Transfer learning improves pMHC kinetic stability and immunogenicity predictions

Abstract

The cellular immune response comprises several processes, with the most notable ones being the binding of the peptide to the Major Histocompability Complex (MHC), the peptide-MHC (pMHC) presentation to the surface of the cell, and the recognition of the pMHC by the T-Cell Receptor. Identifying the most potent peptide targets for MHC binding, presentation and T-cell recognition is vital for developing peptide-based vaccines and T-cell-based immunotherapies. Data-driven tools that predict each of these steps have been developed, and the availability of mass spectrometry (MS) datasets has facilitated the development of accurate Machine Learning (ML) methods for class-I pMHC binding prediction. However, the accuracy of ML-based tools for pMHC kinetic stability prediction and peptide immunogenicity prediction is uncertain, as stability and immunogenicity datasets are not abundant. Here, we use transfer learning techniques to improve stability and immunogenicity predictions, by taking advantage of a large number of binding affinity and MS datasets. The resulting models, TLStab and TLImm, exhibit comparable or better performance than state-of-the-art approaches on different stability and immunogenicity test sets respectively. Our approach demonstrates the promise of learning from the task of peptide binding to improve predictions on downstream tasks. The source code of TLStab and TLImm is publicly available at https://github.com/KavrakiLab/TL-MHC.

Keywords: Machine learning; Peptide immunogenicity; Peptide kinetic stability; Peptide-MHC; Transfer learning.

Fig. 1.

TLStab & TLImm: A BA/EL predictor similar to NetMHCpan4.1 [10] is fine-tuned to stability/immunogenicity tasks. This is achieved by refining the MLP weights through task-specific training.

Fig. 2.

(A) Relationship between experimental ED50 values and stability values on the Ebola virus dataset and the Pox virus dataset. The y-axis depicts the mean stability values of peptides that have a better ED50 than the threshold (x-axis). (B) Relationship between BA predictions from two state-of-the-art tools (plus our pre-trained BA/EL predictor TLBind) and stability values. The y-axis depicts the mean stability values of peptides that have a better predicted BA than the threshold (x-axis). (C) NetMHCpan4.1 (p < 0.001), MHCFlurry2.0 (p < 0.01) and TLBind (p < 0.001) affinity predictions on immunogenic peptides are significantly different when compared to non-immunogenic ones.

Fig. 3.

(A) Pearson’s correlation and Kendall’s tau performance of TLStab against other knowledge transfer approaches on the unbiased 10-fold nested CV experiment. On the left part of the blue dashed line, the performance of BA/EL predictors is depicted (blue bars). On the right side, we show the performance of various knowledge transfer approaches and TLStab (dark yellow bars). (B) Pearson’s correlation and Kendall’s tau performance of TLStab against other approaches on the Ebola virus Dataset. On the left part of the blue dashed line, the performance of BA predictors is depicted (blue bars). On the right side, we show the performance of state-of-the-art pMHC stability tools (teal bars) compared to TLStab (dark yellow bar). (C) Pearson’s correlation and Kendall’s tau performance on the Pox virus Dataset. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 4.

(A) HLA-A*02:01 motifs for the four depicted methods. While the binding affinity predictor has visible presence of negative charge in position 4, the stability prediction methods are less enriched, although the contribution of negative charge to peptide stability has been previously studied [63]. (B) HLA-A*01:01 motifs for the four depicted methods. The NetMHCstabpan motif has substantially lower presence of D3 and E3, although there is experimental evidence of formed bonds that contribute to peptide stability.

Fig. 5.

(A) AUPRC performances of different approaches on the convalescent donor labels. Baseline AUPRC Is depicted In dashed gray. (B) AUPRC performances of different approaches on the unexposed donor labels. (C) Pearson’s correlation and Kendall’s tau performances of different approaches on the response frequencies of convalescent donors. (D) Pearson’s correlation and Kendall’s tau performances of different approaches on the response frequencies of unexposed donors.