#19 · ML / AI methods

Generative models steer biomolecular design via alignment, energy biasing, and sampling

The dominant thread is generative or learned sampling of biomolecular structure and sequence, and the newer element is explicit control over that generation. Metadiffusion adds inference-time meta-energy biasing to pretrained biomolecular diffusion models to produce diverse conformational ensembles. A discrete protein diffusion study of antibody design analyzes DPO alignment, reporting that structure aligns early while properties improve slowly, along with collapse mechanisms. Reading TEA leaves scores Metropolis sampling with a 20-letter alphabet derived from protein language model embeddings. Cross-Chirality Generalization injects axial features into E(3)-equivariant vectors so L-L training transfers to latent-diffusion D-peptide design. Structure prediction work (IntelliFold-2, NbForge) and PRIZM's low-N selection of zero-shot models sit adjacent to this thread. CTMD metadynamics addresses docking-hit enrichment through physics-based sampling rather than generative modeling.