Counting is almost all you need.

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• The T-cell receptor (TCR) repertoire reflects an individual's infection history and immune status.
• Existing methods for disease detection using TCR repertoires have limitations in accuracy and complexity.

Purpose of the Study:

• To develop and evaluate novel computational methods for predicting disease history from TCR repertoires.
• To compare the performance of a simple counting method against leading algorithms and more complex models.

Main Methods:

• A TCR repertoire counting method was implemented and compared with existing algorithms.
• A novel attention model, utilizing Variational AutoEncoder (VAE) projections, was developed to improve TCR weighting.
• A Graph Convolutional Network (GCN) approach was proposed as an intermediate between counting and attention models.

Main Results:

• The counting method demonstrated superior performance compared to two leading algorithms.
• The attention model further improved prediction accuracy for Cytomegalovirus (CMV) infection and its associated HLA alleles.
• The GCN approach achieved accuracy comparable to the attention model with increased simplicity.

Conclusions:

• Simple counting methods can be highly effective for analyzing TCR repertoires for disease history.
• Advanced models like attention mechanisms and GCNs offer improved accuracy and/or simplicity in TCR repertoire analysis.
• These findings provide new tools for non-invasively assessing immunological history and disease presence.