Deeper Dive
My project addresses a central challenge in computational drug discovery. Molecular dynamics (MD) simulations can provide detailed information about how proteins move and how drugs interact with them, but conventional simulations are often too slow to capture rare events such as cryptic pocket opening and changes in ligand binding modes. I became interested in this problem while working on a separate drug discovery project involving potential inhibitors of the Receptor for Advanced Glycation End Products. After designing candidate molecules, I wanted to use MD simulations to determine whether they would actually bind to the protein, but the simulations were far too expensive for my laptop. This introduced me to the field of enhanced sampling, and I started to see the limitations of existing techniques.
I developed diffusion hybridized molecular dynamics (dhMD), in which a diffusion model proposes protein backbone torsion updates that are then screened with a Metropolis-Hastings criterion. The correction allows the learned model to make useful, state-dependent proposals while preserving the correct equilibrium distribution. On alanine dipeptide, dhMD produced a 68-fold improvement in effective sample size per GPU hour for the slow phi dihedral compared with classical MD. I then extended the idea by integrating diffusion into Binding Modes of Ligands Using Enhanced Sampling (BLUES), an enhanced sampling technique that works by randomly rotating the ligand to increase sampling of ligand binding modes. I created diffusion hybridized BLUES (dhBLUES), which converged to equilibrium ligand binding mode populations more than four times faster than BLUES on T4 lysozyme with toluene.
One of the most difficult problems arose when I tried to calculate the forward and reverse proposal probabilities required by the Metropolis Hastings criterion. If the model proposed directly from the current state, the simulations would become biased. I chose to make the model predict the parameters of an explicit proposal distribution from which the actual move could be sampled rather than determine the move itself. This allowed me to calculate the forward and reverse probabilities exactly and preserve the validity of the acceptance criterion. I encountered another problem in dhMD when large proposals caused molecular overlaps and extremely high energies, leading to very low acceptance rates. That failure motivated dhBLUES, which uses Nonequilibrium Candidate Monte Carlo so that the proposed ligand motion happens through smaller steps while allowing the surrounding environment to relax.
This work contributes to better and more efficient drug discovery. Protein conformations and ligand binding modes are necessary for understanding how strongly a molecule binds and how selective it is. If simulations can reach equilibrium statistics using substantially less computational time, researchers can evaluate more molecules and spend less time waiting for simulations. This can make simulation methods accessible to academic laboratories that cannot afford enormous computational budgets. Faster sampling also allows researchers to study important events that conventional simulations rarely observe, including cryptic pocket opening and rearrangements of solvent. The goal of this work was not only to make existing sampling techniques faster, but to show a general framework in which generative models can propose molecular transitions that accelerate simulations. This diffusion hybridization approach can be applied similarly to other enhanced sampling methods.