Predicting coiled coils by use of pairwise residue correlations.

Bonnie Berger, David B. Wilson, Eva Wolf, T Tonchev, Mónica Milla, Peter S. KimProceedings of the National Academy of Sciences 1995 · DOI

TL;DR

This paper introduces PAIRCOIL, a novel computational method for predicting coiled-coil protein domains. By analyzing pairwise amino acid correlations within known coiled coils, PAIRCOIL significantly improves accuracy over existing methods, particularly in distinguishing true coiled coils from other alpha-helical structures and reducing false positives.

What Problem Does This Paper Address?

Proteins are the workhorses of biology, and understanding their structure is crucial for understanding their function. Many proteins fold into complex three-dimensional shapes, and a common structural motif is the "coiled coil." These are formed when two or more alpha-helices wrap around each other, creating a stable, rod-like structure found in proteins like keratin and myosin. Identifying these coiled-coil regions within a protein's amino acid sequence is important for predicting its overall structure and function.

Previous methods, like Parry's algorithm and its implementation NEWCOILS, relied on identifying specific patterns of hydrophobic residues within a repeating "heptad" sequence. While useful, these methods suffered from a significant number of "false positives" – incorrectly identifying non-coiled-coil regions as coiled coils. This was particularly problematic for distinguishing coiled coils from other alpha-helical structures, leading to ambiguity in protein structure prediction. This paper addresses this limitation by developing a more sophisticated method that leverages the subtle, correlated interactions between amino acids within coiled coils.

Methods

The core innovation of this paper is the development of the PAIRCOIL program, which moves beyond simple residue frequency analysis to incorporate pairwise residue correlations. Instead of just looking at the probability of a single amino acid appearing at a specific position in a coiled coil, PAIRCOIL analyzes the probability of two amino acids appearing together at specific positions.

The method works by:

  1. Building a database of known coiled-coil sequences: This database was compiled from various protein families known to form coiled coils (myosins, tropomyosins, intermediate filament proteins).
  2. Tabulating pairwise residue frequencies: For every pair of amino acids, the program counts how often they appear together at specific distances within the heptad repeat positions of the coiled-coil database.
  3. Normalizing probabilities: These frequencies are then normalized against the frequencies found in a general protein sequence database (Genpept) to account for background occurrence. This helps to highlight correlations that are specific to coiled coils.
  4. Calculating residue propensities: A "propensity" score is calculated for each residue, incorporating the influence of its neighboring residues at specific distances (1, 2, and 4 positions apart). This is done using a geometric mean to ensure that a strong signal across multiple distances is required for a high score.
  5. Window-based scoring: A 30-residue window is used to scan the protein sequence, and the maximum window score is taken as the score for the central residue. This window size was chosen based on the typical length of stable coiled coils.

This approach is an extension of Parry's algorithm but incorporates the crucial element of pairwise dependencies, which were traditionally used in fields like code-breaking.

Datasets

  • Coiled-coil database: A custom database of 58,217 amino acid residues from known two-stranded coiled-coil sequences, derived from Genpept and the Protein Identification Resource (PIR). Specific regions of sequences were excluded to avoid uncertainties at boundaries or due to known structural features.
  • PDB-minus database: A subset of the Brookhaven Protein Data Bank (PDB) from which known coiled coils and highly similar sequences were removed. This served as a benchmark for non-coiled-coil structures.
  • PIR-minus database: The PIR database with myosins, tropomyosins, and IF proteins removed. This was used to evaluate PAIRCOIL's performance against a broad set of protein sequences, excluding those commonly found in coiled coils.

Key Insights and Ideas

  • Pairwise correlations are key to coiled-coil specificity: The paper strongly suggests that the specific way amino acids interact with each other at certain distances is more discriminative for coiled coils than just the presence of individual amino acids.
  • "Asymmetries" in correlations reveal structural preferences: The observation of different correlations for "LaLa" (Leu at 'a' followed by Leu at 'd') versus "LaLa" (Leu at 'd' followed by Leu at 'a') highlights how the directionality and spacing within the coiled coil structure are critical.
  • PAIRCOIL effectively filters false positives: The most striking insight is PAIRCOIL's ability to identify sequences that NEWCOILS incorrectly flags as coiled coils, particularly those that are actually other types of alpha-helices. This is a significant step forward in refining protein structure prediction.
  • The database of correlations itself is informative: The visualization of pairwise correlations (Fig. 1) offers a rich resource for understanding which amino acid pairs stabilize or destabilize coiled coils, potentially guiding future protein design and engineering efforts.

Key Results

  • Significantly improved prediction accuracy: PAIRCOIL is reported to be "significantly better" than existing methods at distinguishing coiled coils from non-coiled coils.
  • Elimination of false positives: The paper highlights that PAIRCOIL produces "no obvious false positives or false negatives" when tested against the PDB-minus database. For example, the polylysine sequence (Lys-Lys-Lys), which scores highly with NEWCOILS, scores very low with PAIRCOIL.
  • Reduced overlap with non-coiled-coil databases: When comparing coiled-coil scores against PDB-minus and PIR-minus databases, PAIRCOIL shows much less overlap than NEWCOILS, indicating better discrimination.
  • Quantitative improvement: The paper provides a likelihood equation for PAIRCOIL scores, showing that a score of 3.24 corresponds to a 50% likelihood of being a coiled coil. In contrast, a NEWCOILS score of 1.35 also corresponds to 50% likelihood, but with a much higher rate of false positives.
  • Identification of biologically relevant false positives: The paper lists numerous examples of proteins that NEWCOILS predicts as coiled coils with high confidence, but PAIRCOIL scores them very poorly, suggesting these are indeed false positives that might have led to misinterpretations of protein function.

Limitations and Future Work

  • Database bias: The authors acknowledge that the database of coiled-coil sequences might contain inherent biases, although they suggest that the success of methods based on these sequences implies these biases are not "great."
  • Generalizability to three-stranded coils: The current method is based on two-stranded coiled coils. The authors suggest that the methods could be readily applied to three-stranded coiled coils with a new database.
  • Enhancing other prediction methods: The authors believe that pairwise correlation methods can enhance other structure prediction techniques, such as profile methods that rely on hydrophobicity and solvent accessibility.
  • Further exploration of correlation data: The detailed pairwise correlation data (Fig. 1) is presented as a valuable resource for future research into the specific interactions that stabilize or destabilize coiled coils.

Summary updated

Abstract

A method is presented that predicts coiled-coil domains in protein sequences by using pairwise residue correlations obtained from a (two-stranded) coiled-coil database of 58,217 amino acid residues. A program called PAIRCOIL implements this method and is significantly better than existing methods at distinguishing coiled coils from alpha-helices that are not coiled coils. The database of pairwise residue correlations suggests structural features that stabilize or destabilize coiled coils.
BibTeX
@inproceedings{berger1995predicting,
  author = {Bonnie Berger and David B. Wilson and Eva Wolf and T Tonchev and Mónica Milla and Peter S. Kim},
  title = {{Predicting coiled coils by use of pairwise residue correlations.}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {1995},
  doi = {10.1073/pnas.92.18.8259},
}
Figure