Applications of Explainable Artificial Intelligence in Bioinformatics and Their Experimental Validation

A sequence model can identify a base substitution predicted to alter a splice junction, and a cancer model can assign a tumour to a pathway-defined group. Researchers often interpret these outputs as explanations, although each calls for a different experimental test. This narrative review examines how explainable artificial intelligence is used in regulatory genomics, RNA processing, cancer genomics, single-cell and spatial transcriptomics, protein science, molecular property prediction and microbial genomics. It compares the biological entity being explained, the reference used to calculate an effect and the evidence available beyond the original model. Base substitutions and matched molecular pairs permit explicit contrasts. Pathway scores, tissue neighbourhoods and internal protein features require additional tests to separate the proposed mechanism from correlated measurements. Recent evaluations of foundation models also show why comparisons with simple predictors and genuinely held-out biological units matter before interpretation. Across applications, the clearest route to experimental follow-up is to specify an expected molecular event, its direction and a result that would challenge the prediction. This approach separates faithful descriptions of model behaviour from claims supported by an independent assay and clarifies what an explanation contributes to the selection of the next experiment.