Generation of chemically modified peptides using peptide folding information
Peptide therapeutics are a fast-growing modality that remains hard to design, because activity depends jointly on sequence, local chemistry, folding, and developability constraints such as toxicity, selectivity, and synthesizability. Existing methods sit at two extremes that each fall short, so this project proposes a unified, folding-aware framework that couples sequence optimization over natural amino acids with residue-level optimization over a curated library of synthesizable modifications. It uses folding predictions, conformer-derived descriptors, and function scoring as active signals in a hierarchical discrete generative search, building on prior results showing that efficient representations can rival heavier architectures.
- Bridges sequence-level and atom-level methods
- Uses folding and function scores as active guidance
- Generative AI based on hierarchical discrete optimization
