Maria-I: A Deep-Learning Approach for Accurate Prediction of MHC Class I Tumor Neoantigen Presentation

Background: Therapeutic cancer vaccines targeting neoantigens have shown promise in early phase clinical trials for inducing tumor-specific T-cell responses in diverse tumor types. Identification of the small number of optimal cancer vaccine targets is essential to limit cost and improve efficacy of patient-specific vaccine design. Therefore, there is a need to narrow neoantigen selection to those peptides with the highest potential to induce anti-tumor immune responses, and to maximize the associated clinical benefits. The presentation of these antigens by Major Histocompatibility Complex (MHC) class I is key for prioritizing candidates. Previous algorithms to predict antigen presentation have had limited success and have primarily been trained on in-vitro binding affinity assays, or are commercial platforms not widely available for use.

Methods: We previously developed MARIA (MHC Analysis with Recurrent Integrated Architecture) as a method for predicting MHC-II neoantigens [Chen B et al. 2017 ASH, and Chen et al. 2019, Nature Biotechnology, in press]; here we adapt and develop the same MARIA design for MHC-I and benchmark its performance. Specifically, we used deep-learning for predicting the likelihood of neoantigen presentation trained on presented lymphoma peptides as identified by mass-spectrometry (LC-MS/MS). MARIA-I was initially trained on >47,000 ligands we previously identified by mass-spectrometry through antigen presentation profiling of patients with mantle cell lymphomas [Khodadoust et al. 2017, Nature]. Recurrent neural network (RNN) layers within MARIA-I integrate 4 key features: peptide sequence-MHC key residues relationship, peptide-MHC binding, peptide cleavage pattern, and gene expression.

Results: We internally validated MARIA-I using 10-fold cross validation, where we observed an average AUC of 0.99 for the integrated model (Fig 1a); the peptide-MHC key residues relationships conferred the largest contribution to MARIA-I performance. We also further tested MARIA-I using an external dataset of ~19,000 diverse peptides from MHC-I antigen presentation profiles of four cancer cell lines (ACC1143, HCT116, HCC1937, and SUP-B15). MARIA-I maintains above 0.92 AUC across four diverse cancer types (AUC = 0.92, 0.92, 0.96, and 0.95, respectively). Since presentation of immunogenic peptides is essential for activation of T cell effector functions, we separately tested MARIA-I's ability to identify CD8 tumor infiltrating lymphocyte responses using corresponding data from 62 patients and 7587 neoantigens [Parkhurst et al. 2019, Cancer Discovery]. Despite MARIA-I not being trained on T cell reactivity data, neoantigens known to elicit positive T cell responses had significantly higher MARIA-I scores (Fig 1b, p=6.92e-5, Mann-Whitney U test).

Conclusion: MARIA-I enables accurate prediction of neoantigen presentation for MHC class I at scale. Given its generalizability, we expect MARIA-I to yield insights for development of therapeutic cancer vaccines as well as applications in transplantation and autoimmune pathology.

Figure 1. MARIA-I performance on antigen presentation and identifying immunogenic peptides. a) MARIA-I and sub-model predictors evaluated on 10% held-out validation set. Merging peptide sequence, gene expression, binding, and cleavage scores yielded improvements in performance as compared to each predictor's individual ability. b) MARIA-I assigns significantly higher scores to peptides known to elicit positive CD8 T cell response (p=6.92e-5, Mann-Whitney U test).

View largeDownload slide

View largeDownload slide

Close modal