Current challenges for unseen-epitope TCR interaction prediction and a new perspective derived from image...

Area of Science:

• Immunoinformatics
• Computational Biology
• Machine Learning in Immunology

Background:

• T-cell receptor (TCR)-epitope recognition prediction has advanced for specific epitopes.
• Predicting all possible TCR-epitope pairs is difficult due to sequence diversity and limited data.

Purpose of the Study:

• To review current challenges in generic TCR-epitope recognition prediction.
• To introduce ImRex, a novel feature representation for TCR-epitope interaction.
• To evaluate validation strategies for TCR-epitope predictive models.

Main Methods:

• Examined validation strategies for assessing generalization performance on seen and unseen epitopes.
• Developed ImRex (interaction map recognition) using physicochemical properties of amino acids in CDR3 and epitope sequences.
• Utilized convolutional neural networks with the combined sequence representation.

Main Results:

• ImRex enables prediction for a subset of unseen epitopes not drastically different from training data.
• Appropriate feature engineering and rigorous benchmarking are crucial for model development and validation.
• Identified challenges including negative data selection, imbalanced epitope distribution, and TCR chain exchangeability.

Conclusions:

• Extrapolation to novel epitopes remains a significant challenge in TCR-epitope recognition.
• ImRex offers a feasible approach for predicting recognition of similar, unseen epitopes.
• Rigorous validation and feature engineering are essential for robust TCR-epitope predictive models.