Artificial intelligence in drug research: A nested-cohort bibliometric analysis of mechanistic interpretation and model interpretability

Artificial intelligence (AI) supports prediction and explanation in drug research. We examined the representation and development of mechanistic interpretation and model interpretability within this literature. Publications from 2000 to 2025 were analysed using Web of Science Core Collection (WoSCC), with PubMed providing a complementary perspective. Cohort A comprised eligible AI drug research; B was its subset addressing substantive drug-mechanism analysis and/or drug-related model interpretation. Comparisons contrasted B with A-only using temporal models, lexical analyses, network permutations, matched institutional analyses, and false-discovery-rate correction. WoSCC included 22,542 A publications, including 1324 in B; PubMed included 17,743 and 1048, respectively. From 2020 to 2025, B/A increased from 3.67% to 10.30% in WoSCC and from 3.17% to 10.64% in PubMed. Full-period trends were nonlinear (both P < 0.001). Explanation-related terminology also increased within A-only. Excluding query-linked terms, crude term-association estimates correlated across databases (Spearman ρ = 0.900; approximate 95% CI, 0.883-0.914); 320 WoSCC and 245 PubMed terms remained significant after publication-year and document-type adjustment and multiplicity correction. WoSCC direct-citation and bibliographic-coupling networks showed modest positive cohort assortativity (0.0746 and 0.0865; both q = 0.0005). In WoSCC, B showed greater institutional concentration than matched A-only samples (q < 0.001), without significant interinstitutional collaboration differences. Principal temporal and network directions persisted in screening sensitivity analyses. Mechanistic interpretation and model interpretability represent a growing emphasis within AI drug research, characterised by cross-database lexical concordance and, in WoSCC, modest network differentiation and greater institutional concentration.