In progress · NSF Engineering Research Initiation (ERI)
Award #2553227 · $199,868
May 1, 2026–April 30, 2028 (estimated end)
Principal Investigator: Jesus A Beltran Verdugo
Antimicrobial peptides (AMPs) offer a promising alternative to conventional antibiotics due to their broad-spectrum activity and low risk of resistance. However, discovery of novel AMPs remains challenging due to the vast peptide sequence space, the expense of experimental screening, and incomplete understanding of the structural and physicochemical determinants of activity. Computational approaches have emerged to accelerate design and discovery. Despite rapid and substantial progress in this area, current methods still face critical drawbacks, including labeling noise, severe data imbalance, and limited peptide representation.
This project proposes a novel multimodal peptide language model (PLM) that integrates sequence, structural, and physicochemical features into a unified embedding space. By combining complementary modalities, the framework will capture the biochemical and structural properties underlying antimicrobial activity while remaining robust to noisy or incomplete biological data. To further address labeling challenges, this project combines PLM with Positive-Unlabeled (PU) learning, enabling the calibrated detection of hidden positives among peptides currently annotated as untested or non-antimicrobial in public repositories such as UniProt. This multimodal PU learning framework will provide a scalable, interpretable, and reproducible approach for identifying new AMPs from previously unexplored regions of peptide space.
Jackson Copeland: Developing a reliable benchmark for training and evaluating AI models, addressing uncertain activity labels and class imbalance.
Damarrion Morgan-Harper: Investigating peptide sequence representations and positive–unlabeled learning for AMP classification; first author of the accepted IEEE BIBM 2026 workshop paper below.
Mark Raspopov: Investigating structural representations of peptides as part of the work toward combining sequence, physicochemical, and structural information.
Alejandro Lopez: Developing PepSpace, an interactive tool for exploring peptide embeddings and activity information in 2D or 3D; coauthor of the accepted AMP classification paper.
Morgan-Harper, D. L., Lopez, A., Copeland, J., & Beltran, J. A. (2026). Antimicrobial peptide classification using peptide language model embeddings and positive-unlabeled learning. Accepted workshop paper, 2026 IEEE International Conference on Bioinformatics and Biomedicine (BIBM).
The study combines frozen ProtBERT-BFD sequence embeddings with positive–unlabeled learning to classify antimicrobial peptides without treating every untested sequence as inactive.
Lopez, A., & Beltran Verdugo, J. A. (2026, September 17–19). PepSpace: Interactive visual analytics of peptide embedding spaces for antimicrobial discovery. Poster presentation, 38th Annual Great Minds in STEM Conference, Albuquerque, NM.
Related alumni research: Mohsen Alavian’s 2026 master’s thesis, AI Methods for Antimicrobial Peptide Analysis, explored multilabel peptide activity classification.