#21 · Structural bio & protein design

Generative design of biomolecules and molecules paired with explicit feasibility scoring

The dominant thread is generative proposal of biological sequences and molecules coupled to computational scoring. MOZAIC grows drug-like fragments through in silico reaction rules and Conformational Space Annealing, optimizing QED, synthetic accessibility, and binding affinity. SynFrag scores synthetic accessibility with a fragment-assembly autoregressive model, while SepTop free-energy benchmarking on Cyclophilin D addresses affinity prediction for fragment merging and linking. InversePep applies diffusion-based inverse folding to generate peptide sequences for target backbones, and InstructNA designs DNA aptamers de novo from HT-SELEX-trained nucleic acid language models without structural input. What is new is the pairing of generation with explicit scoring of feasibility and performance. The items differ in maturity and modality, so the shared pattern is methodological rather than a common biological target.