(394g) Deep Learning-Guided Anti-Bacterial Peptide Design

Antibacterial peptides, a promising approach to treating a growing number of antibiotic-resistant diseases, often kill bacterial cells by binding to and altering the cell membrane. While antibacterial peptides typically leverage a combination of hydrophobic and electrostatic interactions to modulate peptide-cell membrane interactions, the role that peptide structure plays in governing potency is poorly understood. Neural networks offer a promising approach to learning this structure-function relationship. Unfortunately, publicly available datasets for these anti-bacterial activity are “data-poor” containing only 102-103 datapoints, limiting their utility in training neural networks.

In this work, we present a new transfer learning approach for anti-bacterial activity that involves first build a generalized, multi-task graph network to simultaneously predict peptide retention times on RPLC, HILIC, and IEX columns, tasks for which large (~105 datapoints) datasets are available publicly. We then finetune our model on a small (~103), publicly available antibacterial dataset. Finally, we apply our model to discover new antibacterial peptides by further refining our model on a small, internally generated, standardized dataset of point mutants. We demonstrate our model's ability to predict the outcome of new point mutations and apply our model to designing new candidate anti-bacterial peptides.