Year: 2024
Type: article
Abstract: Artificial intelligence is transforming genomic medicine, achieving state-of-the-art performance in variant pathogenicityprediction, polygenic risk scoring, clinical trial matching, and therapeutic recommendation. Yet clinical adoption remainsconstrained by a fundamental concern: deep learning models behave as black boxes that cannot explain why they makespecific predictions, limiting clinician trust, regulatory approval, and patient autonomy in healthcare decisions.Explainable AI (XAI) techniques -- attention visualisation, SHAP values, counterfactual explanations, mechanisticinterpretability, and causal inference -- aim to render model reasoning transparent while preserving predictiveperformance. The challenge in genomic medicine is particularly acute: explanations must be biologically meaningful(connecting to genes, pathways, molecular mechanisms), clinically actionable (supporting therapeutic decisions), andlegally defensible (satisfying FDA and EU AI Act transparency requirements). We present the Explainable Genomic AIFramework (EGAIF), evaluating five XAI approaches across four genomic medicine applications (variant pathogenicityclassification, polygenic risk score interpretation, drug response prediction, and clinical trial matching). Our ExplainableGenomic AI Score (EGAS) measures explanation fidelity to model behaviour, biological plausibility, clinical utility, usercomprehension, and regulatory compliance. Mechanistic interpretability achieves the highest EGAS (0.926) throughdirect mapping of neural network internals to biological pathways, while counterfactual explanations achieve the highestclinical actionability (0.960) by showing clinicians which genetic changes would alter predictions
Author: Laura Dubois
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
Funder: Austrian Science Fund
OpenAlex ID: W7203875403