An AI Agent for Immune Receptor Fingerprint‑Based Diagnosis of Infection of Unknown Origin

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

Keywords: AI Agent, multi-task representation learning, Conditional sequence generation, Immune repertoire modeling, Epitope inference, Clinical diagnostics

TL;DR: Generative allele-aware epitope inference plus proteome retrieval turns TCR “fingerprints” into ranked pathogen hypotheses with calibrated confidence for IUO diagnosis.

Abstract: When routine tests fail to find a pathogen, diagnosing infections of unknown origin stalls. We instead read the patient's immune response for AI-readable clues. We formalize a new machine learning task: inferring plausible epitopes directly from immune-receptor repertoires and localizing their pathogen sources. To address this problem, we introduce a Transformer-based multi-sequence novel representation-learning model that jointly models T-cell receptors, human leukocyte antigen , and antigenic peptides, and we pretrain it across six tasks; the model achieves best or second-best performance across all six tasks against strong baselines. Building on this, we develop an end-to-end, clinically oriented agent that operates in a perceive--plan--act loop, orchestrating epitope generation, HLA-personalized filtering, consistency checks, and retrieval, with clinician-in-the-loop threshold adaptation; when evidence conflicts, it performs calibrated abstention and logs an interpretable decision trace. End-to-end on clinical-style repertoires with diagnostic report generation, the agent outperforms discriminative-pairing and direct-retrieval baselines. Upon publication, we will release all code, models, and pathogen indices under a research license, together with de-identified evaluation data.

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

Submission Number: 37