PCalign: a method to quantify physicochemical similarity of protein-protein interfaces

Shanshan Cheng, Yang Zhang, Charles L. Brooks IiiBMC Bioinformatics 2015 · OpenReview

TL;DR

This paper introduces PCalign, a novel coarse-grained method for quantifying the physicochemical similarity between protein-protein interfaces. PCalign overcomes limitations of existing methods by handling low-resolution structural data and accounting for both geometric and chemical properties, enabling the discovery of functional relationships masked by apparent structural dissimilarity.

What Problem Does This Paper Address?

Understanding how proteins interact is fundamental to biology, from cellular signaling to disease mechanisms. Protein-protein interfaces are the key sites of these interactions. While comparing these interfaces can reveal functional relationships that aren't obvious from evolutionary history alone, existing computational methods often struggle with several challenges:

  • Resolution Dependence: Many methods rely on atomic-level detail, making them unsuitable for large macromolecular complexes often solved with lower-resolution techniques like cryo-electron microscopy (cryo-EM). This limits our ability to study a vast range of important biological assemblies.
  • Sequence Order: Protein interfaces are formed by residues that might not be adjacent in the primary sequence. Methods that strictly adhere to sequence order can miss these important spatial arrangements.
  • Chemical Properties: Simply comparing geometric shapes isn't enough; the chemical nature of the amino acids at the interface plays a crucial role in binding and function. Existing methods often neglect this aspect or handle it inadequately.

The authors aim to bridge this gap by developing a method that is robust to resolution, sequence-order independent, and integrates both geometric and chemical information to provide a more comprehensive measure of interface similarity.

Methods

PCalign employs a multi-step approach:

  1. Extract Interfacial Residues: The method first identifies residues involved in the protein-protein interface. For high-resolution structures, this is done by checking for heavy atom proximity. For lower-resolution structures (where only Cα atoms are available), a distance-dependent criterion is used, where the cutoff distance is dynamically adjusted based on the types of amino acids involved to account for varying side-chain sizes.
  2. Identify Initial Alignments: To find the best spatial superposition of two interfaces, PCalign uses a geometric hashing algorithm. This technique treats each interface as a set of Cα atoms (color-coded by amino acid chemical type) in 3D space. It's "sequence-order independent" because it doesn't assume any predefined order of residues. The algorithm identifies transformations (rotations and translations) that maximize the number of spatially and chemically matching residue pairs. It also accounts for the fact that interfaces are typically formed by two distinct binding fragments, ensuring that fragment A aligns with A' and B with B'.
  3. Iterative Refinement: After an initial alignment, the method refines it iteratively. It uses a maximum weight matching algorithm (specifically the Hungarian algorithm) on a bipartite graph to identify structurally equivalent residue pairs. The "weight" of a match considers both the spatial proximity (Euclidean distance between Cα atoms) and the chemical similarity of the residue pair. This process is repeated until no further significant improvement in the alignment is achieved.
  4. Scoring Function (PC-score): A normalized scoring function, the PC-score, is derived from the optimal alignment. This score quantifies the degree of overlap in terms of both geometric and chemical characteristics. It's designed to be between 0 and 1, with 1 representing identical interfaces. The raw score is further scaled to remove dependency on interface size.

Datasets

The paper utilizes several datasets for its analysis:

  • PDB (Protein Data Bank): Used to derive statistics for the distance-dependent criterion for identifying interfacial residues in low-resolution structures.
  • SCOPPI (Structural Classification Of Protein-Protein Interfaces): A database of protein-protein interfaces grouped by evolutionary relationships. This dataset was manually curated to select 609 dimers from 124 protein families, ensuring structural dissimilarity between monomers from different groups. This dataset was used to benchmark PCalign against existing methods.
  • Viral Capsid Proteins: Specifically, 18 T=3 viruses were used to compare PCalign's Q-score with a previously established Q-score metric.
  • Viral Mimicry Examples: Three specific examples of viral mimicry (M3 protein-CCL2, SV5 V protein-DDB1, NiV-G-ephrin-B2) were used to illustrate the method's ability to detect functional relationships masked by low sequence/structural similarity.

Key Insights and Ideas

  • Bridging the Resolution Gap: The most significant insight is PCalign's ability to handle both high- and low-resolution structural data. This is crucial for studying large biological assemblies like viral capsids, which are often only available at lower resolutions.
  • Beyond Geometry: By incorporating chemical similarity alongside geometric alignment, PCalign provides a more nuanced and biologically relevant measure of interface similarity. This allows for the detection of functional relationships that might be missed by purely geometric approaches.
  • Detecting Convergent Evolution: The method is particularly adept at identifying interfaces that have evolved similar functional roles independently (convergent evolution), even if the underlying proteins are not closely related. This is a key application for understanding how viruses hijack host machinery or how different biological systems solve similar problems.
  • Robustness to Noise: The "backbone" dataset, created by perturbing Cα atom positions, demonstrates PCalign's robustness to noise and low-quality structural data.

Key Results

  • Validation with Q-score: PCalign's similarity score showed a high correlation (0.93) with the established Q-score for quasi-equivalent viral capsid protein-protein interfaces, validating its ability to capture geometric properties.
  • Performance Comparison:
    • In predicting highly related interfaces (using SCOPPI data), PCalign achieved an AUC of 0.970 (original data) and 0.955 (backbone data), performing comparably to the state-of-the-art method Ialign (AUC 0.980) and outperforming I2I-SiteEngine (AUCs 0.831-0.909).
  • When comparing PCalign and Ialign on unrelated interfaces (Figure 3), PCalign demonstrated a statistically significant advantage in terms of higher coverage and lower RMSD, especially when considering only geometric properties.
  • PCalign outperformed Ialign in finding better structural alignments in 50,838 cases (with chemical term) and 57,639 cases (without chemical term), with odds ratios of 1.5 and 2.1 respectively.
  • Viral Mimicry Detection: PCalign successfully identified significant interface similarities in three viral mimicry cases where sequence and structural similarity were low, highlighting its utility in uncovering functional relationships masked by evolutionary divergence.

Limitations and Future Work

  • Sensitivity to Local Motifs: The authors acknowledge that PCalign's focus on overall interface similarity might lead to a lack of sensitivity for detecting small, localized structural motifs within larger, otherwise dissimilar interfaces. The normalization process can dilute these local signals.
  • Computational Cost: While PCalign is faster than I2I-SiteEngine, it is slightly slower than Ialign due to its more complex algorithm for sampling a larger alignment space. However, the authors deem this cost sufficiently low for large-scale studies.
  • Future Work: The authors express interest in applying PCalign to analyze large macromolecular assemblies like viral capsids to gain deeper insights into their higher-order organization and unique structural/functional characteristics. They also suggest that the method could be a valuable tool for exploring the repertoire of protein-protein interfaces and understanding their structural relationships.

Summary updated

Abstract

Structural comparison of protein-protein interfaces provides valuable insights into the functional relationship between proteins, which may not solely arise from shared evolutionary origin. A few methods that exist for such comparative studies have focused on structural models determined at atomic resolution, and may miss out interesting patterns present in large macromolecular complexes that are typically solved by low-resolution techniques. We developed a coarse-grained method, PCalign, to quantitatively evaluate physicochemical similarities between a given pair of protein-protein interfaces. This method uses an order-independent algorithm, geometric hashing, to superimpose the backbone atoms of a given pair of interfaces, and provides a normalized scoring function, PC-score, to account for the extent of overlap in terms of both geometric and chemical characteristics. We demonstrate that PCalign outperforms existing methods, and additionally facilitates comparative studies across models of different resolutions, which are not accommodated by existing methods. Furthermore, we illustrate potential application of our method to recognize interesting biological relationships masked by apparent lack of structural similarity. PCalign is a useful method in recognizing shared chemical and spatial patterns among protein-protein interfaces. It outperforms existing methods for high-quality data, and additionally facilitates comparison across structural models with different levels of details with proven robustness against noise.
BibTeX
@article{cheng2015pcalign,
  author = {Shanshan Cheng and Yang Zhang and Charles L. Brooks III},
  title = {{PCalign: a method to quantify physicochemical similarity of protein-protein interfaces}},
  journal = {BMC Bioinformatics},
  year = {2015},
  url = {https://openreview.net/forum?id=r5ym2drKxn},
}
Figure