Deep generative models for T cell receptor protein sequences.

Area of Science:

• Computational biology
• Immunoinformatics
• Machine learning in immunology

Background:

• Adaptive immune repertoires are crucial for immune responses and disease.
• Traditional models of T cell receptor (TCR) repertoire diversity often rely on V(D)J recombination and selection models.
• Understanding repertoire generation and variation is key for applications like vaccination and diagnostics.

Purpose of the Study:

• To explore the application of deep learning, specifically variational autoencoders (VAEs), for modeling T cell receptor (TCR) repertoires.
• To assess the capability of VAEs in capturing the complex distributions of immune repertoire sequences.
• To evaluate VAEs' performance in learning biological rules of VDJ recombination and generalizing to new data.

Main Methods:

• Fitting deep neural network-parameterized VAE models to large-scale T cell receptor (TCR) repertoire sequencing data.
• Utilizing VAEs for cohort frequency estimation and learning VDJ recombination rules.
• Comparing VAE-generated sequences with those from traditional recombination-selection models.

Main Results:

• Simple VAE models achieved accurate cohort frequency estimation of TCR repertoires.
• The VAEs successfully learned the underlying rules governing VDJ recombination.
• VAE models demonstrated strong generalization capabilities to unseen TCR sequences.
• VAE-generated sequences exhibited characteristics similar to real biological sequences.
• VAE-like models could effectively distinguish between real and simulated TCR sequences.

Conclusions:

• Deep learning VAEs offer a powerful and effective approach for modeling adaptive immune repertoires.
• VAEs can capture complex sequence distributions and learn fundamental biological processes like VDJ recombination.
• This approach holds promise for advancing immunoinformatics, enabling better understanding of immune responses and development of novel therapeutics.