Defining functional distance using manifold embeddings of gene ontology annotations

Gilad Lerman, Boris E. ShakhnovichProceedings of the National Academy of Sciences 2007 · DOI

TL;DR

This paper introduces a novel method for quantifying functional distances between protein domains by leveraging manifold embedding techniques on Gene Ontology (GO) annotations. By creating a "functional domain universe," they demonstrate that these new distances accurately reflect established measures of sequence, structure, and phylogenetic similarity, offering deeper insights into protein evolution.

What Problem Does This Paper Address?

While we have robust methods for measuring similarity in protein sequence and structure, precisely defining functional relationships has been a significant challenge. Protein function is often described using hierarchical ontologies like the Gene Ontology (GO), but these structures aren't inherently designed for distance calculations. Existing methods, like shortest path algorithms on the GO graph, suffer from limitations such as being overly sensitive to the ontology's inherent biases and arbitrary characteristics. This lack of a rigorous functional distance metric hinders our ability to fully understand how protein function relates to sequence, structure, and evolutionary history.

Methods

The core of this paper lies in applying manifold embedding techniques, specifically diffusion-type kernels, to GO annotations. The process involves:

  1. Representing GO as a Graph: The GO's molecular function component is treated as a directed acyclic graph (DAG), where nodes are annotations.
  2. Defining Local Similarities: A kernel is constructed where elements represent local similarities between GO annotations. This is based on the idea that GO terms shared by fewer protein sequences should have smaller local distances (higher similarity).
  3. Manifold Embedding: Several diffusion-type kernels (based on local linear embedding, diffusion powers, and the pseudoinverse of the graph Laplacian) are used to embed these local similarities into a higher-dimensional space. This process captures the underlying geometry of the GO graph.
  4. Calculating Functional Distances: A distance metric is derived from the kernel, representing the Euclidean distance between embedded GO annotations. This metric is then extended to protein domains by considering their GO annotations as subgraphs.
  5. Evaluating Accuracy: The computed functional distances are correlated with established measures of sequence similarity (BLAST), structural similarity (DALI Z-scores), and phylogenetic similarity (mutual information of phylogenetic profiles).

Datasets

  • Gene Ontology (GO): Used as the primary source for functional annotations. The molecular function component of the GO DAG was utilized.
  • DALI Domain Dictionary: Used for structural comparisons.
  • NRDB (Non-Redundant Protein Database): Used to derive sequences for phylogenetic profile construction.
  • ASTRAL: Used to determine SCOP fold annotations for protein domains.

The paper does not introduce a new benchmark dataset but rather uses existing, well-established biological data sources for evaluation.

Key Insights and Ideas

  • GO is not inherently a distance metric: The paper highlights that while GO is excellent for describing function, its hierarchical structure doesn't directly translate to meaningful distance calculations.
  • Manifold embedding can capture functional relationships: By treating GO annotations as nodes in a graph and applying manifold embedding, the authors can create a metric that reflects functional similarity.
  • Diffusion kernels are powerful: The diffusion-type kernels, particularly those based on graph Laplacians and diffusion powers, prove to be effective in generating functional distances that correlate well with other biological similarity measures.
  • Functional distance reveals evolutionary patterns: The newly defined functional distances allow for the visualization of relationships between protein domains, revealing insights into both divergent (e.g., DNA binding specificity) and convergent evolution.
  • Kernel choice matters for resolution: Different diffusion kernels offer varying resolutions, allowing researchers to choose a kernel appropriate for analyzing short-range or long-range functional relationships.

Key Results

  • Strong correlations: Functional distances derived from diffusion-type manifold embeddings showed significant correlations with sequence, structure, and phylogenetic similarity. For instance, correlations with structure and phylogenetic similarity were fit to an exponential decay curve with a mean lifetime (τ) ranging from 0.23 to 0.54, indicating good agreement.
  • Superiority over shortest path: The proposed functional distances performed significantly better than metrics based on shortest path algorithms on the GO DAG, which showed no significant correlation with homology or phylogenetic similarity.
  • Biological insights: The "functional domain universe" graph, built using these distances, successfully clustered domains by fold and function, illustrating relationships like the proximity of DNA-binding and transcription factor domains.
  • Identification of evolutionary links: The method identified domains that act as functional bridges between clusters, potentially representing evolutionary intermediates.
  • Confirmation of convergence: The method confirmed known cases of convergent evolution, such as the convergence of tRNA synthases, where proteins with close functional distances had no apparent structural similarity.

Limitations and Future Work

  • Kernel choice dependency: The observed correlations are dependent on the choice of kernel, suggesting that further research into optimal kernel selection for specific applications might be beneficial.
  • Computational efficiency for large datasets: While the paper mentions discussing efficient numerical evaluation in the SI Text, the scalability of these methods to extremely large biological datasets could be a consideration.
  • Generalizability to other ontologies: The authors suggest that the methods can be generalized to other ontologies, which represents a promising avenue for future work.
  • Deeper evolutionary analysis: The paper proposes that further application of evolutionary trace methods to the identified functional relationships could uncover specific residues responsible for functional differences and dynamics.
  • Uncovering more insights: The authors believe that further exploration of the "protein domain universe" will undoubtedly reveal more insights into the relationship between structure and function evolution.

Summary updated

Abstract

Although rigorous measures of similarity for sequence and structure are now well established, the problem of defining functional relationships has been particularly daunting. Here, we present several manifold embedding techniques to compute distances between Gene Ontology (GO) functional annotations and consequently estimate functional distances between protein domains. To evaluate accuracy, we correlate the functional distance to the well established measures of sequence, structural, and phylogenetic similarities. Finally, we show that manual classification of structures into folds and superfamilies is mirrored by proximity in the newly defined function space. We show how functional distances place structure-function relationships in biological context resulting in insight into divergent and convergent evolution. The methods and results in this paper can be readily generalized and applied to a wide array of biologically relevant investigations, such as accuracy of annotation transference, the relationship between sequence, structure, and function, or coherence of expression modules.
BibTeX
@inproceedings{lerman2007defining,
  author = {Gilad Lerman and Boris E. Shakhnovich},
  title = {{Defining functional distance using manifold embeddings of gene ontology annotations}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2007},
  doi = {10.1073/pnas.0702965104},
}
Figure