#39 · Structural bio & protein design

Learned models generate and prioritize ligands conditioned on structure and synthesis

The shared element across these four preprints is conditioning learned models for ligand generation or prioritization on structural, target, or synthesis information, though the items differ substantially in task, target class, and architecture. PocketVE applies a pocket-conditioned variance-exploding diffusion model to 3D structure-based design and adds classifier-free property guidance. Fixed-Dimensional Latent Flow generates variable-size 3D molecules through two-stage latent flow matching with an autoregressive Transformer, without equivariant layers. OmniSyn unifies target-aware generation and H2L optimization within a synthesis-native LLM conditioned on protein sequence. MEL-Steered PharmacoNet adapts a pharmacophore prescreening model to individual GPCR targets using V-SYNTHES fragment-docking data. What is new is the combination of property guidance, geometric or synthesis constraints, and target adaptation across these approaches.