical trial data. Their Octo Virtual Cell model goes further, simulating what would happen to a patient’s tumor microenvironment if a specific gene or protein were knocked down, enabling in-silico hypothesis testing before committing to a clinical study.
Bear and Alfa describe two primary use modes: reverse translation (starting from patient data to identify novel drug targets) and prospective patient stratification for ongoing Phase 2 and Phase 3 trials. The conversation covers how the models cluster patients into biological subpopulations that cut across traditional cancer type classifications, the challenge of building training datasets before any model could be trained, and how re-analyzing responder versus non-responder cohorts from past trials can surface hypotheses for more targeted follow-on study designs.
📺 Source: Latent Space · Published April 20, 2026
🏷️ Format: Interview
