Deep learning pan‐specific model for interpretable MHC‐I peptide binding prediction with improved attention mechanism

Abstract

Accurate prediction of peptide binding affinity to the major histocompatibility complex (MHC) proteins has the potential to design better therapeutic vaccines. Previous work has shown that pan-specific prediction algorithms can achieve better prediction performance than other approaches. However, most of the top algorithms are neural networks based black box models. Here, we propose DeepAttentionPan, an improved pan-specific model, based on convolutional neural networks and attention mechanisms for more flexible, stable and interpretable MHC-I binding prediction.With the attention mechanism, our ensemble model consisting of 20 trained networks achieves high and more stabilized prediction performance. Extensive tests on IEDB's weekly benchmark dataset show that our method achieves state-of-the-art prediction performance on 21 test allele datasets. Analysis of the peptide positional attention weights learned by our model demonstrates its capability to capture critical binding positions of the peptides, which leads to mechanistic understanding of MHCpeptide binding with high alignment with experimentally verified results. Furthermore, we show that with transfer learning, our pan model can be fine-tuned for alleles with few samples to achieve additional performance improvement. Deep-AttentionPan is freely available as an open-source software at https://github.com/ jjin49/DeepAttentionPan.

Positional SHAP (PoSHAP) for Interpretation of Machine Learning Models Trained from Biological Sequences

“…This pattern is in congruence with the experimental data for the binding of A*11:01 (49). Jin et al (28) reported anchor sites for MHC alleles from attention-based models. PoSHAP analysis matched these anchor sites based on the PoSHAP heatmap (Figure 4A) and the range of the SHAP values per position (Supplemental Figure 5).…”

“…All three of these have significantly greater SHAP contributions when lysine is at position nine. This may reflect that there is some flexibility with the earlier root site when lysine is bound, and may demonstrate that the model had learned the length of the binding motif between the second position and the C-terminus (28) (Supplemental Table 2, Supplemental Figure 5).…”

“…Both classification and regression models are used to learn which peptides bind to each MHC allele, for example see O’Donnell et al , Zeng and Gifford, and Liu et al (25–27) However, because many reports forgo model interpretation, the learned biochemical relationships remain unknown. Other works determine relationships learned by their model, for instance in Jin et al (28), and Hu et al (29) used CNNs with an attention mechanism to determine the weights of the inputs on the final prediction. This method has been successful in recapitulating the experimentally defined binding motifs, but requires that the model be constructed with attention layers.…”

Positional SHAP (PoSHAP) for Interpretation of machine learning models trained from biological sequences

“…Both classification and regression models are used to learn which peptides bind to each MHC allele, for example see O'Donnell et al, Zeng and Gifford, and Liu et al [28][29][30] However, because many reports forgo model interpretation, the learned biochemical relationships remain unknown. Other works determine relationships learned by their model, for instance both Jin et al [31], and Hu et al [32] used CNNs with an attention mechanism to determine the weights of the inputs on the final prediction.…”

Predicting SARS-CoV-2 epitope-specific TCR recognition using pre-trained protein embeddings

“…A recent study [33] of protein language models has shown that the output attentions of BERT-based protein models can capture biologically relevant protein properties. An attention-based deep learning model for peptide-MHC-I binding predictions [34] has shown that the attentions learned by the predictive model can capture critical amino acid positions of the peptides, which help stabilize the peptide-MHC-I bindings.…”