Choice of Adaptive Sampling Strategy Impacts State Discovery, Transition Probabilities, and the Apparent Mechanism of Conformational Changes

Abstract

Interest in atomically-detailed simulations has grown significantly with recent advances in computational hardware and Markov state modeling (MSM) methods, yet outstanding questions remain that hinder their widespread adoption. Namely, how do alternative sampling strategies explore conformational space and how might this influence predictions generated from the data? Here, we seek to answer these questions for four commonly used sampling methods: 1) a single long simulation, 2) many short simulations run in parallel, 3) adaptive sampling, and 4) our recently developed goal-oriented sampling algorithm, FAST. We first develop a theoretical framework for analytically calculating the probability of discovering select states on simple landscapes, where we uncover the drastic effects of varying the number and length of simulations. We then use kinetic Monte Carlo simulations on a variety of physically inspired landscapes to characterize the probability of discovering particular states and transition pathways for each of the four methods. Consistently, we find that FAST simulations discover each target state with the highest probability, while traversing realistic pathways. Furthermore, we uncover the potential pathology that short parallel simulations sometimes predict an incorrect transition pathway by crossing large energy barriers that long simulations would typically circumnavigate. We refer to this pathology as “pathway tunneling”. To protect against this phenomenon when using adaptive-sampling and FAST simulations, we introduce the FAST-string method. This method enhances sampling along the highest-flux transition paths to refine an MSMs transition probabilities and discriminate between competing pathways. Additionally, we compare the performance of a variety of MSM estimators in describing accurate thermodynamics and kinetics. For adaptive sampling, we recommend simply normalizing the transition counts out of each state after adding small pseudo-counts to avoid creating sources or sinks. Lastly, we evaluate whether our insights from simple landscapes hold for all-atom molecular dynamics simulations of the folding of the λ-repressor protein. Remarkably, we find that FAST-contacts predicts the same folding pathway as a set of long simulations but with orders of magnitude less simulation time.

The Cap-Snatching SFTSV Endonuclease Domain Is an Antiviral Target

“…Then 10 sweeps of a k-medoids update step was used to center the clusters on the densest regions of conformational space. Following clustering, an MSM was built by row-normalizing the observed transition counts, at a lag-time of 2 ns, with a small pseudo-count as a prior (Zimmerman et al, 2018). Radius of gyration and the solvent accessible surface areas for each state in the MSM were calculated using MDTraj and were weighted based on their determined populations (McGibbon et al, 2015).…”

High-Resolution Mining of SARS-CoV-2 Main Protease Conformational Space: Supercomputer-Driven Unsupervised Adaptive Sampling

“…This strategy belongs to the family of counts based adaptive sampling algorithms, where one only exploits the number of passages in the different states (micro or macro) visited in the previous iterations to choose which state to restart trajectories from. These are known to be efficient for pure exploration purpose (as it is the case here), even though more refined algorithms exist when some information is available as to where the sampling should be guided 25 . However, contrary to what is usually done in the context of Markov State Models (MSMs), 24 the states are not defined by applying a clustering algorithm to the already explored structures, but are the projection on the n principal components generated by PCA (here, n = 4 as we discussed) of all the previous data.…”

Conformational distributions of isolated myosin motor domains encode their mechanochemical properties

“…Markov state models were then fit for each variant by applying a 1/ n pseudocount to each element of the transition counts matrix and row-normalizing, as recommended in Zimmerman, et al (Zimmerman et al, 2018). Lag times were chosen by the implied timescales test and by examining the equilibrium probability distribution for unrealistically overpopulated states (suggesting insufficient sampling of a particular transition or internal energy barriers).…”