Lipid nanoparticles (LNPs) are increasingly recognized as effective vehicles for diverse therapeutic agents, including nucleic acids, small molecules, and gene therapies. The design and optimization of LNPs require detailed molecular-scale characterization, which is experimentally challenging. Molecular dynamics (MD) simulation is a powerful computational method for investigating the structural dynamics of biomolecules, such as proteins, lipid nanoparticles. However, simulating those systems often requires a long time to capture the rare biological events, even on state-of-the-art supercomputers. This downside of MD simulations is a persistent challenge in the computational field. To address these limitations, machine learning—and more specifically, deep learning (DL)—has emerged as a promising paradigm to complement the traditional simulation approaches for intricate biological systems such as LNPs. This thesis integrates two complementary research directions to address limitations in computational modeling: (1) application of atomic-level MD simulations to investigate and optimize the design of LNPs as drug delivery carriers and (2) development of DL models to generate synthetic biomolecular trajectories that reproduce the statistical features of MD data by reconstructing the free energy surface. For starting, a basic DL model, a multilayer perceptron (MLP), was trained and validated on the data by using overdamped Langevin dynamics on a one-dimensional double-well potential surface. The model has been tested on the training data set to reconstruct the free energy surface. The benchmarked model will next be applied to the MD trajectories of LNPs to capture the stochasticity of these large multicomponent systems. This strategy of combining both MD simulation and DL frameworks can accelerate the characterization of LNPs and reduce computational demands.
Fauzia Haque (Sun,) studied this question.