#30 · ML / AI methods

Deep models for single-cell perturbation, annotation, and foundation-model interpretation

The dominant thread is learned models applied to single-cell data under heterogeneous conditions, including unseen perturbations, heterogeneous batch compositions, and out-of-distribution cell types. MapPFN uses in-context learning, pre-trained on synthetic causal biological priors, to predict single-cell perturbation effects. scBatchProx refines precomputed single-cell embeddings with federated-inspired batch-client optimization to stabilize cell-type discriminability across heterogeneous batch compositions. A hierarchical cross-entropy loss for scRNA-seq cell type annotation improves out-of-distribution performance by 12-15% across models. scGeneLens dissects how single-cell foundation models represent cells, combining sparse block attention, attention propagation, and integrated gradients. What is new is the pairing of predictive modeling with explicit interpretability of foundation models. The longitudinal EHR world model is a distinct clinical domain, included as a related use of latent trajectory prediction rather than single-cell data.