Designing meaningful continuous representations of T cell receptor sequences with deep generative models.

Area of Science:

• Immunology
• Bioinformatics
• Machine Learning

Background:

• T Cell Receptor (TCR) antigen binding is crucial for adaptive immunity.
• The high diversity of TCRs and protein interaction complexity hinder effective low-dimensional representations for analysis.
• Current TCR analysis methods face limitations in capturing the full spectrum of TCR interactions.

Purpose of the Study:

• To develop a novel computational framework for learning low-dimensional representations of TCR sequences.
• To enable high-quality de novo generation of TCR sequences.
• To provide a robust method for TCR clustering and analysis.

Main Methods:

• Developed TCR-VALID, a capacity-controlled disentangling variational autoencoder.
• Trained the model on a large dataset of approximately 100 million TCR sequences.
• Quantified the properties of the learned representations, including dimensionality, continuity, and disentanglement.

Main Results:

• TCR-VALID generates low-dimensional, continuous, and disentangled TCR representations.
• The model enables high-quality de novo generation of TCR sequences.
• TCR-VALID representations facilitate fast and accurate TCR clustering, outperforming existing methods.

Conclusions:

• TCR-VALID offers a powerful new approach for T Cell Receptor sequence analysis and representation learning.
• The developed framework advances the field of protein representation learning in low dimensions.
• This method enhances our ability to study TCR-antigen interactions and immune responses.