Active Learning of the Conformational Ensemble of Proteins using Maximum Entropy VAMPNets

Abstract

Rapid computational exploration of the free energy landscape of biological molecules remains an active area of research due to the difficulty of sampling rare state transitions in Molecular Dynamics (MD) simulations. In recent years, an increasing number of studies have exploited Machine Learning (ML) models to enhance and analyze MD simulations. Notably, unsupervised models that extract kinetic information from a set of parallel trajectories have been proposed, including the variational approach for Markov processes (VAMP), VAMPNets, and time-lagged variational autoencoders (TVAE). In this work, we propose a combination of adaptive sampling with active learning of kinetic models to accelerate the discovery of the conformational landscape of biomolecules. In particular, we introduce and compare several techniques that combine kinetic models with two adaptive sampling regimes (least counts and multi-agent reinforcement learning-based adaptive sampling) to enhance the exploration of conformational ensembles without introducing biasing forces. Moreover, inspired by the active learning approach of uncertainty-based sampling, we also present MaxEnt VAMPNet. This technique consists of restarting simulations from the microstates that maximize the Shannon entropy of a VAMPNet trained to perform soft discretization of metastable states. By running simulations on two test systems, the WLALL pentapeptide and the villin headpiece subdomain, we empirically demonstrate that MaxEnt VAMPNet results in faster exploration of conformational landscapes compared to the baseline and other proposed methods.

Energetics of substrate transport in proton-dependent oligopeptide transporters

“…Recently, several machine learning based adaptive sampling approaches have been reported. [38][39][40] In this case, we already knew the key reaction coordinates involved in the transport cycle based on our previous simulation study of the PepT So . 30 Therefore, the least count based adaptive sampling combined with the known reaction coordinates is expected to be provide similar efficiency as the current state-of-the-art adaptive sampling methods.…”

Adaptive Sampling Methods for Molecular Dynamics in the Era of Machine Learning

“…We will focus on three studies that have applied DL models to adaptive sampling, 2,76,77 but of course related works exist. 71,78,79 A connection between these three works is that all of them apply a similar simulation-training cycle that consists of launching the simulations, training the model on them, and then using the model to select seeds for new trajectories to restart the loop.…”

“…71,78,79 A connection between these three works is that all of them apply a similar simulation-training cycle that consists of launching the simulations, training the model on them, and then using the model to select seeds for new trajectories to restart the loop. Two of these works 2,76 use different types of variational autoencoders (VAEs) 80 as their base models, while the third one 77 uses primarily VAMP-Nets. 61 Specifically, DeepDriveMD 76 employs a convolutional VAE (CVAE), while latent space-assisted adaptive sampling (LAST) 2 employs a VAE parametrized by a feed-forward network with fully connected layers.…”

“…MaxEnt VAMPNet 77 is the third approach to be discussed. In this technique, rather than identfying outliers in latent space, VAMPNet 61 is used to classify the discovered conformations into metastable states.…”

“…The rationale for selecting the structures based on their entropy is that the structures that cannot be clearly placed into a metastable state correspond to transition states or poorly sampled regions of the phase space. When applied to a small peptide (sequence WLALL 86 ), MaxEnt showed a 3× acceleration in conformational landscape discovery compared to a combination of VAMPNet and count-based adaptive sampling 77 and approximately 60% higher landscape discovery compared to reinforcement-learning-based approaches. 47,48 Interestingly, the trajectories collected with MaxEnt also produced converged MSMs, while those collected with count-based sampling alone did not.…”