Multiscale Molecular Modeling of Biomolecular Self-Assembly: From Molecular Dynamics to Machine Learning

The self-assembly of large biomolecular aggregates underlies a wide spectrum of biological processes, from cytoskeletal filament dynamics to the formation of viral capsids and engineered protein-based biomaterials. These processes are inherently multiscale, involving thousands of interacting subunits, complex conformational rearrangements, and cooperative transitions that span nanometers (nm) to microns (μm), and microseconds (μs) to minutes (min). Capturing such long time- and length-scale events poses significant challenges for both experiment and computation. Experimental approaches often provide only static or ensemble-averaged snapshots, while conventional all-atom molecular dynamics (MD) simulations are computationally prohibitive for assemblies of this large size and complexity. Consequently, it is challenging to elucidate the molecular-level mechanisms that governs the pathways of large-scale biomolecular systems, motivating the development of new computational methodologies and solutions to bridge the gap between atomistic detail and emergent mesoscale phenomena. Multiscale coarse-graining (CG) theory offers a powerful framework to address these challenges by systematically connecting all-atom resolution with coarse-grained models that capture collective behavior over extended spatiotemporal scales. Coarse-graining is indispensable for the study of ultra-large biomolecular systems, as it can reduce the number of degrees of freedom while preserving essential structural and dynamics features. However, the fidelity of CG simulations depends critically on how atomic sites are mapped to CG sites. Existing mapping strategies, many of which rely on sequence contiguity, often fail to represent functionally important domains when residues that move collectively are distant in sequence but adjacent in three-dimensional space. This limitation restricts the ability of CG models to capture cooperative conformational changes and emergent pathways in massive protein complexes. A central challenge, therefore, is to design mapping methods that are both systematic and physically grounded, and that can also leverage emerging machine learning (ML) techniques to extend coarse-graining frameworks to even larger assemblies and more complex dynamics. In this dissertation, I present a body of work that bridges molecular dynamics, multiscale coarse-graining theory, and machine learning for applications to extremely large biomolecular assembly systems. Specifically, we introduce a generalized K-means clustering coarse-graining (KMC-CG) methodology that systematically identifies optimal mappings by incorporating both spatial proximity and correlated motions of residues. This approach removes the constraints of sequence-based algorithms and yields physically intuitive, transferable CG models of unprecedented robustness for very large biomolecules. In addition, we explore how ML concepts, such as clustering, dimensionality reduction, and data-driven acceleration, can be integrated into coarse-graining workflows to enhance efficiency and accuracy. The utility of these methods is demonstrated in two distinct but complementary contexts. First, in collaboration with experimental partners, we employ multiscale simulations with artificial intelligence (AI)–driven approaches to guide the design of megamolecule self-assembly systems, where computational predictions of building block architecture and metal–ligand coordination inform synthesis and characterization. Second, we investigate the dynamic instability of microtubules (MTs), developing multiscale models that accelerate relaxation dynamics and reveal nucleotide-dependent differences in structure, mechanics, and effective interactions at the growing tip. Building on this foundation, we are further constructing a low-resolution CG model for the microtubule system and, subsequently, a polymerizable multiscale model of MT tips that couples CG molecular dynamics with the kinetic Monte Carlo (KMC) framework for subunit addition, GTP hydrolysis, and dissociation. Collectively, these studies highlight how methodological innovations—combining molecular dynamics, coarse-graining, and machine learning—enable efficient and accurate simulations of ultra-large biomolecular assemblies, advancing both the rational design of synthetic self-assembling systems and the mechanistic understanding of dynamic cytoskeletal processes that are fundamental to cellular organization.