Beyond XAI: a claim-specific translation matrix for cancer biomarkers and therapeutic targets

Explainable artificial intelligence (XAI) is increasingly used in molecular and multimodal oncology to identify features that influence machine-learning predictions. A model-important feature may subsequently be described as a biomarker, mechanistic driver or therapeutic target, although these labels represent distinct scientific claims requiring different evidence. We propose the XAI Biomarker Translation Matrix, a claim-specific conceptual framework separating explanation evidence (E), biological evidence (B) and translational evidence (T). A claim gate routes XAI-prioritised candidates into three validation pathways: diagnostic, prognostic or treatment-predictive biomarkers; mechanistic therapeutic targets; and tumour-selective delivery targets. The explanation dimension is XAI-specific and addresses whether attribution is defined, stable, faithful and reproducible, whereas the biological and translational dimensions connect the attributed feature to claim-appropriate evidence. Claim-evidence concordance requires terminology to remain no stronger than the evidence demonstrated for the intended use, and falsification-first translation asks investigators to prespecify the observation most capable of challenging the proposed interpretation. Complementing established AI-reporting, biomarker-validation and target-assessment standards, the matrix adds an XAI-specific explanation-evidence layer and an explicit claim gate linking attribution to claim-specific evidence and terminology: XAI prioritises candidates; the claim determines the validation pathway.