An Integrated 2D and 3D Graph Machine Learning Platform for Peptide Discovery

Peptides are short chains of amino acids that play essential roles in living systems and underpin a growing range of technologies, from new medicines and antibiotics to biomanufacturing and advanced biomaterials. The costly and time-consuming process of designing peptides or finding novel naturally occurring ones with useful properties remains slow. This project will develop computer-aided artificial intelligence (AI) tools that will first learn directly from the known molecular structure of peptides and their properties, and then help speed up the process of designing peptides with the desired predicted function. This will help researchers explore novel naturally occurring and synthetic peptides to identify promising candidates before committing to experiments. By accelerating peptide discovery, the work is expected to advance the design of next-generation biomaterials, for example, antimicrobial peptides. The project contributes to significant societal and economic importance and is central to the national priorities of artificial intelligence and biotechnology. The open-source software to be developed by the team is expected to lower barriers for researchers across disciplines. The project team plans to integrate new material into undergraduate and graduate courses and engage K–12 students through hands-on outreach activities designed to inspire the next generation of scientists and engineers. This project develops an integrated graph machine learning platform that captures both the two-dimensional (2D) topological and three-dimensional (3D) geometric structures of peptides to enable accurate prediction and design of peptide properties. The research is organized around three thrusts. The first thrust designs novel graph neural network architectures tailored to peptides, including amino acid-aware message passing that respects the hierarchical organization of atoms within amino acids, graph transformers that capture long-range interactions, and two-tower models that fuse graph representations with pretrained protein language models to combine structural and evolutionary information. The second thrust incorporates 3D structural information by constructing a new, large-scale dataset of peptide conformers through molecular dynamics-based sampling, developing geometry-aware graph neural networks that encode bond lengths, dihedral angles, and backbone curvature, and creating multimodal 2D-3D foundation models that can infer 3D-dependent properties even when only 2D information is available at test time. The third thrust will focus on predicting and validating chromatographic retention times relevant to peptide manufacturing and antibacterial activity to aid discovery. Through the development of an open-source software package, the project advances the state of the art in artificial intelligence for biology, generates reusable datasets and biotechnology tools for the research community, and provides a foundation for faster, lower-cost discovery of peptides with targeted functions. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.