Geometric deep learning improves generalizability of MHC-bound peptide predictions

Area of Science:

• Computational biology
• Immunoinformatics
• Structural bioinformatics

Background:

• Peptide-MHC interactions are critical for immunity, impacting autoimmunity, pathogen defense, and tumor surveillance.
• Accurate prediction of peptide-MHC binding is essential for advancing cancer immunotherapies.
• Current sequence-based prediction methods face challenges in generalizability across diverse MHC alleles.

Purpose of the Study:

• To address the generalizability limitations of current sequence-based MHC-bound peptide prediction methods.
• To improve the accuracy and efficiency of predicting peptides that bind to MHC molecules.
• To explore the potential of structure-based computational methods for MHC-peptide binding predictions.

Main Methods:

• Developed structure-based methods utilizing geometric deep learning (GDL) to predict MHC-bound peptides.
• Introduced a self-supervised learning approach on protein structures (3D-SSL) to enhance data efficiency.
• Evaluated method generalizability across unseen MHC alleles and resilience to binding data biases.

Main Results:

• Structure-based GDL methods showed improved generalizability across diverse MHC alleles compared to sequence-based approaches.
• The 3D-SSL method outperformed sequence-based methods trained on significantly more data, demonstrating superior data efficiency.
• Structure-based GDL methods proved resilient to biases present in binding affinity datasets, as shown in a Hepatitis B virus case study.

Conclusions:

• Structure-based methods, particularly GDL, offer a promising avenue for enhancing the generalizability and data efficiency of MHC-bound peptide predictions.
• These findings have significant implications for fields requiring accurate prediction of molecular interactions, such as cancer immunotherapy and T-cell receptor specificity prediction.
• The study underscores the potential of leveraging structural information and advanced machine learning techniques to overcome limitations in current immunoinformatics tools.