Deep Neural Networks Predict MHC-I Epitope Presentation and Transfer Learn Neoepitope Immunogenicity

Abstract

In this study, we present computational methods for predicting MHC-I epitope presentation with high accuracy and improved prediction of immunogenic neoepitopes. The BigMHC method comprises a large (51 million parameters), pan-allelic deep neural network trained on peptide-MHC presentation data from mass spectrometry assays and transfer learned on data from assays of antigen-specific immune response. Compared with NetMHCpan-4.1, PRIME, MHCflurry 2.0, MixMHCpred 2.1, MHCnuggets 2.3.2, and TransPHLA, BigMHC significantly improved the accuracy of predicting presented epitopes on a test set of 946,008 peptide-MHC complexes (AUROC=0.9838, AUPRC=0.9506). After transfer learning on immunogenicity data, BigMHC predicted immunogenic neoepitopes with the highest performance on independent test sets of neoepitopes validated for immunogenicity by IFN-γ release and MANAFEST assays. Moreover, BigMHC yielded significantly higher precision in identifying 215 immunogenic neoepitopes from a large (n=45,410,552) background of non-immunogenic epitopes, making BigMHC effective in clinical settings. All the data and codes are freely available at https://github.com/KarchinLab/bigmhc.

IMGT/RobustpMHC: Robust Training for class-I MHC Peptide Binding Prediction

“…To answer (1), we compare PerceiverpMHC with recent pseudo-sequence approaches and DeepAttentionPan (the single full sequence base approach ) on 6 datasets in section 1.1.In section 1.3 we show that the learned representations of RobustpMHC can be transferred to datasets that are even significantly different than the training dataset. Finally, we present ablation studies in section 1.4 to answer the research questions (2)(3)(4). With regard to (4), we show that across 8 different datasets RobustpMHC is either state-of-the-art approach or at par with state-of-the-art approach for that dataset in section 1.2.…”

“…We initially evaluate whether neural networks can inherently learn which amino acids are crucial for binding given the full MHC sequence or if we need to design hand-crated pseudo-sequences. To this end, we evaluate the performance of PerceiverpMHC on four publicly available benchmarks: independent and external set from Anthem [33] dataset, Neoantigen [10] and HPV [8] datasets and compare with the state-of-the-art pseudo-sequences based approaches like TransPHLA [10], capsNet [21], NetMHCpan 4.1 [44] and BigMHC [2]. The independent set from Anthem dataset [33] contains 112 types of HLA alleles, whereas the external set contains five HLA alleles.…”

“…Neoepitope dataset: For evaluating the transfer learning capability of our approach, we follow the BigMHC [2]. We use the training (positive = 1,407; negative = 4,778), and validation (positive = 173; negative = 515) split for transfer learning our projection block keeping the weights of Perceiver IO block frozen.…”

Discovering and overcoming the bias in neoantigen identification by unified machine learning models

“…Strikingly, the model achieved the best performance on both benchmarks when only the HLA allele was used. The performance of HLA-only model on IEDB benchmark had greatly surpassed (AUROC:+3.18% ∼ +15.95%, AUPRC:+5.13% ∼ +14.01%) all 15 models evaluated in Ref [26] (Fig. 3E).…”

“…We conducted a 10-fold cross-validation in consistency with Ref [22]. In section 4.2, to ensure the consistency and fairness necessary for bias exploration and ablation study, we adopted the same dataset curated by Ref [26].…”