Repurposing Natural Products with Artificial Intelligence: A Framework-Oriented Review of Molecular Similarity, Target Networks, and Disease-Repositioning Logic

Download PDF

Natural products remain important sources of chemically diverse and biologically active molecules, yet repositioning them toward new therapeutic contexts requires a careful distinction between computational hypothesis generation and pharmacological validation. This framework-oriented review examines how artificial intelligence and computational pharmacology can support natural-product repurposing through three connected reasoning domains: molecular similarity, target-network inference, and disease-repositioning logic. The review integrates methodological and application-oriented evidence to organize how natural-product identity can be transformed into computable molecular representations, compared within chemical space, linked to putative targets, mapped onto disease-associated networks, and integrated with omics, pharmacological, exposure, and validation evidence. The proposed framework emphasizes that molecular fingerprints, descriptors, molecular graphs, SMILES-based encodings, and learned embeddings can support analogue retrieval and chemical-neighbourhood analysis, but their outputs depend strongly on representation choice and applicability domain. Target prediction, target fishing, drug–target interaction modeling, protein-network mapping, disease-module analysis, knowledge graphs, and transcriptomic reversal can extend beyond structural resemblance, but each introduces assumptions about data provenance, target context, biological directionality, and causal relevance. Candidate prioritization is therefore framed as an evidence-calibration process rather than a proof of therapeutic utility. The review highlights the need to distinguish similarity, prediction, prioritization, validation, and translation; to identify where evidence layers converge or conflict; and to require experimental, exposure-aware, and translational assessment before repositioning claims are considered credible. Future progress depends on better natural-product data standardization, uncertainty-aware models, context-specific networks, leakage-resistant benchmarks, reproducible workflows, and prospective validation.