Generative Models for Therapeutic Peptide Optimization​

1:10-1:40 PM ET

The therapeutic potential of peptides can be further unfolded with the help of artificial intelligence. Non-natural amino acids (NNAAs), amino acids not encoded by DNA, have been used to enhance many peptide properties from binding affinity, protease and plasma stability to half-life. Incorporating NNAAs has enabled stepping beyond the chemical space dictated by the 20 natural amino acids and expand it to a small molecule-like chemical space.

In this work, we present a novel generative AI-based tool, PepINVENT as an extension to the small molecule molecular design platform, REINVENT [1].  Our tool aims to facilitate de novo design of peptides with enhanced properties through reinforcement learning driven generation. With its chemistry-aware generative capabilities, PepINVENT explores chemical spaces beyond those of enumerated amino acid libraries. It produces novel, diverse and chemically feasible amino acids with chirality mutations, backbone modifications, and sidechain alterations.

In addition to the novel amino acid designs, our generative framework offers structural flexibility by exploring diverse peptide topologies or constraining the generation to a specific topology. We have developed a range of learning objectives that are specific for multi-parameter optimization of properties on both amino acid- and peptide-level and applied it to several therapeutically relevant peptides. To enable this work in the laboratory, we have also developed tools for chem-informatically protect any novel amino-acid to be used in solid-phase synthesis and to score the synthesizability of the so obtained molecule by a suitable retrosynthesis score.​

​