Integrating molecular dynamics and machine learning to identify potential apo-state conformational and solvent-exposure signatures associated with resistant KRAS mutants

From Descriptor Learning to Binding Stability: An Explainable Machine Learning Pipeline for EGFR Double-Mutant Inhibitor Discovery

An integrated computational workflow combining explainable machine learning, virtual screening, molecular dynamics simulations, and binding free-energy calculations to identify novel inhibitors of this drug-resistant EGFR variant may support the development of new therapeutic strategies for overcoming resistance in EGFR-driven cancers.

Allosteric co-binding restores tyrosine kinase inhibitor affinity in T790M-mutant EGFR through conformational reprogramming.

Drug resistance in epidermal growth factor receptor (EGFR)-mutant cancers commonly arises from kinase-domain substitutions that remodel the adenosine triphosphate binding pocket and reduce complementarity to orthosteric inhibitors, with the T790M gatekeeper mutation posing a major challenge. This study evaluated whether pre-occupying a proximal allosteric pocket with selected phytochemicals could bias mutant EGFR toward drug-compatible conformations and improve inhibitor binding. A two-phase computational workflow was employed: (i) molecular docking of gefitinib and erlotinib to wild-type and mutant EGFR; and (ii) allosteric pre-docking of phytochemicals followed by redocking of the ATP-site inhibitor. Top-ranked complexes were advanced to 200-nanosecond all-atom molecular dynamics simulations in explicit solvent and end-state binding free-energy estimation using Molecular Mechanics Generalized Born Surface Area (MM/GBSA). Docking predicted stronger binding to wild-type EGFR and reduced affinity for the T790M mutant, whereas co-binding produced compound-dependent improvements. Simulations suggested partial stabilization of the protein-ligand complexes, characterized by reduced root mean square deviation, damped hinge and αC-helix motions, reduced solvent exposure, and radii of gyration approaching wild-type behavior. Binding free energies improved from -12.6 to -17.6 kcal mol-1 (Genistein) and -19.58 kcal mol-1 (Tupichinols C) for gefitinib, and from -13.4 to -18.1 and -20.1 kcal mol-1, respectively, for erlotinib. Absorption, distribution, metabolism, excretion, and toxicity screening supported the developability of the leading candidates. This integrated framework provides structural, dynamic, and energetic criteria for prioritize cooperative allosteric-orthosteric co-binding chemotypes for experimental validation.

Exploring conformational dynamics of the HER2 DFG-flip using machine-learning-guided metadynamics for type II inhibitor design.

Protein kinases regulate cellular proliferation and survival through tightly controlled signalling mechanisms, and their dysregulation is a major driver of cancer. Human epidermal growth factor receptor 2 (HER2) is a clinically validated kinase target, yet structural and drug-discovery efforts have largely focused on type I inhibitors binding the active DFG-in conformation. The absence of experimentally resolved structures for HER2 in the inactive DFG-out state has limited structure-based development of inactive-state type II inhibitors. Here, we employ molecular dynamics (MD) simulations coupled with machine-learning-guided well-tempered metadynamics (MetaD) to characterize the conformational landscape underlying the DFG-in to DFG-out transition in HER2. Deep learning-based collective variable was constructed by the Deep Targeted Discriminant Analysis (DeepTDA) method employing unbiased MD data from both metastable states. DeepTDA enabled efficient sampling of the DFG-flip pathway, capturing multiple recrossing events and allowing reliable estimation of the free energy difference between the two states within accessible simulation time. The inactive DFG-out conformation is found to be energetically favoured by approximately 25 kJ/mol relative to the DFG-in state, in close agreement with experimental and computational data from other apo kinases. In addition to the canonical active conformation observed crystallographically, we identify an alternative DFG-in conformer that may contribute to the low intrinsic kinase activity reported for HER2. A DFG-up intermediate conformer is also observed, resembling crystal structures reported for Aurora-A kinase. A combined covalent docking and molecular dynamics approach followed to assess the MetaD-predicted DFG-out structure for ligand binding reliability, yielding comparable predicted/experimental binding affinity. Together, these results provide a structurally grounded inactive-state HER2 model and a mechanistic framework for structure-based design of HER2 inhibitors targeting inactive kinase conformations, highlighting the applicability of deep learning CVs to systems of complex free energy landscape.

Harnessing machine learning, docking and molecular dynamics for the virtual screening of compounds as CDK4/6 dual inhibitors.

Machine learning–driven identification of PIM2 kinase inhibitors through QSAR modeling and molecular dynamics simulations

The proto-oncogene serine/threonine kinase PIM2 is a critical regulator of cell proliferation, survival, and tumor progression and represents an attractive therapeutic target for several cancers. In this study, an integrated machine learning–guided computational pipeline was developed to identify potential PIM2 inhibitors by combining quantitative structure–activity relationship (QSAR) modeling, virtual screening, molecular docking, molecular dynamics (MD) simulations, and pharmacokinetic prediction. Bioactivity data for PIM2 inhibitors were retrieved from the ChEMBL database, yielding 5953 compounds. After data cleaning, structural standardization, and removal of duplicates and invalid entries, a curated dataset of 1584 compounds was obtained for QSAR modeling. To address dataset imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied before model development. Twelve molecular fingerprint descriptors were generated and used to construct 180 QSAR models using five machine learning algorithms, including Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), k-Nearest Neighbors (KNN), and Multilayer Perceptron (MLP). Among these models, the Random Forest–fingerprint model demonstrated the best predictive performance, achieving a mean R2 of 0.971 with low prediction errors (RMSE = 0.271; MAE = 0.125) across training, testing, and cross-validation datasets. The optimized model was subsequently applied to virtual screening of multiple chemical libraries, including FDA-approved drugs, natural product databases, and commercial compound collections. Several promising candidates were identified, including TCMBANKIN000009 (emetine), Amb28533044 (4,6′-Anhydrooxysporidinone), NPC170963 (Lysophosphatidylcholine (15:0)), NPC262768 (Endosulfan), and NPC469603 (8-hydroxyircinialactam A). Molecular docking showed that these compounds bind within the ATP-binding pocket of PIM2 kinase, forming interactions with key residues such as Lys62, Asp125, Asp128, and Glu168. Subsequent molecular dynamics simulations confirmed the stability of selected complexes, demonstrating reduced residue fluctuations, stable protein compactness, and persistent intermolecular interactions during the simulation. Furthermore, ADMET prediction suggested favorable pharmacokinetic and toxicity profiles for several compounds. Collectively, these findings highlight the potential of the identified molecules as promising PIM2 inhibitor candidates, providing valuable leads for future experimental validation and anticancer drug development.

Targeting WEE1 kinase: an integrated machine learning–cheminformatics framework for ultra-large-scale virtual screening and novel inhibitor discovery