TCR-H: explainable machine learning prediction of T-cell receptor epitope binding on unseen datasets

Area of Science:

• Immunoinformatics
• Computational Biology
• Machine Learning

Background:

• Current artificial-intelligence and machine-learning (AI/ML) models for predicting T-cell receptor (TCR)-epitope specificity often fail to generalize to unseen data.
• This limitation hinders their real-world applicability in understanding immune responses.

Purpose of the Study:

• To develop a robust AI/ML model for predicting TCR-epitope specificity that generalizes to unseen data.
• To enhance the explainability of TCR-epitope specificity predictions.

Main Methods:

• Developed TCR-H, a supervised classification Support Vector Machines model utilizing physicochemical features.
• Trained TCR-H on the largest available dataset, using only experimentally validated non-binders as negative data points.
• Employed the SHAP (Shapley additive explanations) eXplainable AI (XAI) method for model interpretation.

Main Results:

• TCR-H achieved an AUC of ROC of 0.87 for epitope 'hard splitting' (unseen epitopes).
• The model demonstrated strong performance with an AUC of 0.92 for TCR hard splitting and 0.89 for 'strict splitting' (unseen epitopes and TCRs).
• SHAP analysis identified key physicochemical features driving model predictions.

Conclusions:

• TCR-H significantly improves the prediction of TCR-epitope specificity, particularly for unseen data.
• The model's generalizability and explainability represent a substantial advancement in the field.
• This work paves the way for more reliable computational tools in immunology.