Abstract
Identifying transitional states is crucial for understanding protein conformational changes that underlie numerous biological processes. Markov state models (MSMs), built from Molecular Dynamics (MD) simulations, capture these dynamics through transitions among metastable conformational states, and have demonstrated success in studying protein conformational changes. However, MSMs face challenges in identifying transition states, as they partition MD conformations into discrete metastable states (or free energy minima), lacking description of transition states located at the free energy barriers. Here, we introduce Transition State identification via Dispersion and vAriational principle Regularized neural networks (TS-DAR), a deep learning framework inspired by out-of-distribution (OOD) detection in trustworthy artificial intelligence (AI). TS-DAR offers an end-to-end pipeline that can simultaneously detect all transition states between multiple free minima from MD simulations using the regularized hyperspherical embeddings in latent space. The key insight of TS-DAR lies in treating transition state structures as OOD data, recognizing that they are sparsely populated and exhibit a distributional shift from metastable states. We demonstrate the power of TS-DAR by applying it to a 2D potential, alanine dipeptide, and the translocation of a DNA motor protein on DNA, where it outperforms previous methods in identifying transition states.
Introduction
Proteins are dynamic molecular machines that transition between various "metastable" states to perform biological functions. Understanding the transition states (TS)—the high-energy, fleeting structures located at the peaks of energy barriers—is fundamental for drug design and enzyme engineering. However, identifying these states is notoriously difficult. Standard tools like Markov State Models (MSMs) are designed to cluster data into stable minima. Because transition states are sparsely populated and short-lived in simulations, MSMs often force them into stable clusters, effectively "blurring" the energy barrier.
To address this, we present TS-DAR, a deep learning framework that treats rare protein transition states as "out-of-distribution" (OOD) data. By mapping molecular dynamics into a specialized hyperspherical latent space, the model can simultaneously and automatically identify multiple transition states between stable protein shapes.
Methodology: The TS-DAR Framework
TS-DAR shifts the perspective of transition state detection by framing it as an outlier detection problem. The framework utilizes a neural network encoder to project protein conformations onto a hyperspherical latent space.
The model is optimized using a dual-loss function:
- VAMP-2 Loss: A kinetic loss based on the variational approach for Markov processes that compresses conformations belonging to the same stable state together.
- Dispersion Loss: A novel regularization term that pushes the centers of these stable state clusters as far apart as possible on the hypersphere.
By tightly packing stable states and maximizing the distance between them, any data point that lands in the "empty space" between clusters is flagged with a high OOD score. These outliers represent the transition states.
Experimental Results
We validated TS-DAR on three systems of increasing complexity: a 2D Müller potential, alanine dipeptide, and the AlkD DNA motor protein.
Simultaneous Barrier Detection
In the Müller potential—a classic energy landscape with three minima—TS-DAR successfully identified two distinct transition regions. Traditional methods often fail here, either misidentifying energy basins as barriers or failing to capture multiple transitions simultaneously.
Efficiency and Mechanistic Discovery
For Alanine dipeptide, TS-DAR identified the transition state using only 3 ns of simulation data, a significant improvement over the ~77 ns required by Transition Path Sampling (TPS).
In the complex AlkD-DNA system (over 725,000 conformations), TS-DAR revealed that the rate-limiting step in DNA repair involves a specific protein rotation that breaks approximately five hydrogen bonds. This provides a molecular explanation for kinetic rates that were previously observed but not mechanistically understood.
Discussion and Future Work
The "aha moment" of TS-DAR is the realization that transition states exhibit a distributional shift from metastable states. Using a hypersphere as a regularizer is critical; as a finite, closed space, it prevents the dispersion term from pushing clusters to infinity, instead forcing a uniform arrangement that reflects the system's underlying kinetics.
While highly efficient, TS-DAR requires that trajectories cross transition barriers at least a few times; it cannot identify states never visited during simulation. A promising future direction is using the OOD score as a "heat map" for adaptive sampling. By initiating new simulations from high-OOD conformations, researchers can more efficiently explore the most critical, unknown regions of a protein's conformational landscape.
Summary updated