Contrastive learning of T cell receptor representations.

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Predicting T-cell receptor (TCR) and ligand interactions is crucial but challenging due to sparse specificity-labeled data.
• Language models pre-trained on unlabeled data show promise for addressing data bottlenecks in other fields.
• Optimal pre-training strategies for protein language models in TCR specificity prediction are not well-established.

Purpose of the Study:

• To introduce SCEPTR, a novel TCR language model for data-efficient transfer learning in TCR specificity prediction.
• To develop and evaluate a pre-training strategy combining autocontrastive learning and masked-language modeling for TCRs.
• To advance computational methods for understanding TCR-ligand interactions.

Main Methods:

• Developed SCEPTR (simple contrastive embedding of the primary sequence of T-cell receptors), a specialized TCR language model.
• Implemented a novel pre-training strategy integrating autocontrastive learning and masked-language modeling.
• Compared SCEPTR performance against existing protein language models and sequence alignment-based methods.

Main Results:

• SCEPTR achieved state-of-the-art performance in TCR specificity prediction, demonstrating data-efficient transfer learning capabilities.
• The proposed pre-training strategy significantly improved SCEPTR's predictive accuracy.
• SCEPTR outperformed existing protein language models and a variant lacking autocontrastive learning.

Conclusions:

• SCEPTR provides an effective computational tool for predicting TCR specificity, addressing data scarcity challenges.
• Autocontrastive learning is a valuable paradigm for decoding the rules governing TCR specificity.
• This work paves the way for improved understanding and prediction of immune system interactions.