Advancing cancer vaccine design using deep learning to predict peptide-HLA class I binding and immunogenicity

Melanoma is an aggressive form of skin cancer with high inter- and intratumor heterogeneity. Despite advances in immune checkpoint inhibitor therapies, patients fail to achieve long-term remission. Cancer vaccines offer a promising, personalized approach by targeting neoantigens, tumor-specific peptides that bind patient Human Leukocyte Antigen (HLA) types. To predict peptide-HLA (pHLA) binding, the Hacohen lab and other groups have previously used various machine learning architectures, including neural networks. However, current models are constrained by poor predictive power, shallow architectures, lack of consideration for protein structure, or sparse datasets. This proposal aims to advance cancer vaccine design through the development of deep learning models that accurately predict pHLA binding and immunogenicity, with a strong focus on efficacy in melanoma. In Aim 1, I will improve training dataset quality using active learning, the method of identifying peptides with the highest model uncertainty and prioritizing them for experimental validation. To design a peptide library for active learning, I will use a two-pronged approach to identify putative false positives, putative false negatives, and universally poorly classified peptides. In collaboration with the Wu and Carr labs, these peptides will be validated as binders or non-binders experimentally using a novel high- throughput E. coli-based user-defined pHLA production assay. Using the updated dataset, I will then develop a new pHLA binding predictor by finetuning Evolutionary Scale Modeling 2 (ESM2), a large language model trained on protein sequences rich in evolutionary and structure information, to the specific task of pHLA binding. The model will be validated using melanoma patient data under the Clinical Proteomics Tumor Analysis Consortium with the Carr lab. Given that most neoepitopes, even those that bind HLA, do not lead to T cell activation (i.e. immunogenicity), in Aim 2, I will extend this work to predict pHLA immunogenicity. Prior work has shown that pHLA complex stability is correlated to immunogenicity, but datasets on pHLA stability are sparse. With assistance from the Hacohen, Carr, and Wu labs, a separate high-throughput E.coli-based assay will be used to generate the largest known pHLA thermal stability dataset to date, with 5,000 peptides across 15 HLA-Class I alleles. Using this data, I will train a generative diffusion model to produce peptide sequences that have high predicted pHLA complex stability. This model will be experimentally validated by the Hacohen, Carr, and Wu labs in melanoma using known pHLA-T cell receptor pairs such as that of MART1, melanoma antigen recognized by T cells 1, across a gradient of temperatures. Put together, this integrative approach between computational model generation and directed data generation with novel high-throughput assays can transform the performance of current pHLA binding and immunogenicity models, with broad-ranging impact for improving therapeutic design of cancer vaccines in melanoma. The fellowship will be done in a collaborative research environment focused on interdisciplinary mentorship, and professional and academic development. Project Number: 1F30CA310069-01 | Fiscal Year: 2026 | NIH Institute/Center: National Cancer Institute (NCI) | Principal Investigator: Jessika Baral | Institution: HARVARD MEDICAL SCHOOL, BOSTON, MA | Award Amount: $43,056 | Activity Code: F30 | Study Section: Special Emphasis Panel[ZRG1 F09C-H (22)] View on NIH RePORTER: https://reporter.nih.gov/project-details/11315297