Development of Target-specific Machine Learning Models for in silico Prediction of Antidiabetic Phytocomplex Components

Abstract Aim. To develop and validate target-specific machine learning models for in silico prediction of multi-target biological activity of natural compounds relevant to type 2 diabetes mellitus (T2DM) pathogenesis, and to apply them for screening the components of a multi-component herbal mixture. Methods. Fourteen LightGBM classification models with isotonic probability calibration were built to predict activity against targets from five categories related to T2DM pathogenesis: antidiabetic (DPP-4, PPARγ, PTP1B), anti-inflammatory (COX-2, 5-LOX, TNF-α), antioxidant (Keap1, xanthine oxidase), antihypertensive (AT1 receptor, PDE5, renin), and hypolipidemic (CETP, fatty acid synthase, PPARα). Training sets were obtained from the ChEMBL database (pIC50 ≥ 6.0, active; ≤ 5.0, inactive). Each molecule was described by 2225 features: 10 physico-chemical descriptors, MACCS keys (167 bits), and Morgan fingerprints (ECFP4, 2048 bits). The chemical composition of 11 plants was determined by GC-MS and LC-MS. Results. All 14 models showed ROC-AUC of 0.917–0.994 (mean 0.965), F1 of 0.923–0.986, and MCC of 0.570–0.891. Validation with reference drugs confirmed correct identification of clinically used inhibitors. Screening of 11 plants revealed complementary activity profiles: anti-inflammatory activity dominated in mint, oregano, and elderberry; antidiabetic activity in V. uliginosum and V. myrtillus. Conclusions. The proposed target-specific approach identifies the multi-target bioactivity of herbal complexes across multiple T2DM targets and supports the rational design of phytocompositions. Keywords : in silico, machine learning, type 2 diabetes mellitus, herbal complex, target-specific models, LightGBM.