Unsupervised Graph-Based Learning Predicts Mutations That Alter Protein Dynamics

Robert L. Peach, Dominik Šaman, Sophia N. Yaliraki, David R. Klug, Liming Ying, Keith R. Willison, Mauricio BarahonabioRxiv 2019 · DOI

TL;DR

This paper introduces an unsupervised graph-based learning method to predict how mutations affect protein dynamics. By analyzing protein structures as graphs, they can identify key substructures and score mutations based on their disruption of these dynamics, which was experimentally validated using FRET measurements on Adenylate Kinase (ADK).

What Problem Does This Paper Address?

Proteins are dynamic molecules, and their function is intimately tied to how they move and change shape across various timescales, from atomic vibrations to large-scale domain movements. Understanding how mutations alter these dynamics is crucial for deciphering protein function and engineering new ones. However, traditional methods for studying protein dynamics, like molecular dynamics (MD) simulations, are computationally expensive for capturing long timescales. Furthermore, predicting the precise impact of mutations on these complex dynamics has been challenging. This paper aims to bridge this gap by providing a computationally efficient and data-driven approach to identify functionally relevant protein substructures and predict the impact of mutations on these dynamics.

Methods

The core of this work is an unsupervised graph-based learning method called Markov Stability (MS). Here's how it works:

  • Atomistic Graph Construction: Protein structures (from PDB files) are converted into detailed graphs where atoms are nodes and interactions (covalent bonds, hydrogen bonds, hydrophobic interactions) are weighted edges. This retains fine-grained physico-chemical information.
  • Graph Partitioning with Markov Stability: MS uses a random walk on this graph to identify "communities" or substructures that behave similarly over different timescales. It does this by exploring partitions at various "Markov times" (a dimensionless measure of timescale) and identifying partitions that are both robust (consistent across multiple runs) and persistent (stable over a range of Markov times). This allows for the identification of biologically relevant substructures at multiple scales, from chemical groups to large domains, without prior knowledge or coarse-graining.
  • In Silico Mutagenesis: The MS method is applied to computationally generated "mutant" proteins where individual amino acids are replaced (e.g., with alanine or arginine). The disruption caused by each mutation to the identified graph partitions is quantified.
  • Experimental Validation: The computational predictions are then validated using single-molecule Förster Resonance Energy Transfer (smFRET) experiments on Adenylate Kinase (ADK). FRET measures distances between specific points in the protein, allowing researchers to observe changes in domain positions (e.g., open vs. closed states).

The novelty lies in applying MS to atomistic protein graphs for unsupervised identification of multiscale substructures and then using this framework to systematically score the impact of mutations on these dynamics.

Datasets

  • Protein Structures: Two X-ray crystal structures of Aquifex Adenylate Kinase (ADK) were used: the unliganded open conformation (PDB ID: 2RH5) and the closed conformation (PDB ID: 2RGX).
  • Experimental Data: Single-molecule FRET (smFRET) measurements were performed on wild-type ADK and various D152 and V164R mutants. Luciferase activity assays were also conducted.

The paper does not introduce new benchmark datasets but rather uses existing protein structures and generates new experimental data for validation.

Key Insights and Ideas

  • Unsupervised Discovery of Protein Hierarchy: The MS method can reveal a hierarchical organization of protein substructures at different scales, from local chemical bonds to global domains, purely from the atomistic structure, without any prior assumptions.
  • Mutations as Disruptors of Dynamic Modularity: Mutations don't just change local interactions; they can disrupt the "modularity" of protein domains, affecting how these domains move and interact over long timescales. This disruption can be quantified by how much a mutation affects the robustness of the identified graph partitions.
  • Long-Range Allosteric Effects: The study highlights that mutations in seemingly distant parts of a protein (like D152 in the lid domain) can have significant effects on large-scale dynamics and even enzymatic activity, demonstrating a form of allosteric communication.
  • Predictive Power of Graph Dynamics: The quantitative correlation between the computed disruption of the 3-way partition (lid, core, AMPbd domains) and the experimentally observed shifts in protein conformation (measured by FRET) is a strong validation of the graph-based approach.

Key Results

  • Identification of Key Residues: The MS analysis identified D152 as a critical residue in the lid domain of ADK, whose mutation significantly disrupts the large-scale 3-way partition.
  • Experimental Confirmation of D152 Impact: Mutations at D152 (especially D152A) led to a substantial shift in ADK towards a closed conformation, as observed by smFRET. This effect was graded, with different D152 mutations showing varying degrees of impact.
  • Quantitative Correlation: The predicted loss of robustness of the 3-way partition due to D152 mutations showed a high correlation ($R^2 = 0.93$) with the experimentally measured FRET ratio between closed and open populations.
  • Functional Consequences: All D152 mutations resulted in complete inhibition of catalytic activity, suggesting a critical role for D152 in maintaining structural integrity and function, even though it's not directly involved in substrate binding.
  • Arginine Scan Insights: An arginine scan identified V164R as another significant mutation, also shifting ADK towards a closed conformation, but with a proportional reduction in catalytic activity rather than full inhibition, suggesting a different mechanism of action compared to D152 mutations.

Limitations and Future Work

  • Computational Cost for Large Systems: While more efficient than MD for long timescales, the computational cost of full mutagenesis scans might still be a consideration for very large protein complexes.
  • Force Field Dependence: The study mentions that the diversity of force fields used in MD has prevented consensus, implying that the choice of parameters in graph construction could also influence results, though the paper focuses on a single parameterization.
  • Experimental Validation Scope: While D152 and V164R were experimentally validated, the study performed full alanine and arginine scans computationally, suggesting that further experimental validation of other predicted disruptive mutations would be valuable.
  • Future Work: The authors suggest that this framework can be extended to study other allosteric mechanisms, identify critical residues for protein engineering, and potentially be applied to larger protein systems. The identification of paired residues that might counteract or enhance the effects of mutations like D152A is also mentioned as a future direction.

Summary updated

Abstract

A bstract Proteins exhibit complex dynamics across a vast range of time and length scales, from the atomistic to the conformational. Adenylate kinase (ADK) showcases the biological relevance of such inherently coupled dynamics across scales: single mutations can affect large-scale protein motions and enzymatic activity. Here we present a combined computational and experimental study of multiscale structure and dynamics in proteins, using ADK as our system of choice. We show how a computationally efficient method for unsupervised graph partitioning can be applied to atomistic graphs derived from protein structures to reveal intrinsic, biochemically relevant substructures at all scales, without re-parameterisation or a priori coarse-graining. We subsequently perform full alanine and arginine in silico mutagenesis scans of the protein, and score all mutations according to the disruption they induce on the large-scale organisation. We use our calculations to guide Förster Resonance Energy Transfer (FRET) experiments on ADK, and show that mutating residue D152 to alanine or residue V164 to arginine induce a large dynamical shift of the protein structure towards a closed state, in accordance with our predictions. Our computations also predict a graded effect of different mutations at the D152 site as a result of increased coherence between the core and binding domains, an effect confirmed quantitatively through a high correlation ( R 2 = 0.93) with the FRET ratio between closed and open populations measured on six mutants.
BibTeX
@misc{peach2019unsupervised,
  author = {Robert L. Peach and Dominik Šaman and Sophia N. Yaliraki and David R. Klug and Liming Ying and Keith R. Willison and Mauricio Barahona},
  title = {{Unsupervised Graph-Based Learning Predicts Mutations That Alter Protein Dynamics}},
  year = {2019},
  doi = {10.1101/847426},
  howpublished = {bioRxiv},
}
Figure