Deep Learning Guided Discovery of Synthetically Complex Peptides using Hierarchical Graph Networks

Camille Bilodeau
UVA Chemical Engineering

Time: 2026-09-09, 12:00 - 13:00 ET
Location: Rice 540 and Zoom

Abstract Advances in synthetic chemistry have enabled a new generation of complex peptide therapeutics incorporating modifications such as non-natural amino acids, backbone cyclization, polymer units, and D-amino acids. While these modifications can improve stability, bioavailability, and target specificity, they dramatically expand the peptide design space. Machine learning offers a promising approach for navigating this complexity, but most existing models represent peptides as linear sequences of the 20 canonical amino acids and therefore cannot readily accommodate non-natural residues or complex architectures.

To address this challenge, we developed two complementary tools for machine learning on synthetically complex peptides. PepMNet [1] is a multi-scale graph neural network that represents peptides at both atomic and residue levels, capturing detailed chemical features while preserving higher-level structural information. This hierarchical representation enables non-natural residues and complex topologies to be encoded directly. Across benchmarking tasks including antimicrobial activity and chromatographic retention, PepMNet outperforms sequence-based approaches, particularly for poorly sampled non-natural residues. PepFoundry [2] complements this framework by converting synthetically modified peptide sequences into standardized, machine learning-ready graph representations. As a proof of concept, we applied these tools to antimicrobial peptide discovery in collaboration with the Hughes Lab at the University of Virginia. We curated a dataset of complex peptides containing non-natural side chains, polymer units, and D-amino acids with measured activity against Klebsiella pneumoniae, a WHO priority pathogen [3]. PepMNet was pretrained on large public datasets and fine-tuned on this targeted dataset, then coupled with a genetic algorithm to explore chemical space and identify novel candidates predicted to exhibit strong antimicrobial activity. Top candidates were synthesized and experimentally evaluated using minimum inhibitory concentration assays. Together, PepMNet and PepFoundry provide an integrated framework for representing, predicting, and optimizing synthetically complex peptides, enabling data-driven exploration of chemical spaces inaccessible to conventional sequence-based models.

[1] Garzon Otero, D.; Akbari, O.; Bilodeau, C. Mol. Syst. Des. Eng. 2025, 10, 205–218. [2] Garzon Otero, D.; Akbari, O.; Mandapati, A.; Bilodeau, C. J. Chem. Inf. Model. 2026. [3] WHO bacterial priority pathogens list, 2024.

Bio: Dr. Bilodeau received her B.S. and M.S. from Northwestern University and her Ph.D. from Rensselaer Polytechnic Institute, both in Chemical and Biological Engineering. During her Ph.D., she received the Lawrence Livermore Advanced Simulations and Computation Graduate Fellowship, through which she carried out research at Lawrence Livermore National Laboratory. Dr. Bilodeau was a postdoctoral researcher studying machine learning for molecular design in Klavs Jensen’s group in the Chemical Engineering department at MIT. She joined the University of Virginia as an assistant professor in Chemical Engineering and began her position in January 2023.

Molecular simulations are powerful tools that can predict important physical properties while simultaneously yielding a full molecular picture of the system. In the past decade, the advent of GPU computing has resulted in dramatic improvements in the computational speed of molecular simulations, making it possible to study larger, more complex systems. At the same time, the field of deep learning has experienced a renaissance, with neural networks being successfully employed for reading text, classifying images, and even folding proteins. This convergence of increased hardware and algorithmic capabilities has set the stage for combining the wealth of data arising from rapid molecular simulations with the deep learning tools to mine this data.

In the Bilodeau group, we explore the intersection between molecular simulations, statistical physics, and artificial intelligence to develop tools to discover and design new molecules, surfaces, and proteins with optimized properties. Our core expertise lies in molecular dynamics simulations of soft matter systems and artificial intelligence for molecular property prediction and generation. This interdisciplinary toolset allows us to solve important problems in applications ranging from designing biotherapeutics to developing novel separation materials.