Explainable Artificial Intelligence in Bioinformatics: Applications, Challenges, and Future Opportunities

Explainable Artificial Intelligence in Bioinformatics

Abstract

Machine learning now sits at the centre of bioinformatics, delivering strong performance on tasks from gene-expression classification and protein-structure prediction to digital pathology and drug-target scoring. The best-performing models, however, are the hardest to interrogate, and in biology a prediction that cannot be explained has limited scientific or clinical value. Explainable artificial intelligence (XAI) addresses that gap by producing human-readable accounts of model behaviour. This review surveys the techniques applied to biological data, SHAP, LIME, Grad-CAM, attention, layer-wise relevance propagation, graph explainers and biologically constrained architectures, and their use across genomics, transcriptomics, proteomics, sequence analysis, structural bioinformatics, multi-omics integration and biomedical imaging. Six structured tables and seven figures compare methods, applications, metrics and open problems, each accompanied by a statement of what the comparison contributes. We then weigh these methods against the critical evidence published between 2024 and 2026, which is more sobering than the enthusiasm of the primary literature suggests: attention-derived scores in single-cell foundation models add no predictive value over simple gene-level statistics, and widely used saliency methods can survive randomisation of the very model they claim to explain. Grouping ten challenges by cause rather than by symptom produces the review's central argument. Instability under collinearity, computational cost and the limits of attention are properties of current methods and will yield to better algorithms; the absence of mechanistic ground truth, the scarcity of interventional data and the gap between statistical importance and biological causality will not yield to algorithms at all, and depend instead on experimental and curatorial effort. We close with research priorities: causal and counterfactual explanation, interpretability of biological foundation models, community benchmarks, human-in-the-loop systems, multi-scale explanation and prospective clinical evaluation, and argue that XAI should be judged by whether its explanations survive experimental test, not by how plausible they look.