Conflict-Aware Multimodal Adjudication Agent System for Interpretable Therapeutic Response Prediction from Histology and Spatial Transcriptomics

Therapeutic response prediction from H&E histology and spatial transcriptomics (ST) is commonly treated as a multimodal fusion problem. However, clinically meaningful cases often arise when modalities disagree: histology may indicate immune infiltration, stromal organization, or tumor viability, whereas ST may reveal immune dysfunction, suppressive niches, molecular resistance programs, or spatially restricted therapeutic targets. Rather than treating such disagreement as noise, we frame it as evidence to be tested, revised, and adjudicated. We propose CAMA, a Conflict-Aware Multimodal Adjudication Agent System for interpretable therapeutic response prediction from H&E histology and ST. CAMA separates hypothesis generation from verification: each modality proposes candidate response mechanisms, translates them into testable sub-hypotheses, evaluates them through intra-modal and cross-modal evidence, calibrates them against cohort-level context, and adjudicates the surviving trajectories into a patient-level response decision. This process yields not only a responder/non-responder prediction, but also supporting evidence, conflicting evidence, refined or rejected hypotheses, biological rationale, uncertainty, and an audit trail. Thus, CAMA reframes histology-ST integration from feature fusion to conflict-aware evidence adjudication, enabling biologically grounded and clinically auditable patient-level reasoning under heterogeneous multimodal evidence.