#31 · ML / AI methods

Generative models for biomolecules gain structure-conditioned and dynamics-aware control

The dominant thread is generative modelling conditioned on structural, physical or interaction context. DPDiff conditions protein-specific 3D ligand generation on geometry- and sequence-based interaction priors, and MolPIF unifies continuous coordinates with discrete atom types via parameter-space interpolation flow. Boltzmann generators target equilibrium sampling of molecules and materials as an alternative to molecular dynamics. GeoCoupling learns temporal couplings between sequence and structure in multimodal co-design, and VelocityFM applies rectified flow matching in velocity space to short-horizon protein trajectory prediction. A masked Flood complex yields persistent-homology descriptors of MD-sampled structures, with MSMs guiding marsfm generative training. What appears new is the extension of generators toward dynamics-aware and cross-modal tasks, rather than static structure generation alone.