Repertoire-Based Diagnostics Using Statistical Biophysics

A fundamental challenge in immunology is diagnostic classification based on repertoire sequence. We used the principle of maximum entropy (MaxEnt) to build compact representations of antibody (IgH) and T-cell receptor (TCR{beta}) CDR3 repertoires based on the statistical biophysical patterns latent in the frequency and ordering of repertoires constituent amino acids. This approach results in substantial advantages in quality, dimensionality, and training speed compared to MaxEnt models based solely on the standard 20-letter amino-acid alphabet. Descriptor-based models learn patterns that pure amino-acid-based models cannot. We demonstrate the utility of descriptor models by successfully classifying influenza vaccination status (AUC=0.97, p=4x10-3), requiring only 31 samples from 14 individuals. Descriptor-based MaxEnt modeling is a powerful new method for dissecting, encoding, and classifying complex repertoires.