ImmuneEmbedding: cross-modal recovery of the human immune state from paired TCR and BCR repertoires

Understanding which immune measurements can be predicted from adaptive receptor repertoires is central to representing individual immune state. We introduce ImmuneEmbedding, a modular participant-level representation and evaluation framework linking T-cell receptor (TCR) and B-cell receptor (BCR) repertoires to cellular and molecular measurements. The framework combines interpretable repertoire statistics with abundance-weighted ESM-2 sequence embeddings and evaluates a supervised structural bottleneck as a complementary representation. We study 99 clinical, cellular, transcriptomic, proteomic, and serum targets in COMBAT using donor-grouped cross-validation and donor-cluster bootstrap intervals. On 136 paired-repertoire timepoints, matched linear readouts cover 14 targets with combined structural inputs, compared with 11 for TCR and seven for BCR; three paired-input targets are absent from both single-chain coverage sets. On an expanded set of 140 timepoints, structural features cover 16 targets overall, while combining structure and sequence increases serum coverage to four, compared with three for structure and two for sequence alone. Comparisons with k-mer features, released JL-GloVe embeddings, and attention-based multiple-instance learning reveal distinct target profiles, with TCR attention reaching six serum targets. Across the evaluated ImmuneEmbedding configurations and both analysis sets, 38 targets meet the nominal coverage criterion, including 14 serum analytes. Separate cohort-calibrated TCR transfer experiments achieve cytomegalovirus serostatus AUROCs of 0.874 in confirmation donors and 0.719 across studies with supervised augmentation. These findings map repertoire-associated immune variation and show how receptor pairing, sequence representation, and aggregation shape the measurements accessible from an individual's repertoire.