Cellular membranes consist of mixed collections of lipids, steroids, and membrane proteins. The compositional heterogeneity, along with the morphology the of cellular membranes has been shown to be essential for the regulation of cellular activity through membrane-protein interactions. For example, Parkinson's disease has been attributed to the aggregation of alpha-synuclein after binding to the surface of synaptic vesicles. Another example is that the fusion peptide on the surface of HIV virion specifically targets the phase boundaries on cellular membranes. A better understanding of the natural of membrane-protein interactions plays a key role in tackling these diseases. Our group employs molecular dynamics simulations with enhanced sampling methods to investigate the impact of the heterogeneity and morphology of cellular membranes. The figure above is from an all-atom vesicle simulation run on the Anton2 supercomputer. The spherical voronoi diagram allows us to obtain information on the packing (area per molecule) of the lipids and the nearest neighbors, which relate to the possible phase seperation on highly curved membranes.

Direct dynamics (DD) simulations of chemical reactions reveal their complicated dynamics at atomistic level, elucidates discoveries from experiments that are nonintuitive, and predicts behaviors of chemical reactions whose conditions are difficult to realize. In DD, the interactions between atoms (i.e. energy gradient, corresponding to forces acting on atoms) are calculated on-the-fly with ab initio methods and their positions (i.e. configuration) are propagated iteratively by solving the classical equations of motion over a small time interval. The time-evolution of the configurations of the system makes up a trajectory, which usually involves thousands of ab initio energy gradient calculations. A large number of trajectories are necessary to accurately model the reaction in real-life: for example, in bimolecular collision, trajectories need to be sampled at various impact parameters with reactant molecules initiated at random orientations. The enormous amount of the computation (i.e. millions of ab initio energy gradient calculations) inevitably forces DD to settle with limited number of trajectories which was generated from some low-level ab initio method with a very limited basis set. The lack of accuracy in forces and ergodicity in sampling sometimes could lead to unphysical dynamics of DD. We are developing an active learning algorithm that is coupled with DD to drastically cut down the number of ab initio energy gradient calculations per trajectory as the simulations goes on. With this algorithm, whenever an ab initio energy gradient is calculated, instead of being thrown away, it is stored in a database (i.e. library). The library grows as the simulation progresses. All configurations encountered in a new trajectory will be checked against the library to identify similar configurations whose energy gradients have already been computed. When possible, their gradients are calculated from a machine-learning algorithm that we are currently developing. The library is uniquely indexed to minimize the cost of configurations search and retrieval, as a result the cost of IMRR energy gradient is only a small fraction of normally-employed ab initio energy gradient. With actively learning coupled DD, the average cost of a new trajectory decreases monotonically: as the library grows larger, a smaller portion of the trajectory requires ab initio energy gradient calculations, and eventually, the computation cost of a new trajectory plateaus to the cost of molecular-mechanics like. The figure above shows a conceptual work-flow of the algorithm.

Intestinal permeability of small molecule drug measured in vivo is a reliable indicator of its bioavailability, but the measurement itself involves an invasive procedure. We are developing innovative molecular dynamics simulation protocols of assessing intestinal permeability of small molecule drug, whose accuracy and reliability significantly improves from the current alternative in vitro and in silico methods. This protocol employs the state of art enhanced sampling method to compute a multidimensional free energy surface of the permeation, and a novel statistical model to measure the diffusivity of the drug on this surface. The capability of this protocol can be assessed by the permeation of five small molecule drugs of various size and lipophobicity, and the results demonstrated a very strong correlation to their intestinal permeabilities. More importantly, this protocol reveals the interplay between the structure of the small molecule and its permeability, inspiring new strategies of drug discovery and development.

A fundamental understanding of the dynamics of elemental chemical reactions plays an essential role in probing into the mechanisms of more complicated ones. Gas phase reaction, where the densities of the reactants are extremely low, is the ideal system for studying reaction dynamics and has attracted attention from chemists over a few decades. Direct dynamics simulations of chemical reactions in gas phase have shown great success in revealing their complicated dynamics at atomistic level, elucidating discoveries from experiments that are nonintuitive, and predicting behaviors of chemical reactions whose conditions are difficult to realize. Our group employs direct dynamics simulation to study chemical reactions taking place in deep space to help understanding how large molecules are formed. We also study how different excitations of reactant molecules could impact the dynamics of the reaction. The video above is produced from our research of the bimolecular collision of HBr+ + CO2 where it shows the product, HOCO+ + Br, is formed after the trajectories spending sometime in the intermediates.