Cheminformatics analysis, machine learning-based QSAR model building and virtual screening: identification of dual...

Rho-associated coiled-coil kinase 1 (ROCK1) and Transforming Growth Factor Beta Receptor 1 (TGFBR1) are pivotal mediators of epithelial-mesenchymal transition (EMT), cytoskeleton remodeling, cell adhesion, and cancer metastasis. Their simultaneous inhibition offers a promising therapeutic strategy to disrupt cancer metastasis. This study integrates machine learning (ML)-based predictive QSAR (Quantitative Structure Activity Relationship) modeling, structure-activity relationship (SAR) analysis, ML-based virtual screening to identify dual inhibitors targeting both ROCK1 and TGFBR1. Bioactivity data with SMILES representation of ROCK1 and TGFBR1 inhibitors were sourced from ChEMBL. Molecular descriptors and fingerprints (e.g., Morgan, AtomPairs2D, etc) were computed using RDKit and PaDEL to encode molecular properties. Analysis of chemical similarity using Tanimoto coefficients identified activity-cliff pairs of compounds with high structural similarity but significant activity differences. QSAR models were developed using individual and combined fingerprints, with additional strategies such as excluding activity cliff pairs and stratified sampling based on pIC₅₀ quantiles. Predictive machine learning models, including Random Forest (RF), Gradient Boosting (GB), XGBoost (XGB), K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and Ridge Regression (RR), demonstrated robust performance (R² > 0.6). The best ROCK1 and TGFBR1 models (RF Regressor) were used to screen a library of 1,48,684 compounds, resulting in the identification of a Hit (2-(1H-1,3-benzodiazol-1-yl)-N-[3-(1H-pyrazol-4-yl)phenyl]propanamide) with predicted dual ROCK1 (pIC₅₀ = 7.29) and TGFBR1 (pIC₅₀ = 7.00) inhibitory activity. Further, docking and molecular dynamics simulations of the Hit showed strong and stable interaction with both the targets hinge region residues, suggesting potential dual target inhibition activity.