A Stochastic Landscape Approach for Protein Folding State Classification

Michael Faran, Dhiman Ray, Shubhadeep Nag, Umberto Raucci, Michele Parrinello, Gili BiskerJournal of Chemical Theory and Computation 2024 · DOI

TL;DR

This paper introduces the Stochastic Landscape Classification (SLC), a novel, unsupervised algorithm that automatically identifies and classifies protein states during folding. By analyzing trends in low-dimensional collective variables (CVs) derived from molecular dynamics simulations, SLC effectively segments and clusters protein conformations, offering a powerful new tool for understanding complex biomolecular dynamics.

What Problem Does This Paper Address?

Understanding how proteins fold into their functional three-dimensional structures is fundamental to biology and medicine. Misfolded proteins are implicated in numerous diseases, making the study of protein folding a critical area of research. Molecular dynamics (MD) simulations are powerful tools for observing these processes, but interpreting the vast amounts of data generated can be challenging. Traditional methods often rely on pre-defined collective variables (CVs) that may not fully capture the complex dynamics, or they use machine learning approaches that require extensive training data. This paper addresses the need for an automated, unsupervised method that can accurately classify protein states from MD trajectories without prior knowledge of the number of states or extensive training, while also providing insights into the quality of different CVs.

Methods

The core of the proposed method is the Stochastic Landscape Classification (SLC). It operates in several key steps:

  1. CV Trajectory Input: The process begins with a time series of a chosen collective variable (CV) from an MD simulation.
  2. BEAST Algorithm for Segmentation: The Bayesian Estimator of Abrupt change, Seasonality, and Trend (BEAST) algorithm is applied to the CV trajectory. BEAST identifies points where the statistical properties of the CV change significantly, effectively segmenting the trajectory into distinct periods. This study specifically focuses on trend changes, assuming that within metastable states, the CV fluctuates around a stable mean, while transitions between states exhibit a trend.
  3. Stochastic Coordinate Extraction: For each segment identified by BEAST, statistical features are extracted, including the mean, standard deviation (STD), and average trend. These are termed "stochastic coordinates."
  4. Normalization and PCA: The extracted stochastic coordinates are normalized to ensure comparability across segments. Principal Component Analysis (PCA) is then used for dimensionality reduction, creating a lower-dimensional "Stochastic Landscape."
  5. DBSCAN Clustering: The Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm is applied to the points in the Stochastic Landscape. DBSCAN groups segments with similar statistical properties into clusters, representing distinct protein states.
  6. Label Matching and Output: The identified clusters are optionally matched to known "ground-truth" states using the Kuhn-Munkres algorithm. Finally, these labels are mapped back to the original CV trajectory, providing a time series of classified protein states.

The novelty lies in combining BEAST for data-driven segmentation with DBSCAN for unsupervised clustering based on statistical features of these segments, rather than relying on direct coordinate values or pre-trained models.

Datasets

The study utilizes existing molecular dynamics simulation data for two well-characterized proteins:

  • Chignolin: A small, fast-folding peptide. The analysis used a 106 µs unbiased trajectory.
  • Trp-Cage: A slightly larger peptide with a more complex folding landscape, including a misfolded state. The analysis used a ~268 µs unbiased trajectory, which was downsampled.

The "ground-truth" data for evaluating the SLC's performance was derived from these simulations using specifically designed collective variables, such as TPI-Deep-TDA for Chignolin and 3-state Deep-TDA for Trp-Cage, which are known to effectively distinguish between protein states.

Key Insights and Ideas

  • Unsupervised State Identification: The SLC method demonstrates that protein states can be effectively classified without prior knowledge of the number of states or requiring a supervised learning approach. This is a significant advantage for exploring unknown or complex biological systems.
  • Kinetic Information is Key: By segmenting based on trend changes (via BEAST), the SLC implicitly incorporates kinetic information about transitions between states, which is often missing in purely structural clustering methods.
  • CV Quality Matters: The study highlights that the choice of collective variable significantly impacts classification accuracy. While advanced, machine-learning-derived CVs often yield better results, the SLC can still achieve respectable performance with simpler, intuitive CVs, suggesting its robustness.
  • Statistical Features as Fingerprints: The core idea that segments of a CV trajectory corresponding to the same protein state share similar statistical properties (mean, STD, trend) is a powerful concept for state identification.

Key Results

The SLC approach demonstrated high accuracy in classifying protein states for both Chignolin and Trp-Cage proteins.

  • Chignolin: For Chignolin, the SLC achieved classification accuracies of up to ~90% when using an optimal CV (TPI-Deep-TDA). Even with simpler CVs, the classification metrics (Rand Index, Adjusted Rand Index, etc.) generally showed high values, indicating good agreement with the ground truth.
  • Trp-Cage: For Trp-Cage, which has three states (folded, unfolded, misfolded), the SLC also performed well, particularly with advanced CVs like 3-state Deep-TDA. While some misclassification of the misfolded state was observed, the overall accuracy was high.
  • Robustness: The method showed good performance across various CVs, demonstrating its general applicability. The percentage of unclassified data points was found to correlate negatively with classification accuracy, suggesting that refining the DBSCAN parameters could further improve results.

Limitations and Future Work

  • State Visitation Requirement: A key prerequisite for SLC is that the protein must visit each relevant metastable state multiple times within the trajectory. If a state is rarely visited or has very short dwelling times, classification accuracy may suffer.
  • CV Dependence: While the method is robust, the quality of the chosen CV remains a critical factor. The paper suggests that refining CVs can lead to improved classification.
  • Unclassified Data: The presence of unclassified data points, particularly in more complex systems or with less informative CVs, indicates areas where the clustering might be improved.

Future work suggested by the authors includes:

  • Application to More Complex Systems: Extending the methodology to larger, more complex proteins with multiple states, and to intrinsically disordered proteins (IDPs).
  • Experimental Data Integration: Adapting the SLC to analyze time-series data from experimental techniques like single-molecule force spectroscopy.
  • Algorithm Refinement: Exploring other unsupervised clustering algorithms and optimizing DBSCAN parameters to reduce unclassified data and improve accuracy.
  • Handling Non-Ergodic Dynamics: Developing strategies to apply SLC to short trajectories with non-ergodic dynamics.

Summary updated

Abstract

Protein folding is a critical process that determines the functional state of proteins. Proper folding is essential for proteins to acquire their functional three-dimensional structures and execute their biological role, whereas misfolded proteins can lead to various diseases, including neurodegenerative disorders like Alzheimer's and Parkinson's. Therefore, a deeper understanding of protein folding is vital for understanding disease mechanisms and developing therapeutic strategies. This study introduces the Stochastic Landscape Classification (SLC), an innovative, automated, nonlearning algorithm that quantitatively analyzes protein folding dynamics. Focusing on collective variables (CVs) - low-dimensional representations of complex dynamical systems like molecular dynamics (MD) of macromolecules - the SLC approach segments the CVs into distinct macrostates, revealing the protein folding pathway explored by MD simulations. The segmentation is achieved by analyzing changes in CV trends and clustering these segments using a standard density-based spatial clustering of applications with noise (DBSCAN) scheme. Applied to the MD-based CV trajectories of Chignolin and Trp-Cage proteins, the SLC demonstrates apposite accuracy, validated by comparing standard classification metrics against ground-truth data. These metrics affirm the efficacy of the SLC in capturing intricate protein dynamics and offer a method to evaluate and select the most informative CVs. The practical application of this technique lies in its ability to provide a detailed, quantitative description of protein folding processes, with significant implications for understanding and manipulating protein behavior in industrial and pharmaceutical contexts.
BibTeX
@article{faran2024stochastic,
  author = {Michael Faran and Dhiman Ray and Shubhadeep Nag and Umberto Raucci and Michele Parrinello and Gili Bisker},
  title = {{A Stochastic Landscape Approach for Protein Folding State Classification}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2024},
  doi = {10.1021/acs.jctc.4c00464},
}
Figure