Session T44: Harnessing the Power of Machine Learning in Studying Biomolecular Dynamics

Learning mechanisms of rare events from short-trajectory data

Understanding mechanisms of rare events requires estimating statistics such as expected hitting times, rates, and committors. In systems with well-defined metastable states and free energy barriers, these quantities can be estimated using enhanced sampling methods combined with classical rate theories. However, calculating such statistics for more complex processes with rugged landscapes and/or multiple pathways requires more general numerical methods. In this lecture, I will describe my group's recent efforts to develop both linear (Galerkin) and nonlinear (neural-network) methods for estimating transition-path statistics by combining information from many short molecular dynamics trajectories.   

Target structural based de novo drug generation

De novo drug design explores new chemical space and is not limited by known chemical libraries, which has been considered as the "holy grail" of drug discovery. Previously we have developed LigBuilder software series that can design target binding compounds that can be synthesis reachable with preferable properties. LigBuilder has been used by many academic and commercial users, who reported successful design examples that have been experimentally verified. Along with the rapid development of deep learning methods, various new AI techniques were introduced into the drug design field. We are working on methods that can take advantage of both physical based and AI based models. Deep generative models have gained much attention in de novo drug design in recent years. However, directly generating binding molecules inside the target site remains difficult. We developed DeepLigBuilder that can directly design 3D molecules inside the target binding pocket, by combining a novel 3D deep generative network for drug-like molecules and an optimization module that performs Monte Carlo Tree Search. Compounds designed using DeepLigBuilder for several targets show promising activity in experimental studies.   

Bayesian Hi-C metainference of chromatin structure via molecular dynamics simulations

Hi-C and related experiments provide information on contact frequencies in folding chromatin. While chromatin fold highly heterogeneously, these experiments give ensemble-averaged data. We introduce the Hi-C metainference method to infer the heterogeneous structure ensemble from the Hi-C data with the prior distribution constructed from molecular dynamics simulations and Markov state modeling. We test how the static structural ensemble as well as dynamics can be inferred in this process.   

Graph deep learning locates magnesium ions in RNA

Magnesium ions (Mg2+) are crucial for RNA structure and cellular functions, yet locating them accurately in RNA has been a challenge. We introduce a machine-learning method, originally designed for computer vision, to predict Mg2+ binding sites in RNA by considering geometric and electrostatic RNA features. Through deep learning, we accurately predict Mg2+ density distribution. Validated with five-fold cross-validation on a dataset of 177 Mg2+-containing structures and compared with other methods, our approach demonstrates significantly improved accuracy and efficiency. Saliency analysis reveals essential coordinating atoms and uncovers new Mg2+ binding motifs. Combining this approach with X-ray crystallography can identify metal ion binding sites in RNA, advancing our understanding of RNA structure and function.