#2 · Structural bio & protein design

Generative models steered by RL, fragments, and SMARTS for molecule and linker design

The common thread is constraint-driven generation, in which molecular generators are conditioned on explicit chemical or structural objectives. ALCHIMIA couples RL-guided policy learning over 33 medicinal chemistry transformations with a genetic algorithm and docking for CB2R-directed design, while VeGA-RX and VeGA-SCX use SMARTS and scaffold guidance within SMILES-based generative transformers. LinkLlama fine-tunes Meta Llama 3 on ChEMBL drug-like molecules to produce fragment linkers from text prompts. Q-SFD recasts simultaneous fragment docking as a QUBO with an inter-fragment distance term, roughly doubling top-1 recovery of linkable pose pairs. A separate study evaluates LLM creativity in molecule generation across physicochemical, ADMET, and bioactivity tasks. What is new is the explicit encoding of chemical rules, including transformations, SMARTS patterns, and inter-fragment geometry, alongside learned sampling.