Foundation models are increasingly presented as reusable computational infrastructures for molecular property prediction, compound generation, reaction modelling, biological representation learning, and pharmacoinformatics. Their relevance to quantitative structure–activity relationship modelling, however, cannot be inferred from pretraining scale or benchmark performance alone. This scoping review maps how foundation-model concepts are defined, trained, adapted, transferred, grounded, calibrated, and evaluated across chemically and biologically relevant data modalities. A PRISMA-ScR-compatible and JBI-informed approach was used to define broad mapping questions, eligibility boundaries, information sources, search logic, screening criteria, data-charting fields, and a descriptive thematic synthesis. The evidence was organized by model family, pretraining data, representation type, adaptation mechanism, downstream task, transfer setting, validation stage, uncertainty treatment, chemical-validity assessment, and intended decision context. The literature spans molecular graphs and strings, conformational or multi-view molecular data, reactions, proteins, omics, assay descriptions, and scientific text. Evidence is concentrated in retrospective model development and internal benchmark evaluation, whereas independent external validation, calibration under distribution shift, explicit applicability-domain analysis, prospective experimental confirmation, and decision-context evaluation are less consistently represented. Reported transfer is strongly conditioned by endpoint definition, chemical-space overlap, split strategy, label quality, comparator selection, and access to task-specific data. Scientific grounding remains heterogeneous, ranging from syntactic molecular validity to tool-supported chemical checking, retrosynthetic assessment, and interpretable representation analysis. The review identifies a need for clearer foundation-model definitions, realistic transfer tests, uncertainty evaluation under shift, stronger chemical-validity controls, reproducible workflows, and explicit boundaries between computational performance and pharmaceutical utility. Foundation models may broaden reusable representation learning, but their value for QSAR and pharmacoinformatics remains conditional on evidence appropriate to the intended chemical domain and decision use.
IJPRAS
