Predicting specificity of TCR-pMHC interactions using machine-learning and biophysical models

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• T cell receptor (TCR) recognition of peptide-MHC complexes (pMHCs) is crucial for adaptive immunity.
• Accurate prediction of TCR specificity is vital for understanding immune responses and developing immunotherapies.
• Current predictive models, including machine learning (ML) and physics-based methods, have limitations in generalizing to unseen data.

Purpose of the Study:

• To evaluate the generalization capabilities of existing ML and physics-based methods for TCR-pMHC specificity prediction.
• To develop a novel ML method that improves prediction accuracy for both known and novel peptides.
• To analyze the impact of sequence similarity on model performance.

Main Methods:

• Utilized a proprietary cancer patient dataset for training and validation.
• Compared the performance of traditional ML methods against physics-based approaches.
• Developed and implemented a new ML model incorporating protein foundation models.
• Assessed model performance based on the sequence distance between training and testing datasets.

Main Results:

• ML methods accurately predict TCR specificity for known peptides but fail to generalize to novel peptides.
• Physics-based methods show better performance on novel peptides but underperform on known ones.
• The novel ML method using protein foundation models achieved superior or comparable performance across both known and novel peptides (in- and out-of-distribution).

Conclusions:

• Existing methods have distinct limitations in predicting TCR-pMHC interactions.
• Protein foundation models offer a promising avenue for developing more robust and generalizable TCR specificity prediction tools.
• Further research should focus on method development and strategic data acquisition to overcome current modeling challenges.