Adaptive AI for cell state biology

This talk covers adaptive representation learning for single-cell cell states under shift: catastrophic forgetting in incremental atlas integration, a biology-informed elastic weight consolidation method with Bregman-information replay, test-time adaptation, and reference-guided transition models that link observational and interventional atlases. I will present closed-loop hit discovery as sequential decision-making over perturbations: an in-context, action-conditioned transition model—a dynamics model for rollouts of cellular state under interventions—paired with region-based Bayesian optimization and a TabPFN surrogate to select the next experiment. I will close with how this extends toward closed-loop world models, long-horizon planning over intervention sequences, and learning decision policies from logged discovery campaigns, with patient-relevant outcomes (e.g. CRISPR trials) as a target use case.