General Protein Data Bank-Based Collective Variables for Protein Folding

Albert Ardèvol, Ferruccio Palazzesi, Gareth A. Tribello, Michele ParrinelloJournal of Chemical Theory and Computation 2015 · DOI

TL;DR

This paper introduces a novel method for generating "general" collective variables for protein folding simulations. By using structures from the Protein Data Bank (PDB) as a reference, these new variables can analyze diverse protein systems without needing to be re-derived for each new simulation, overcoming a major limitation of previous dimensionality reduction techniques.

What Problem Does This Paper Address?

Understanding protein folding is crucial for biology and medicine. Molecular dynamics (MD) simulations generate vast amounts of data (trajectories) that are incredibly complex and high-dimensional. To make sense of these trajectories, researchers often use dimensionality reduction techniques to create simpler, "collective variables" that capture the essential motions of the protein. However, existing methods like sketch-map typically require generating these collective variables specifically for each protein system being studied, using data from that system's simulation. This is time-consuming and limits the ability to compare results across different proteins or simulations. This paper addresses the need for a more general approach to collective variables that can be applied across a range of protein systems.

Methods

The core of the paper utilizes the sketch-map algorithm, a dimensionality reduction technique that aims to create a low-dimensional representation of high-dimensional data (like protein configurations) while preserving the relative distances between data points.

The key innovation here is how the "landmarks" for sketch-map are chosen. Instead of selecting landmark configurations directly from the MD trajectory of the protein being studied, the authors use a large dataset of protein structures from the Protein Data Bank (PDB), specifically those determined by NMR. This ensures a broad representation of possible protein conformations.

The process involves:

  1. Selecting PDB structures as landmark configurations.
  2. Generating initial sketch-map projections for these PDB landmarks.
  3. Using these PDB-derived projections as the basis for a new set of collective variables.
  4. Applying these general PDB-based variables to analyze MD trajectories of different proteins.

This approach allows for the creation of "sequence-independent" collective variables.

Datasets

  • Protein Data Bank (PDB): A large repository of experimentally determined protein structures. The authors specifically used NMR-determined structures to capture a wide range of conformational possibilities. They selected 1000 landmark points from segments of these structures.
  • MD Trajectories: The paper evaluates the PDB-based collective variables using MD trajectories from two specific protein systems:
  • A β-hairpin protein (Ace-GEWTYDDATKTFTVTE-NMe) simulated using parallel tempering well-tempered ensemble (PT-WTE) metadynamics.
  • A 9-residue intrinsically disordered peptide (IDP) simulated using two different force fields (AMBER03w and AMBER99SB*-ILDN).

Key Insights and Ideas

  • Generalizability is Key: The most significant insight is that collective variables derived from a diverse, external dataset (like the PDB) can be as effective, and often more transferable, than those derived solely from a specific simulation trajectory.
  • PDB as a Universal Reference: The PDB provides a rich and representative set of protein conformations that can serve as a universal "landmark" set for generating collective variables applicable across many systems.
  • Unbiased Exploration: By using PDB structures, the generated collective variables are less reliant on the specific chemical intuition of the researcher or the limitations of a particular simulation, potentially leading to more surprising discoveries.
  • Force Field Differences Visualized: The PDB-based variables are powerful for comparing the conformational ensembles sampled by different force fields, as demonstrated with the IDP example, revealing distinct structural preferences.

Key Results

  • The PDB-based sketch-map coordinates were shown to be as useful for analyzing MD trajectories as coordinates derived directly from the trajectory data.
  • Histograms generated using PDB-based coordinates successfully identified stable (low free energy) configurations for the studied proteins.
  • A clear one-to-one mapping was established between the low-free energy features identified by trajectory-derived coordinates and those identified by PDB-derived coordinates, confirming the validity of the PDB-based approach.
  • The PDB-based variables were found to be slightly less discriminating than trajectory-based ones for a specific test case, but this was attributed to the broader structural coverage of the PDB dataset, making them more general.
  • The method was successfully applied to visualize and compare the conformational ensembles sampled by different force fields for an intrinsically disordered peptide, highlighting distinct structural behaviors.

Limitations and Future Work

  • Quality of PDB Data: While the PDB is vast, the quality and diversity of NMR structures might influence the generality of the derived variables. The paper notes that the PDB-based coordinates are "slightly less discriminating" in one test, which could be related to the breadth of the PDB dataset.
  • Computational Cost: Generating and applying sketch-map can still be computationally intensive, although the PDB-based approach aims to amortize this cost by creating reusable variables.
  • Future Work: The authors intend to explore using these PDB-based sketch-map variables to enhance sampling in simulations, particularly with methods like metadynamics, by using them as general collective variables. They also plan to make these coordinate sets available for various peptide lengths.

Summary updated

Abstract

New, automated forms of data analysis are required to understand the high-dimensional trajectories that are obtained from molecular dynamics simulations on proteins. Dimensionality reduction algorithms are particularly appealing in this regard as they allow one to construct unbiased, low-dimensional representations of the trajectory using only the information encoded in the trajectory. The downside of this approach is that a different set of coordinates are required for each different chemical system under study precisely because the coordinates are constructed using information from the trajectory. In this paper, we show how one can resolve this problem by using the sketch-map algorithm that we recently proposed to construct a low-dimensional representation of the structures contained in the protein data bank. We show that the resulting coordinates are as useful for analyzing trajectory data as coordinates constructed using landmark configurations taken from the trajectory and that these coordinates can thus be used for understanding protein folding across a range of systems.
BibTeX
@article{ardevol2015general,
  author = {Albert Ardèvol and Ferruccio Palazzesi and Gareth A. Tribello and Michele Parrinello},
  title = {{General Protein Data Bank-Based Collective Variables for Protein Folding}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2015},
  doi = {10.1021/acs.jctc.5b00714},
}
Figure