Building trust in deep learning-based immune response predictors with interpretable explanations.

Area of Science:

• Immunoinformatics
• Computational Biology
• Artificial Intelligence in Immunology

Background:

• Predicting peptide binding to Major Histocompatibility Complex (MHC) class I molecules is critical for vaccine development.
• Current deep learning predictors achieve high accuracy but function as black boxes, hindering trust and understanding.
• Interpretable explanations are needed to validate and trust these complex predictive models.

Purpose of the Study:

• To introduce MHCXAI, a novel framework applying eXplainable AI (XAI) techniques to interpret MHC class I peptide binding predictions.
• To provide human-understandable explanations for the decisions made by MHC class I predictors based on input peptide features.
• To enhance trust and understanding of deep learning models used in predicting immune responses.

Main Methods:

• Applied XAI techniques to analyze outputs from four state-of-the-art MHC class I predictors.
• Utilized a large dataset encompassing diverse peptides and MHC alleles for comprehensive testing.
• Evaluated the reliability and robustness of the generated explanations against ground truth data.

Main Results:

• Successfully generated interpretable explanations for MHC class I predictor outputs.
• Demonstrated the reliability and robustness of the XAI-generated insights.
• Provided feature-based rationales for peptide-MHC class I interaction predictions.

Conclusions:

• MHCXAI significantly improves the interpretability of MHC class I peptide binding predictors.
• The framework fosters greater trust in AI-driven vaccine design by offering validated explanations.
• This work advances the application of XAI in understanding complex biological processes like immune recognition.