An information-theoretic framework for inferring T cell dynamics from single-cell RNA and T cell receptor sequencing.

Abstract

Single-cell T cell receptor (TCR) repertoire sequencing enables simultaneous profiling of gene expression and TCR sequences, providing a coregistered measure of T cell phenotypic diversity and clonal dynamics. However, principled probabilistic frameworks for joint modeling of clonotype and phenotype have remained limited. Here we introduce TCRi, a hierarchical Bayesian information-theoretic framework that infers the joint probability distribution between genetic clonotype and transcriptional phenotype from paired single-cell transcriptomic and TCR sequencing data. TCRi treats phenotype assignment as a probabilistic inference problem: a variational autoencoder embeds single-cell expression into a structured latent space, and a hierarchical Bayesian model learns clonotype-specific phenotype distributions through principled information-sharing across clonotypes, patients and covariates. From these distributions, TCRi computes information-theoretic quantities: clonotypic entropy, phenotypic entropy, phenotypic flux and mutual information between clonotype and phenotype. When TCRi was applied to four real-world datasets comprising 167,118 CD8 T cells from patients with cancer treated with anti-PD-(L)1 immune checkpoint blockade, mutual information emerged as a robust discriminator of therapeutic response.