IMMUNIA: A Multi-LLM Reasoning Agent for Immunoregulatory Surfaceome Discovery

Abstract

Biomarker discovery for immunotherapy often requires reasoning across complex immune contexts. We present IMMUNIA, a multi-large-language-model (multi-LLM) reasoning agent designed to identify immunoregulatory surfaceome genes through interpretable, biologically grounded analysis. The term IMMUNIA originates from the fusion of Immune and Noeia (the Greek concept of perception and understanding), defining an AI system that perceives, reasons, and interprets the immune landscape with human-like cognition. IMMUNIA integrates structured prompting, contextual scoring across immunotherapy, inflammation, and NF-{kappa}B signaling, and consensus reasoning across GPT-4o, GPT-5, and Gemini 2.5 Pro. Benchmarking with positive (HLA) and negative (contactin) controls confirmed model consistency and contextual discrimination. Consensus evaluation prioritized IL1R1, BSG, CD276, ALCAM, B2M, PTPRS, VCAN, and MXRA5 as high-confidence candidates. Among these, PTPRS, VCAN, and MXRA5 emerged as previously unrecognized stromal immune checkpoint-like regulators, shaping tumor-immune crosstalk via phosphatase, ECM, and cytokine signaling networks. IMMUNIA thus establishes a reasoning-centric AI paradigm that bridges computational inference with biological plausibility, offering a scalable approach for precision immunotherapy biomarker discovery.