Attentive Variational Information Bottleneck for TCR-peptide interaction prediction

Area of Science:

• Computational Biology
• Immunology
• Machine Learning

Background:

• Predicting T-cell receptor (TCR) and peptide interactions is crucial in immuno-oncology.
• Existing methods face challenges in accurately modeling complex sequence data.

Purpose of the Study:

• To introduce Attentive Variational Information Bottleneck (AVIB), a novel model for multi-sequence generalization.
• To apply AVIB to the immuno-oncology problem of TCR-peptide interaction prediction.

Main Methods:

• AVIB utilizes multi-head self-attention to approximate posterior distributions over latent encodings.
• The model is conditioned on multiple input sequences for enhanced predictive power.

Main Results:

• AVIB significantly surpasses state-of-the-art methods in TCR-peptide interaction prediction accuracy.
• The learned latent posterior distribution effectively identifies out-of-distribution amino acid sequences in an unsupervised manner.

Conclusions:

• AVIB offers a powerful new approach for sequence-based prediction tasks in computational biology.
• The model's ability to detect anomalous sequences has implications for understanding immune responses and disease.