Quantifying Cross-Attention Interaction in Transformers for Interpreting TCR-pMHC Binding.

Area of Science:

• Immunology
• Computational Biology
• Artificial Intelligence

Background:

• T Cell Receptor (TCR) and peptide-Major Histocompatibility Complex (pMHC) interactions are crucial for adaptive immunity.
• Transformer models show promise in modeling TCR-pMHC binding but lack interpretability.
• Existing explainable AI (XAI) methods are incompatible with the encoder-decoder architectures used in TCR-pMHC modeling.

Purpose of the Study:

• To develop a novel post-hoc XAI method, Quantifying Cross-Attention Interaction (QCAI), for interpreting encoder-decoder transformers in TCR-pMHC modeling.
• To address the limitations of current XAI techniques in understanding the black-box nature of these models.
• To improve mechanistic insights into T cell responses and guide therapeutic development.

Main Methods:

• Proposed QCAI, a new post-hoc method specifically designed for quantifying cross-attention in transformer decoders.
• Compiled TCR-XAI, a benchmark dataset of 274 experimentally determined TCR-pMHC structures for quantitative evaluation.
• Evaluated QCAI's performance by computing physical distances between relevant amino acid residues and comparing residue importance estimations against ground truth.

Main Results:

• QCAI effectively interprets cross-attention mechanisms within transformer decoders used for TCR-pMHC binding.
• The method demonstrates state-of-the-art performance on the TCR-XAI benchmark for both interpretability and prediction accuracy.
• QCAI provides a quantitative evaluation framework for XAI methods in this domain.

Conclusions:

• QCAI offers a significant advancement in explainable AI for TCR-pMHC interaction modeling.
• The method enhances mechanistic understanding of immune responses and aids in the development of targeted immunotherapies.
• QCAI establishes a new benchmark for evaluating XAI methods in computational immunology.