Abstract
State-free reversible VAMPnets (SRVs) are a neural network-based framework capable of learning the leading eigenfunctions of the transfer operator of a dynamical system from trajectory data. In molecular dynamics simulations, these data-driven collective variables capture the slowest modes of the dynamics and are useful for enhanced sampling and free energy estimation. In this work, we employ SRV coordinates as a feature set for Markov state model (MSM) construction. Compared to the current state-of-the-art, MSMs constructed from SRV coordinates are more robust to the choice of input features, exhibit faster implied time scale convergence, and permit the use of shorter lagtimes to construct higher kinetic resolution models. We apply this methodology to study the folding kinetics and conformational landscape of the Trp-cage miniprotein. Folding and unfolding mean first passage times are in good agreement with the prior literature, and a nine macrostate model is presented. The unfolded ensemble comprises a central kinetic hub with interconversions to several metastable unfolded conformations and which serves as the gateway to the folded ensemble. The folded ensemble comprises the native state, a partially unfolded intermediate "loop" state, and a previously unreported short-lived intermediate that we were able to resolve due to the high time resolution of the SRV-MSM. We propose SRVs as an excellent candidate for integration into modern MSM construction pipelines.
Introduction
Analyzing Molecular Dynamics (MD) simulations of protein folding requires dimensionality reduction to identify metastable states and their transition kinetics. Traditional Markov State Models (MSMs) often rely on linear techniques like Time-lagged Independent Component Analysis (TICA). However, TICA's performance is highly sensitive to the initial choice of features and may fail to capture non-linear dynamical processes. This work presents a hybrid approach that integrates State-free Reversible VAMPnets (SRVs)—a deep learning method—into the MSM pipeline to identify the slowest folding modes with high resolution and robustness.
Methodology
State-free Reversible VAMPnets (SRVs)
SRVs utilize twin-lobed neural networks to learn non-linear transformations of MD data. By maximizing the VAMP-2 score, the network identifies the leading eigenfunctions of the system's transfer operator, which correspond to the slowest dynamical processes. This "state-free" approach allows the network to discover optimal reaction coordinates without a predefined discretization of the conformational space.
Hybrid MSM Construction
The learned SRV coordinates serve as the feature set for a standard MSM construction process. These coordinates are clustered (e.g., via k-means) to define microstates, from which a transition matrix is estimated. This hybrid framework combines the non-linear learning power of deep neural networks with the statistical interpretability of MSMs.
Results and Discussion
Robustness to Feature Selection and Lagtime Convergence
A significant advantage of SRVs is their robustness to the choice of input features. While linear methods often require expert-engineered features to capture slow kinetics, SRVs can reconstruct these modes from simpler inputs like C$\alpha$ distances. Additionally, SRV-MSMs exhibit faster convergence of implied timescales, enabling the use of shorter lagtimes (10 ns) compared to traditional TICA-based models.
High-Resolution Folding Landscape of Trp-Cage
The methodology was applied to a 208 $\mu$s simulation of the Trp-cage miniprotein. The SRV-MSM resolved the leading slow modes more accurately than standard MSM approaches, especially when using restricted feature sets. This increased resolution allowed for the identification of a previously hidden intermediate state, $S_0$, which is structurally similar to the native state but contains a misfolded $3_{10}$ helix loop.
The resulting 9-macrostate model reveals a "kinetic hub" in the unfolded ensemble—a molten globule state ($S_8$)—that acts as the primary gateway to the folded ensemble. The calculated folding and unfolding mean first passage times (14.0 $\mu$s and 3.0 $\mu$s, respectively) are in excellent agreement with experimental data and prior high-quality simulations.
Conclusion
Integrating SRVs into the MSM construction pipeline significantly improves the modeling of complex protein dynamics. By providing a robust, non-linear coordinate system, SRVs enable higher kinetic resolution and the discovery of short-lived intermediate states. This approach is highly recommended for modern MD analysis and offers a powerful way to extract maximal insight from long-duration simulations.
Summary updated