Explanation-Driven Visual Interventions in Genomics

Abstract

Artificial Intelligence (AI) is reshaping medical diagnostics, yet its adoption in high-stakes genomic applications remains constrained: the opacity of AI models prevents clinicians from meaningfully verifying, contesting, or learning from automated decisions. This paper presents the GenXAI vision, a system architecture that combines Genomic Language Models (GLMs) with Explainable AI (XAI) and physician-facing visual interfaces to support trustworthy, reproducible microbial species identification from metagenomic data. Applying the paradigm of explanation-driven interventions and Symbiotic AI (SAI), GenXAI goes beyond passive explainability: clinicians can inspect, annotate, and correct XAI attributions. Each correction feeds back into model retraining, creating a continuous mutual improvement loop in which human expertise and AI capability co-evolve. The system is designed from the outset to comply with the EU AI Act and to ensure long-term auditability. We argue that GenXAI instantiates the principles of trustworthiness, reproducibility, and intelligent visual analytics called for by TRI-IVIS in the genomic diagnostics domain.

Full citation