Molecular Dynamics-Enhanced Machine Learning Models for Protein Binding and Stability

This thesis explores the integration of atomistic Molecular Dynamics (MD) with Machine Learning (ML) and Deep Learning (DL) to decode the complex functional profiles of biomolecules, moving beyond the traditional static representations of structural biology. The research introduces molecular dynamics-enhanced machine learning frameworks across three distinct biophysical domains, with a focus on innovative algorithmic strategies for extracting informative patterns from high-dimensional molecular trajectories. First, the study addresses the functional profiling of kinase targets, specifically Cyclin-Dependent Kinases (CDKs) and the broader human kinome, by initially utilizing classical machine learning models and subsequently developing MOLECULE—Molecular-dynamics and Optimized deep Learning for Entropy-regularized Classification and Uncertainty-aware Ligand Evaluation — a self-designed dual-modal deep neural network that combines static chemical fingerprints with dynamic structural fluctuations to classify allosteric and orthosteric inhibitors accurately. This novel approach bridges these complementary molecular representations, embedding molecular motion into a unified predictive architecture. In a second research line, the thesis tackles the challenge of predicting mutational protein stability by coupling plain MD with the Energy Decomposition Method (EDM) to reliably estimate thermodynamic impacts (ΔΔG) based on spectral energy gaps and van der Waals packing. Finally, extending the framework to therapeutic biomolecules, the research achieves the target-independent prediction of nanobody binding. This work demonstrates that the binding mechanism of SARS-CoV-2 neutralizing nanobodies can be classified in a strictly target-free environment by analyzing the energetic distribution, specifically the skewness and kurtosis, of the hypervariable CDR3 loop. Ultimately, this contribution establishes a scalable computational triage methodology that connects instantaneous static predictors with computationally expensive alchemical calculations, providing a rational, cost-aware protocol for molecular prioritization across a hierarchy of predictive models, with direct relevance to the design of selective small-molecule drugs and variant-proof biologics.