Interrogating Thermo-kinetic Bottlenecks in Biomolecular Processes: Designing a Machine Learning Lens for Computational Microscopy

Dutta, Pallab (2026) Interrogating Thermo-kinetic Bottlenecks in Biomolecular Processes: Designing a Machine Learning Lens for Computational Microscopy. PhD thesis, Indian Institute of Science Education and Research Kolkata.

Text (PhD thesis of Pallab Dutta (20RS084)) 20RS084.pdf - Submitted Version Restricted to Repository staff only Download (110MB)

Abstract

Decoding the complex thermo-kinetic rules governing biomolecular processes using standard Molecular Dynamics (MD) frequently encounters scale mismatches and kinetic trapping. Traditional enhanced sampling imposes severe accuracy-cost trade-offs or relies on subjective Collective Variables (CVs). To overcome this, this thesis designs a “machine learning lens”, a suite of AI-driven frameworks to objectively map high-dimensional free energy landscapes (FELs) and decode the structural determinants of rate-limiting transitions. First, Expectation Maximized Molecular Dynamics (EMMD) is introduced. By synergizing localized, unbiased MD with Gaussian and von Mises mixture models, EMMD reconstructs macroscopic FELs strictly from accessible metastable basins, circumventing prohibitive continuous biasing. To ensure landscape accuracy, a robust tripartite CV evaluation pipeline (integrating ISDOM, HCCM, and a modified VAMP score) is established to mathematically balance state distinguishability, orthogonality, and kinetic slowness. This protocol successfully captures the canonical Abl kinase DFG-flip and the complex multi-basin FEL of the SARS-CoV-2 NSP1 C-terminal domain. Transitioning to membrane-embedded systems, an autoencoder-driven CV pipeline successfully delineates the Minimum Free Energy Path of a folding bacteriorhodopsin fragment. Crucially, Graph Attention Networks (GATs) are pioneered to mathematically predict the nonlinear topological deformations of the lipid leaflet in response to the protein’s dynamic conformational flux. Finally, to decode the microscopic origins of these macroscopic transitions, the Residue Importance Projection (RIP) framework leverages the Jensen-Shannon divergence to systematically isolate state-distinguishing residues without heuristic biases. Applied to Abl and Src kinases, RIP maps longrange allosteric networks and provides a structural rationale for oncogenic inhibitor resistance, laying the conceptual foundation for dynamic “RIPhylogenetic” trees. Collectively, this thesis establishes a highly scalable paradigm, enhancing computational microscopy with machine learning to bridge localized thermodynamic sampling with global biological function.

Actions (login required)

View Item