Multi-pathway feature-level interpretability in MHC-I antigen presentation via concept-based modeling.

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Accurate prediction of Major Histocompatibility Complex class I (MHC-I) antigen presentation is crucial for neoantigen discovery and developing immunotherapies.
• Current deep-learning predictors, while accurate, often function as black boxes, limiting their interpretability and trustworthiness.

Purpose of the Study:

• To introduce MHCCBM, a gray-box framework designed to decompose MHC-I antigen presentation into interpretable pathway-level concepts.
• To enhance the trustworthiness of antigen presentation predictors through improved interpretability.

Main Methods:

• MHCCBM decomposes antigen presentation into distinct modules: proteasomal cleavage, TAP transport, peptide-MHC binding affinity, and chaperone dependency.
• The framework supports representational multimodality, integrating sequence-based, structure-informed, or empirical predictors.
• The reference implementation uses ESM-2 for peptide processing and chaperone dependency, MHCflurry for binding affinity, and logistic regression for concept combination.

Main Results:

• The MHCCBM framework achieves accuracy comparable to state-of-the-art predictors.
• Provides MHC-I pathway-level interpretability that aligns with established cellular mechanisms.
• Demonstrates flexibility by allowing substitution of individual predictive modules.

Conclusions:

• MHCCBM offers a more interpretable and trustworthy alternative to black-box models for MHC-I antigen presentation prediction.
• The modular design facilitates integration of diverse predictive approaches and enhances understanding of the antigen presentation pathway.
• This framework holds significant potential for advancing neoantigen discovery and immunotherapy development.