RNA-binding residues prediction using structural features

Huizhu Ren, Ying ShenBMC Bioinformatics 2015 · OpenReview

TL;DR

This paper introduces a novel method for predicting RNA-binding residues in proteins by leveraging new structural features derived from "template patches" of protein surfaces. These features, combined with existing sequence-based information and an ensemble learning approach, significantly improve the accuracy of identifying these crucial interaction sites, offering a more efficient way to study RNA-protein complexes.

What Problem Does This Paper Address?

RNA-protein complexes are fundamental to numerous biological processes, from gene regulation to translation. Understanding how these complexes function hinges on identifying which amino acid residues on a protein are responsible for binding to RNA. While experimental determination of protein structures is the gold standard, it's costly and time-consuming. Therefore, computational methods for accurately predicting RNA-binding residues are highly sought after. Previous computational approaches have primarily relied on sequence-based features, but with the increasing availability of 3D structures, there's a growing interest in incorporating structural information to enhance prediction accuracy. This paper aims to bridge this gap by developing and validating new structural features for RNA-binding residue prediction.

Methods

The core innovation of this paper lies in the introduction of "structural features" derived from "template patches." The process involves:

  1. Template Patch Extraction: Patches of protein surface residues surrounding known RNA-binding residues from a training set are extracted. These are categorized as 2-amino acid (2-aa) or 3-amino acid (3-aa) patches based on the number of residues involved.
  2. Clustering and Representative Patches: These template patches are then clustered using hierarchical clustering (specifically, complete-linkage) based on their structural similarity, measured by a least-squares distance metric. The centroids of these clusters, termed "representative patches," capture distinct structural patterns of RNA-binding sites.
  3. Structural Feature Construction: For a given residue in a target protein, its surrounding surface patches are compared against these representative patches. The accumulated least-squares distances to each representative patch form the new structural feature vector for that residue. This vector quantifies how well the residue's local surface structure matches the identified structural motifs of RNA-binding sites.
  4. Ensemble Learning: These novel structural features are combined with established sequence-based features (like PSSM and residue interface propensity). An ensemble learning technique, utilizing models from the ENTOOL package (specifically ridge regression, perceptron, and multilayer perceptron), is employed to predict the binding status of residues.

Datasets

The primary dataset used for training and evaluation is RB344, a non-redundant dataset comprising 344 proteins categorized into 13 types (RNAse, SRP, aptamer, dsRNA, exosome, mRNA, ribosomal, small, snRNP, splicing, tRNA, viral, and other). RNA-binding residues were defined using two criteria: (i) any atom within 5 Å of a nucleotide atom, or (ii) residues involved in specific interactions (van der Waals, hydrogen bonding, hydrophobic, or electrostatic) with nucleotides. The paper also references the Protein Data Bank (PDB) as a source for RNA-protein complex structures.

Key Insights and Ideas

  • Structural Motifs as Binding Units: The paper proposes that RNA-binding sites are not just random collections of residues but are assembled by specific structural "patches" or motifs. Identifying these motifs is key to accurate prediction.
  • RNA Structure Preference: The clustering of template patches reveals distinct structural patterns, suggesting that RNA molecules might exhibit preferences for certain protein surface structures when binding. This is a significant implication for understanding protein-RNA recognition.
  • Feature Engineering Power: The novel structural features, derived from comparing local surface patches to representative motifs, prove to be effective, contributing to improved prediction accuracy beyond sequence-based methods alone.
  • Ensemble Learning Advantage: Combining multiple classifiers through ensemble learning further boosts prediction performance, demonstrating the robustness of the approach.

Key Results

  • The proposed structural features, when combined with PSSM scores, achieved an AUC of 0.67 and an MCC of 0.24.
  • When all features (structural, PSSM, and residue interface propensity) were used, the performance improved to an AUC of 0.68 and an MCC of 0.26.
  • The new structural features alone showed competitive performance, with an AUC of 0.62 and MCC of 0.18, slightly lower than PSSM scores alone (AUC 0.64, MCC 0.19) but demonstrating their independent value.
  • Compared to other established methods like BindN+, PPRInt, and DRNA on the RB344 dataset, the proposed method achieved an AUC of 0.68, matching BindN+ and PPRInt. While BindN+ had a slightly higher MCC (0.28 vs. 0.26), the proposed method showed the highest sensitivity.
  • The method demonstrated good performance across various protein categories, particularly excelling in aptermer and splicing classes for MCC.

Limitations and Future Work

  • Patch Size Limitation: The current study focuses on 2-aa and 3-aa template patches. The authors acknowledge that larger template patches might capture more comprehensive structural information.
  • Scale Factor Assumption: The structural similarity calculation assumes no scale change between patches, which might be a simplification.
  • Future Directions: The authors plan to extract larger template patches from binding sites and explore their structural patterns further. They also aim to investigate the observed RNA structure preferences in more detail.

Summary updated

Abstract

RNA-protein complexes play an essential role in many biological processes. To explore potential functions of RNA-protein complexes, it’s important to identify RNA-binding residues in proteins. In this work, we propose a set of new structural features for RNA-binding residue prediction. A set of template patches are first extracted from RNA-binding interfaces. To construct structural features for a residue, we compare its surrounding patches with each template patch and use the accumulated distances as its structural features. These new features provide sufficient structural information of surrounding surface of a residue and they can be used to measure the structural similarity between the surface surrounding two residues. The new structural features, together with other sequence features, are used to predict RNA-binding residues using ensemble learning technique. The experimental results reveal the effectiveness of the proposed structural features. In addition, the clustering results on template patches exhibit distinct structural patterns of RNA-binding sites, although the sequences of template patches in the same cluster are not conserved. We speculate that RNAs may have structure preferences when binding with proteins.
BibTeX
@article{ren2015rna,
  author = {Huizhu Ren and Ying Shen},
  title = {{RNA-binding residues prediction using structural features}},
  journal = {BMC Bioinformatics},
  year = {2015},
  url = {https://openreview.net/forum?id=ac7REO1Vv0},
}
Figure