#35 · Structural bio & protein design

Context-conditioned generators design molecules, peptides, and cell morphology profiles

The common thread is conditioning generative models on explicit biological or chemical context rather than unguided sampling. FuncBind uses neural fields and score-based generative models to produce target-conditioned all-atom molecules, including small molecules. POTFlow couples a lead-conditioned flow-matching model with optimal transport and secondary-structure priors to generate therapeutic peptides targeting ATP5A in glioblastoma. HIL-DD embeds expert feedback in an equivariant rectified flow model for faster 3D molecule generation. TRIDENT conditions cell morphology synthesis on compound and L1000 RNA profiles. What appears new is the explicit coupling of generators to pocket, lead, prior, or expert-feedback information within the design loop. The set is heterogeneous in modality, spanning small molecules, peptides, and cell morphology, so the shared element is methodological rather than system-specific.