Seminar, Hosted by Kempner Institute, Harvard University
Adaptive AI for cell state biology
Helmholtz Munich
Abstract
Single-cell atlases keep growing and shifting, and models of cell state have to adapt without forgetting what they learned. This talk covers adaptive representation learning for single-cell data under distribution shift: catastrophic forgetting during incremental atlas integration, a biology-informed elastic weight consolidation method with Bregman-information replay, test-time adaptation, and reference-guided transition models that connect observational and interventional atlases. It then frames closed-loop hit discovery as sequential decision-making over perturbations, pairing an in-context, action-conditioned transition model of cellular state under interventions with region-based Bayesian optimisation and a TabPFN surrogate to choose the next experiment. The talk closes with extensions towards closed-loop world models, long-horizon planning over sequences of interventions, and learning decision policies from logged discovery campaigns, with patient-relevant outcomes such as CRISPR trials as a target application.
Topics
- Single-cell atlases, search World Wide
- Bayesian optimisation, broader topic: optimisation
- adaptive representation learning, broader topic: representation learning
- single-cell data
- distribution shift
- catastrophic forgetting
- incremental atlas integration, broader topic: integration
- elastic weight consolidation, broader topic: consolidation
- Bregman-information replay, broader topic: replay
- test-time adaptation, broader topic: adaptation
- reference-guided transition models, search World Wide
- region-based Bayesian optimisation, broader topic: optimisation