From Model Explanations to Drug Discovery Evidence: A Critical Review
Explanations can inform which molecular analog is synthesized, which protein contact is investigated, and which repurposing hypothesis is tested. This structured critical narrative review examines the evidence supporting such uses across drug-discovery tasks. A literature-derived evidence-oriented explainability framework (EOEF) provides a conceptual guide for assessing specified explanation claims, connecting established evaluation concepts to explanation objects, reference provenance, comparison conditions, retained coverage, and intended use. Seven scoped records from six selected studies illustrate these relationships. Published comparisons show that contact supervision improves structural recovery under tested novelty conditions, whereas measured structure–activity relationships, structural calculations, and external prediction support different scientific questions. Counterfactual uncertainty filtering increases agreement with a computational property oracle on a retained subset; gain and remaining coverage must be read together. Predictive feature sets can remain useful despite variable individual rankings. Human-evaluation findings likewise depend on the information supplied by the tested interfaces. These comparisons distinguish evidence for model investigation, domain hypotheses, and candidate prioritization. A recommended claim–evidence–scope record makes the relevant conditions traceable, helping connect an explanatory claim to the observation and comparison needed for a discovery choice.
Authors
- Mohd Tahir Ismail (ORCID: https://orcid.org/0000-0003-2747-054X)
- Xuyang Lu (ORCID: https://orcid.org/0009-0009-2280-7781)
Institutions
- Universiti Sains Malaysia (MY)
Publication Details
- Journal
- Pharmaceuticals
- Published
- 2026-11-16
- DOI
- https://doi.org/10.3390/ph19111817
- Primary Topic
- Explainable Artificial Intelligence (XAI)
- Type
- article
- Field-Weighted Citation Impact
- 0.00