Unveiling Switching Function of Amino Acids in Proteins Using a Machine Learning Approach

Parisa Mollaei, Amir Barati FarimaniJournal of Chemical Theory and Computation 2023 · DOI

TL;DR

This paper introduces a novel machine learning (ML) framework to analyze the dynamics of individual amino acid residues within proteins. By leveraging molecular dynamics (MD) simulations, the approach identifies "switching" residues that adopt distinct conformational states and classifies them into stable (SS) and unstable (US) types, revealing a correlation between SS residue dynamics and overall protein function.

What Problem Does This Paper Address?

Understanding the precise structural dynamics of individual amino acid residues within proteins is a significant challenge. While experimental techniques like X-ray crystallography and NMR provide snapshots or average behaviors, they often lack the temporal and spatial resolution to capture the rapid, fleeting conformational changes that occur at the gigahertz range. Molecular dynamics (MD) simulations offer a powerful computational alternative, providing atomistic detail over time. However, analyzing the vast datasets generated by MD simulations, especially for rare conformational switching events, is computationally intensive and complex. This paper aims to bridge this gap by developing an automated, ML-driven method to identify and characterize these dynamic residue behaviors, linking them to broader protein functions.

Methods

The authors developed a two-stage framework:

  1. Data Preprocessing:
    • Atom Selection: A subset of key atoms within each amino acid residue is selected to reduce the dimensionality of the data while retaining essential structural information.
  2. Angle Calculation: Angles formed by combinations of three selected atoms are calculated for each frame of the MD trajectory.
  3. Density of States (DoS): Histograms of these angles are generated to represent the "density of states" for each angle, illustrating the preferred conformational angles of the residue. Residues exhibiting bimodal DoS (two distinct peaks with no intermediate states) are flagged as potential "switch" residues.
  4. Machine Learning Processing:
    • Feature Engineering: 14 features are engineered from the DoS patterns to capture the characteristics of bimodal switch histograms.
  5. Classification of Switch Residues: A Random Forest (RF) model is trained to classify residues as either "switch" or "non-switch" based on these features.
  6. Classification of Stability (SS vs. US): A separate framework is developed to distinguish between Stable Switch (SS) and Unstable Switch (US) residues. This involves calculating an "Instability Ratio" (ratio of transitions between states to trajectory length). A Logistic Regression model is then trained using this ratio to classify SS and US residues.

Datasets

The framework was applied to trajectory data from two proteins:

  • Fs peptide protein: A well-established model system for protein folding studies. The trajectory used was 500 ns long.
  • β₂AR receptor: A G protein-coupled receptor (GPCR) involved in signal transduction. Trajectories were generated for both inactive and active crystal structures.

The paper does not introduce new benchmark datasets but rather applies the developed framework to existing MD simulation data.

Key Insights and Ideas

  • Automated Identification of Dynamic Residues: The ML framework provides an automated and efficient way to identify residues that undergo significant conformational switching, a task that is challenging with manual analysis of MD trajectories.
  • Distinguishing Dynamics from Statics: The DoS provides a static picture of preferred angles, but the ML approach goes further by analyzing the dynamics of these switches. This allows for the classification of switch residues into stable (SS) and unstable (US) types, revealing that residues with similar DoS can have vastly different dynamic behaviors.
  • Correlation with Global Protein Function: A crucial finding is the correlation between the dynamics of SS residues and the overall functional properties of the protein. This suggests that these specific dynamic residues play a key role in mediating protein function.
  • Computational Efficiency: By first identifying SS residues, subsequent analysis can focus on a smaller subset of residues, making the investigation of protein-wide correlations more computationally tractable.

Key Results

  • Switch Residue Classification: The Random Forest model achieved a high classification accuracy of 96.94% in identifying switch residues.
  • Stability Classification: The Logistic Regression model achieved an accuracy of 98.97% in distinguishing between Stable Switch (SS) and Unstable Switch (US) residues.
  • Fs Peptide Protein: Seven residues were classified as switch residues, with one (A9) identified as an SS residue. A strong correlation was observed between the A9 residue's dynamics and the protein's Root Mean Square Deviation (RMSD), a measure of protein folding.
  • β₂AR Receptor: The S143 residue was identified as an SS residue, showing a strong correlation with the receptor's activation states. The authors demonstrated that as the angle of S143 changes, the average activation level of the receptor increases significantly. Conversely, US residues showed no such correlation.

Limitations and Future Work

  • Focus on Bimodal Switches: The current framework primarily focuses on residues exhibiting bimodal (two-state) switching. The authors acknowledge that the approach could be extended to n-modal switching.
  • Computational Cost of MD: While the ML framework makes analysis more efficient, the initial generation of MD trajectories, especially for long timescales or large proteins, remains computationally expensive.
  • Future Directions: The authors suggest that this framework can be applied to protein engineering, drug discovery, and understanding complex biological processes like signal transduction in GPCRs. Further investigation into the specific roles of SS residues in various protein functions is warranted. The ability to correlate residue dynamics with global properties opens avenues for designing more effective therapeutics by targeting these key dynamic elements.

Summary updated

Abstract

Dynamics of individual amino acids play key roles in the overall properties of proteins. However, the knowledge of protein structural features at the residue level is limited due to the current resolutions of experimental and computational techniques. To address this issue, we designed a novel machine-learning (ML) framework that uses Molecular Dynamics (MD) trajectories to identify the major conformational states of individual amino acids, classify amino acids switching between two distinct modes, and evaluate their degree of dynamic stability. The Random Forest model achieved 96.94% classification accuracy in identifying switch residues within proteins. Additionally, our framework distinguishes between the stable switch (SS) residues, which remain stable in one angular state and jump once to another state during protein dynamics, and unstable switch (US) residues, which constantly fluctuate between the two angular states. This study also illustrates the correlation between the dynamics of SS residues and the protein's global properties.
BibTeX
@article{mollaei2023unveiling,
  author = {Parisa Mollaei and Amir Barati Farimani},
  title = {{Unveiling Switching Function of Amino Acids in Proteins Using a Machine Learning Approach}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2023},
  doi = {10.1021/acs.jctc.3c00665},
}
Figure