Machine learning-guided discovery of a novel natural CDK7 inhibitor for triple-negative breast cancer

Triple-negative breast cancer (TNBC) is an aggressive breast cancer subtype with limited targeted therapeutic options; while sacituzumab govitecan and PARP inhibitors (for BRCA-mutated TNBC) have received regulatory approval in specific patient subgroups, the majority of TNBC patients continue to lack effective targeted therapy. Cyclin-Dependent Kinase 7 (CDK7; UniProt P50613) is a crucial regulator of transcription and a promising therapeutic target. Natural compounds are a rich source of bioactive scaffolds, though they remain largely unexplored as CDK7 inhibitors. We applied a combined computational pipeline of machine learning (ML), virtual screening, molecular docking, molecular dynamics (MD) simulation, and MM-GBSA binding free energy analysis. We developed an ensemble ML model (LightGBM, XGBoost, CatBoost, and Random Forest; cross-validation R² = 0.7402) using a curated dataset of 935 CDK7 bioactivity compounds (spanning active through weakly active/inactive, pIC₅₀ range 3.18-9.51). The model evaluated 738,824 COCONUT natural compounds and identified 300 candidates for docking against CDK7 (PDB: 6XD3). The top five compounds underwent 10 ns MD simulations and MM-GBSA analysis, and the top hit was further examined with a 500 ns MD simulation. COC_081808, an indolocarbazole alkaloid (staurosporine family), emerged as the top candidate, exhibiting the best docking score (-12.28 kcal/mol), single-frame MM-GBSA (-34.18 kcal/mol), and trajectory MM-GBSA (-46.51 ± 3.82 kcal/mol). It formed stable hydrogen bonds with ASP155, ASN141, and ASN142 and π-π interactions with PHE91 and PHE93. Comparative MD showed the lowest ligand RMSD (1.01 ± 0.21 Å) and high hydrogen-bond occupancy. The 500 ns simulation verified stable protein-ligand interactions, with a protein RMSD of 1.87 Å, constant compactness, and the predominant low-energy conformational basin from PCA. COC_081808 is a computationally prioritized candidate with consistent multi-level evidence supporting potential CDK7 binding; experimental validation, including biochemical IC₅₀ determination, cellular assays, and kinome-wide selectivity profiling, remains essential before drawing therapeutic conclusions.