Abstract Background: The advent of immune checkpoint inhibitors has improved morbidity and mortality for some cancers, and recent breakthroughs in gene & cell therapy have shed light on curing some types of blood cancers. However, many cancers remain intractable and the development of novel, effective and safe therapies continue to be a priority. Cancer vaccines as a cancer immunotherapy approach has seen a resurgence in recent years, due to the success of mRNA vaccines for the COVID-19. However, the accurate prediction of immunogenicity of cancer vaccines remains elusive. Methods: Our models predict the probability of a given peptide derived from the protein of interest to be presented by MHC-I or MHC-II. For MHC-I antigen presentation model development, over 17 million entries in the dataset were collected from published literature and available databases, e.g., IEDB, with peptide lengths ranging from 8 to 11. The peptides were restricted to 150 unique MHC-I alleles. Similarly, ~4 million entries with peptide lengths ranging from 13 to 21 were collected for MHC-II antigen presentation model development, and the peptides were restricted to 19 unique MHC-II alleles. To develop advanced antigen presentation models, a language model was chosen as the backbone network and contrast learning was used to better discriminate the peptide-MHC match versus mismatch. Overall, both MHC-I and MHC-II presentation models were constructed with about 30 million parameters. To validate the model prediction accuracy, automated peptide synthesis and surface plasmon resonance (SPR) technologies were applied. Results: Using open-sourced data, our developed AI models surpassed the performance of state-of-the-art prediction algorithms, the latest versions of NetMHCpan and MixMHCpred, for both MHC-I and MHC-II antigen presentation. Furthermore, to validate the algorithm accuracy and the peptide immunogenicity, 28 predicted patentable peptides derived from mutated TP53 protein were synthesized and their binding to respective common HLA alleles were validated using SPR. We found that greater than 80% of the peptides display binding affinities that are stronger than the positive control, suggesting that AI significantly improves neoantigen peptide vaccine design. Conclusions: We developed advanced AI algorithms to rapidly design shared neoantigen T cell epitopes with predicted strong binding affinity to MHC-I and MHC-II. We envision that the epitopes predicted and designed by our AI algorithms possess great potential in advancing the field of off-the-shelf cancer vaccine development and hold the promise of significantly benefiting patients, once translated into the clinic. Citation Format: Genwei Zhang, Jiewen Du, Xiangrui Gao, Tianyuan Wang, Zhenghui Wang, Qingxia Zhang, Tongren Liu, Dong Chen, Ruohan Zhu, Yalong Zhao, Chi Han Samson Li, Melvin Toh, Lipeng Lai. Towards the efficient design of shared neoantigen peptide cancer vaccines using artificial intelligence [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 3525.
Read full abstract