AI-powered literature-based scoring and Z-score normalization for multi-property profiling of dual-effect natural compounds

Abstract Phytochemicals with dual therapeutic effects particularly those exhibiting both antidiabetic and anticancer activity represent promising candidates for drug repurposing and multi-target drug discovery. We present an AI-assisted pipeline for ranking natural compounds with dual antidiabetic and anticancer potential. The system integrates semantic similarity (S-PubMedBERT), stance detection (BioBERT-NLI), co-occurrence analysis from PubMed and Scopus, and clinical trial profiling. Scores for antidiabetic and anticancer activity were derived as functions of antioxidant and anti-inflammatory relevance common mechanistic intermediates. A hybrid Z-score normalization strategy was used: intra-group baselining for large chemical families (n ≥ 5) and metformin-referenced standardization for smaller groups. Importantly, the system is designed as a literature-driven evidence-synthesis layer—structuring and weighing published experimental and clinical findings—rather than a de novo predictor of bioactivity in the absence of literature support. Quantitatively, the stance classifier achieved [F1 = 0.xx / accuracy = 0.xx] on n = [xxx] manually annotated evidence sentences, and the prioritization recovered [k of m] canonical multifunctional compounds within the top-[K] candidates (precision@K = 0.xx). The pipeline identified lead phytochemicals (e.g., curcumin, resveratrol, quercetin) consistently exceeding Z > 1 across both therapeutic axes. This scalable, literature-driven framework integrates biomedical NLP and weighted evidence to prioritize multifunctional compounds for drug discovery.