ImmunoTrace: A Meta-Agent for Immune History Tracking

Jiahao MA, Hongzong LI, Ye-Fan Hu, Jian-Dong Huang

Keywords: AI Agent, Retrieval-Augmented Modeling, Contrastive Learning, Probabilistic Evidence Fusion, Immune Exposure

TL;DR: ImmunoTrace is an AI agent that links a single-time-point TCR repertoire (with optional HLA) to proteome-scale peptide libraries.

Abstract: The adaptive immune system encodes an individual's exposure history in the T-cell receptor (TCR) repertoire. We present ImmunoTrace, an AI agent for immune history tracking that estimates past pathogen exposure from a single time-point repertoire by linking TCRs and HLA alleles to proteome-scale peptide libraries. A shared protein language model encodes TCR CDR3 sequences, HLA pseudo-sequences, and candidate peptides. Three high-capacity projection heads adapt these embeddings, and two cross-attention modules explicitly model TCR–peptide and HLA–peptide interactions. The fused representation is passed to a deep classifier to produce binding probabilities, while a contrastive branch with an InfoNCE objective and a learnable temperature sculpts the embedding space; we jointly optimize the contrastive and BCE losses while partially fine-tuning ESM2. For subject-level tracking, scores are calibrated into probabilities and evidence is aggregated across the repertoire with a probabilistic fusion scheme, yielding pathogen-level exposure estimates together with interpretable peptide-level evidence. On a multi-pathogen benchmark that includes Treponema pallidum (syphilis) and Neisseria gonorrhoeae (gonorrhea), ImmunoTrace surpasses strong baselines, generalizes under protein and HLA distribution shifts, maintains well-calibrated predictions, and scales to proteome-sized libraries with practical latency. We will release code and data-preparation recipes to facilitate reproducibility.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 51