A Practical Guide to Transition State Analysis in Biomolecular Simulations with TS-DAR

Eshani C. Goonetilleke, Bojun Liu, Yue Wu, Michael S. O’Connor, Xuhui HuangThe Journal of Physical Chemistry B 2025 · DOI

TL;DR

This paper introduces TS-DAR, a novel deep learning framework that uses out-of-distribution detection to automatically identify transition states in biomolecular simulations. By mapping molecular conformations to a hyperspherical latent space and employing specific loss functions, TS-DAR offers a more comprehensive view of protein dynamics than traditional methods, aiding in drug discovery and understanding disease mechanisms.

What Problem Does This Paper Address?

Proteins function by undergoing conformational changes, which often involve navigating through high-energy, short-lived "transition states." While methods like Markov State Models (MSMs) are good at identifying stable protein states (metastable states), they struggle to pinpoint these crucial transition states. This gap makes it difficult to fully understand the mechanisms behind protein function, drug interactions, and the effects of mutations. The authors aim to bridge this gap by developing a computational framework that can systematically and accurately identify all transition states within complex biomolecular simulations.

Methods

The core of the TS-DAR framework is a deep learning model that maps protein conformations from molecular dynamics (MD) simulations into a low-dimensional "hyperspherical latent space." This space is designed to retain the essential kinetic information of the protein's movements.

The TS-DAR model employs a two-part loss function:

  • VAMP-2 Loss: This component ensures that conformations belonging to the same metastable state are clustered together in the latent space, effectively capturing the system's slowest dynamics.
  • Dispersion Loss: This function encourages the metastable state centers to be evenly distributed across the hypersphere. This is crucial because transition states are inherently "out-of-distribution" compared to the well-characterized metastable states. By forcing the metastable states apart, the model can more easily identify conformations that lie between them as transition states.

The framework also incorporates "out-of-distribution" (OOD) detection principles. Conformations with high OOD scores, indicating they deviate significantly from the identified metastable states, are flagged as potential transition states. The paper also details a workflow involving MD sampling, feature selection (using methods like spectral OASIS or MOSAIC), TS-DAR modeling, and optional Markov State Model construction for kinetic analysis.

Datasets

The paper uses three primary datasets to demonstrate the TS-DAR framework:

  • Alanine Dipeptide: A small, well-studied molecule often used as a benchmark for conformational dynamics. The authors used trajectories with a saving interval of 1 ps.
  • Villin Headpiece (HP35): A larger protein system for which they used a 300 μs all-atom trajectory with a saving interval of 200 ps. For this dataset, they followed a protocol involving building a microstate MSM to determine the appropriate number of macrostates.
  • Protein Phosphatase 2A (PP2A): A key enzyme involved in cellular processes. They used ten 100 ns all-atom trajectories with a saving interval of 10 ps. For this system, they employed spectral OASIS for feature selection, reducing the initial 26,565 pairwise distances to 1,000 features.

Key Insights and Ideas

  • Transition States as "Out-of-Distribution" Data: The core insight is reframing transition states not as rare events to be found, but as data points that are fundamentally different from the common, stable states. This allows the application of powerful OOD detection techniques from machine learning.
  • Hyperspherical Latent Space for Kinetic Information: Mapping conformations onto a hypersphere is a clever way to preserve kinetic information while achieving dimensionality reduction. The angular separation on the hypersphere directly relates to the kinetic landscape.
  • Automated and Systematic Transition State Identification: TS-DAR moves beyond manual identification or relying on specific reaction coordinates. It aims to automatically find all transition states, providing a more complete picture of the conformational landscape.
  • Bridging Deep Learning and Biomolecular Dynamics: The paper effectively demonstrates how advanced deep learning techniques can be adapted and applied to solve long-standing challenges in computational biophysics.

Key Results

  • TS-DAR successfully identified transition states in all tested systems, including alanine dipeptide, villin headpiece, and protein phosphatase 2A.
  • For the AlkD translocation system (mentioned in the introduction), TS-DAR uncovered novel insights into the role of protein-DNA hydrogen bonds in the rate-limiting step.
  • The paper claims TS-DAR outperformed previous methods like MaxEnt-VAMPNets and MSM-committor in terms of accuracy and efficiency for transition state identification.
  • The framework allows for the construction of Markov State Models using the identified states, enabling accurate kinetic modeling and validation via Chapman-Kolmogorov tests.
  • For PP2A, TS-DAR could distinguish subtle conformational differences between open and closed active site states, highlighting its ability to capture fine-grained dynamics.

Limitations and Future Work

  • Setting the OOD Threshold: While the paper provides a default threshold calculation, users may need to empirically optimize it or use prior knowledge to set it, especially for complex systems.
  • Committor Analysis for Validation: The authors recommend using committor analysis to verify the identified transition states, indicating that TS-DAR provides candidates that require further validation.
  • Feature Selection: While methods like spectral OASIS are discussed, the paper suggests that more advanced techniques like equivariant neural networks could offer more efficient and robust feature extraction, especially for large systems.
  • Enhanced Sampling: The authors propose using the latent space representations derived from TS-DAR as collective variables for enhanced sampling methods like metadynamics, which could further accelerate the exploration of rare events.
  • Non-Markovian Models: Integrating TS-DAR with generalized master equation (GME) based models is suggested for a more accurate representation of non-Markovian dynamics.

Summary updated

Abstract

Conformational changes essential for protein function involve transitions through multiple short-lived, high-energy states within the complex free energy landscape. While existing methods, such as Markov State Models and non-Markovian approaches built from molecular dynamics (MD) simulations, can effectively capture metastable states, they struggle to identify transition states. Transition state identification via dispersion and variational principle regularized neural networks (TS-DAR) is a computational framework that utilizes out-of-distribution detection to systematically identify all transition states involved in specific biomolecular conformational changes. TS-DAR leverages a deep learning model to map protein conformations from MD simulations onto a hyperspherical latent space. This low-dimensional representation retains the critical kinetic information of biomolecular conformational changes. To distinguish metastable states from transition states, TS-DAR utilizes a VAMP-2 and dispersion loss function, enabling the automated identification of transition state conformations. This framework provides a comprehensive view of protein conformational landscapes and facilitates studies on drug binding, enzyme activity, and mutation effects. This tutorial aims to guide researchers through the implementation and application of TS-DAR, highlighting its utility in computational biophysics. The code for this tutorial is available at: https://github.com/xuhuihuang/ts-dar-tutorials.
BibTeX
@article{goonetilleke2025practical,
  author = {Eshani C. Goonetilleke and Bojun Liu and Yue Wu and Michael S. O’Connor and Xuhui Huang},
  title = {{A Practical Guide to Transition State Analysis in Biomolecular Simulations with TS-DAR}},
  journal = {The Journal of Physical Chemistry B},
  year = {2025},
  doi = {10.1021/acs.jpcb.5c06097},
}

Loading related content...

Figure