Computational Methods For Predicting Protein–Protein Interactions

Sylvain Pitre, Md Alamgir, James R. Green, Michel Dumontier, Frank Dehne, Ashkan GolshaniAdvances 2008 · DOI

TL;DR

This paper provides a comprehensive overview of computational methods for predicting protein-protein interactions (PPIs), a crucial aspect of understanding cellular functions. It highlights the limitations of experimental methods and explores various computational approaches, including genomic, evolutionary, structural, domain-based, and primary sequence-based techniques, with a focus on the PIPE algorithm developed by the authors.

What Problem Does This Paper Address?

Proteins rarely act alone; they form complex networks of interactions (PPIs) that are fundamental to nearly all cellular processes. While experimental methods exist to discover these interactions, they are often time-consuming, expensive, and prone to significant discrepancies and false positives. This makes it challenging to build a complete and accurate picture of the cellular interactome. The paper addresses the need for more efficient and reliable methods to predict PPIs, which can complement experimental data, help validate findings, and guide future research.

Methods

The paper categorizes computational PPI prediction methods into five main groups:

  • Genomic Methods: These leverage information from whole genome sequencing, such as gene order conservation and gene fusion events, to infer functional relationships between proteins.
  • Evolutionary Relationship: This approach uses phylogenetic profiles, which describe the presence of homologous genes across different organisms, to identify proteins that are likely functionally linked.
  • Protein Structure: Methods in this category utilize the 3D structure of proteins to predict interactions by assessing how well potential interacting partners fit together (docking) or by analyzing conserved structural features at interaction interfaces.
  • Domain-Based: These methods focus on protein domains, which are conserved functional units within proteins. They infer interactions based on the presence of shared or complementary domains in different proteins.
  • Primary Protein Structure: This category explores methods that use the linear sequence of amino acids to predict interactions, often employing machine learning techniques like Support Vector Machines (SVMs) to identify sequence patterns or "signatures" indicative of interaction.

A key focus is on the PIPE (Protein-Protein Interaction Prediction Engine) algorithm developed by the authors. PIPE uses primary protein structure data and measures the co-occurrence of subsequence pairs in known interacting proteins to predict novel PPIs.

Datasets

The paper references several databases for experimental PPI data, which are crucial for training and validating computational methods. These include:

  • BIND (Biomolecular Interaction Network Database)
  • DIP (Database of Interacting Proteins)
  • BioGRID (The Biological General Repository for Interaction Datasets)
  • MIPS (Munich Information Center for Protein Sequences)
  • MINT (Molecular INTeraction database)

The authors' PIPE algorithm was trained and tested using a dataset of 15,118 PPI pairs of S. cerevisiae compiled from DIP and MIPS. They also mention using positive and negative validation datasets, including a positive database of 100 known protein pairs.

Key Insights and Ideas

  • Experimental limitations drive computational need: The inherent costs, time investment, and significant data discrepancies in experimental PPI discovery necessitate robust computational approaches.
  • Multiple lines of evidence are powerful: Combining different types of data (genomic, evolutionary, structural, domain, sequence) often leads to more accurate predictions.
  • Predicting interactomes is a grand challenge: The sheer number of potential protein pairs in an organism makes comprehensive prediction computationally intensive and requires efficient algorithms.
  • False positives are a persistent challenge: Despite advancements, distinguishing true interactions from false positives remains a major hurdle in computational PPI prediction.
  • PIPE offers a promising sequence-based approach: The authors' PIPE algorithm demonstrates competitive accuracy and a lower false positive rate compared to some traditional experimental methods, leveraging only primary protein structure.

Key Results

  • PIPE's performance: The PIPE algorithm achieved an overall accuracy of 75% for yeast PPI prediction, with a false positive rate of 11% and a false negative rate of 39%. This is presented as being on par with other biochemical techniques and having lower false positive/negative rates than TAP-tagging and Y2H analysis.
  • Identification of novel interactions: PIPE was used to identify and validate a novel PPI between YGL227W and YMR135C, which was subsequently confirmed by experimental TAP-tagging.
  • Internal architecture prediction: PIPE was also applied to predict the internal architecture of the vid30 complex, identifying four proteins that formed the core of the complex.
  • Improved PIPE: Later work by the authors led to a significant improvement in PIPE's speed (over 16000-fold) and specificity (~99.9%), though with a reduced sensitivity (~15%), enabling an all-to-all examination of the entire yeast genome.

Limitations and Future Work

Limitations:

  • Data inaccuracies propagate: Computational methods are inherently dependent on the quality of experimental data, meaning inaccuracies in the input data will likely be reflected in the predictions.
  • False positives and negatives: Despite improvements, false positives and negatives remain a significant challenge, making it difficult to distinguish true interactions from noise, especially given the vast excess of potential negative interactions.
  • Gold standard development: The lack of a universally accepted "gold standard" dataset makes it difficult to objectively assess the accuracy of different prediction tools.
  • Computational cost: Sophisticated approaches, especially those aiming to predict entire interactomes, are computationally expensive and time-consuming.

Future Work:

  • Integration of multiple data types: Future tools are expected to incorporate a wider range of data categories (genomic, proteomic, structural, etc.) for more robust predictions.
  • Elimination of false positives: Research will likely focus on developing more sophisticated filters and methods to reduce false positive predictions, potentially by incorporating contextual clues like GO ontology.
  • Improved hardware and algorithms: Continued advancements in hardware performance and algorithmic optimizations will be crucial for tackling the computational demands of large-scale interactome prediction.
  • Consideration of cellular processes: Future strategies may incorporate other cellular processes like post-translational modifications and protein degradation into PPI prediction models.

Summary updated

BibTeX
@inproceedings{pitre2008computational,
  author = {Sylvain Pitre and Md Alamgir and James R. Green and Michel Dumontier and Frank Dehne and Ashkan Golshani},
  title = {{Computational Methods For Predicting Protein–Protein Interactions}},
  booktitle = {Advances},
  year = {2008},
  doi = {10.1007/10\_2007\_089},
}
Figure