Global statistical models of protein coevolution reveal higher-order sectors beyond those obtained from structure alone

Carina Shiau, Haobo Wang, Young Lee, Sergey OvchinnikovbioRxiv 2022 · DOI

TL;DR

This paper introduces a novel spectral clustering method to identify "protein sectors" – interconnected networks of coevolving residues – directly from protein sequence data. These sectors reveal evolutionary relationships that go beyond direct physical contact and can even distinguish between structurally similar proteins, suggesting they capture deeper evolutionary histories.

What Problem Does This Paper Address?

Proteins evolve under evolutionary pressure, and this pressure leaves signatures in their amino acid sequences. While single-residue conservation has long been used to infer functional importance, recent work has focused on pairwise coevolution – how mutations in one residue are compensated by mutations in another. This pairwise information can reveal direct physical interactions. However, proteins are complex systems, and evolutionary pressures can create higher-order correlations involving multiple residues that are not necessarily in direct physical contact. Identifying these higher-order relationships is crucial for a more complete understanding of protein function and evolution. Existing methods for finding these "protein sectors" have been debated, particularly regarding whether they truly represent direct couplings or simply indirect correlations. This paper aims to develop a robust method to extract these higher-order coevolutionary signals and explore their biological significance.

Methods

The core of this work is a spectral clustering algorithm designed to extract protein sectors. The process involves several key steps:

  1. Global Statistical Modeling (GREMLIN): The authors use the GREMLIN algorithm to model multiple sequence alignments (MSAs) and estimate a sequence coevolution matrix. This matrix captures pairwise coevolutionary signals.
  2. Average Product Correction (APC): This step is applied to the coevolution matrix to remove noise and disentangle structural/functional signals from background noise and phylogeny.
  3. Singular Value Decomposition (SVD): SVD is used to decompose the AP-corrected coevolution matrix. This helps to separate functionally significant coevolution signals from those that might be due to statistical noise. The authors focus on the top singular values and their corresponding singular vectors.
  4. Spectral Cleaning: This involves selecting significant singular values and left-singular vectors by comparing them to randomized simulations, effectively reducing the dimensionality and noise.
  5. Hierarchical Clustering (HDBSCAN): The cleaned, low-dimensional data is then clustered using HDBSCAN to identify the protein sectors.
  6. Bootstrapping and Non-Negative Matrix Factorization (NMF): To ensure robustness against sampling variations in the MSA, the authors employ bootstrapping. The results from multiple bootstrap runs are combined using NMF to generate a consensus clustering and derive residue probabilities for sector membership.

Datasets

The study utilizes MSAs from 120 protein families. These MSAs were constructed from the UniProt database and filtered for quality and size (at least 1000 sequences). Additionally, the authors used protein crystal structures (PDB) to assess the spatial connectivity of the identified sectors. For comparing structurally related families, they used 63 pairs of protein families with high structural similarity.

Key Insights and Ideas

  • Beyond Direct Contact: Protein sectors are not solely defined by direct physical contacts between residues. They represent networks of coevolving residues that are evolutionarily coupled, even if they are spatially distant in the 3D structure.
  • Structure-Independent Identification, Structure-Dependent Formation: The method for identifying sectors relies solely on sequence coevolution data, independent of 3D structural information. However, the identified sectors form spatially connected entities in the protein's tertiary structure.
  • Differentiating Evolutionarily Related Proteins: Surprisingly, protein families with high structural similarity do not necessarily share similar protein sectors. This suggests that protein sectors can capture nuances in evolutionary history and functional divergence that are not evident from structural similarity alone.
  • Coevolutionary Information is Richer than Contact Maps: The study demonstrates that coevolutionary signals contain information that cannot be extracted from protein contact maps alone, highlighting the power of global statistical models.

Key Results

  • The proposed spectral clustering model successfully extracts protein sectors from MSAs of 120 protein families.
  • These sectors are shown to be structurally connected in 3D space, forming compact entities.
  • A significant finding is that structurally similar protein families (e.g., 3AK8 and 3BT5) do not share similar protein sectors. This was quantified by comparing sector overlap metrics against null distributions, showing no significant overlap.
  • The analysis of graph connectivity (using scaled beta indices) revealed that coevolution-based networks are significantly more densely connected than networks derived from contact maps.

Limitations and Future Work

  • Phylogenetic and Functional Implications: The authors acknowledge that the phylogenetic and functional implications of protein sectors are yet to be fully explored. They propose future work to investigate whether sectors correlate with functional domains or reveal groups of residues that diverged in evolutionary history.
  • Interpretation of Sectors: While the study shows that sectors are distinct and informative, their precise biological meaning (e.g., representing specific functional modules or allosteric pathways) requires further investigation.
  • Model Sensitivity: The spectral clustering model is noted to be sensitive to slight deviations in the coevolution matrix, which can arise from sampling variations. Bootstrapping is used to mitigate this, but further refinement of thresholding and parameter selection might be beneficial.

Summary updated

Abstract

1 Abstract Recent methods have shown promise in using pairwise sequence coevolution predictions to illuminate physical interactions and functional relationships between pairs of protein residues. As a result, there has been an increased interest in identifying higher-order correlations between sequence positions in an effort to further understand how the multiple sequence alignment (MSA) encodes the conserved biological properties of a protein family. To this end, we propose a robust and generalizable spectral clustering model that can extract interconnected networks of coevolving residues – termed “protein sectors” – using pairwise sequence coevolution predicted by a global statistical model. We assess the statistical and evolutionary origins for protein sectors extracted from the MSA for 120 protein families. We show that protein sectors are extracted from a subset of densely connected components in the sequence coevolution matrix, many of which are not present in the pairwise residue contact graph that is constructed from the protein crystal structure, revealing the existence of networks of paired residues that are not necessarily in direct physical contact but are nonetheless evolutionarily coupled. We found that protein sectors form structurally connected entities in three-dimensional space, despite sector identification being independent of protein crystal structure. Interestingly, protein families with high structural similarity do not share similar protein sectors, suggesting that nuances in the sequence coevolution matrix can differentiate between the evolutionary histories of structurally-related protein families.
BibTeX
@misc{shiau2022global,
  author = {Carina Shiau and Haobo Wang and Young Lee and Sergey Ovchinnikov},
  title = {{Global statistical models of protein coevolution reveal higher-order sectors beyond those obtained from structure alone}},
  year = {2022},
  doi = {10.1101/2022.05.27.493723},
  howpublished = {bioRxiv},
}
Figure