Predicting essential proteins based on subcellular localization, orthology and PPI networks

Gaoshi Li, Min Li, Jianxin Wang, Jingli Wu, Fang-Xiang Wu, Yi PanBMC Bioinformatics 2016 · DOI

TL;DR

This paper introduces SON, a novel computational method for predicting essential proteins by integrating information from protein-protein interaction (PPI) networks, subcellular localization, and orthology. SON significantly outperforms existing methods, particularly in identifying low-connectivity essential proteins, which are often missed by traditional approaches.

What Problem Does This Paper Address?

Identifying essential proteins – those critical for a cell's survival and development – is crucial for understanding fundamental biological processes, developing new synthetic cells, and discovering novel drug targets. Traditional experimental methods for this task (like gene knockouts) are time-consuming, expensive, and inefficient. While computational methods have emerged, they often struggle with accuracy, especially for proteins with fewer interactions (low connectivity). Existing computational approaches primarily rely on either the topological features of PPI networks (topology-based) or the intrinsic properties of protein sequences (sequence-based). However, these methods have limitations: topology-based methods can be misled by noisy PPI data and struggle with low-connectivity proteins, while sequence-based methods might miss important functional context. This paper aims to bridge this gap by proposing a more robust prediction method that leverages multiple data sources to improve accuracy and coverage.

Methods

The core of this work is the development of the SON (Subcellular localization, Orthology, and PPI network) algorithm. SON integrates three key sources of information:

  • Network Centrality based on Edge Clustering Coefficient (NC): This metric quantifies a protein's importance within the PPI network, considering not just direct connections but also the interconnectedness of its neighbors. The raw NC score is normalized to NNC (Normalized Network Centrality) to fall within a [0,1] range.
  • Subcellular Localization Score (SL): This score is derived from the correlation between a protein's subcellular location and its essentiality. The authors analyze how often essential vs. non-essential proteins appear in different cellular compartments and use this to calculate a Subcellular Localization Correlation Coefficient (LCC) for each location. A protein's overall subcellular localization score (SL) is the sum of LCCs for all locations it inhabits. This score is then normalized to NSL (Normalized Subcellular Localization Score).
  • Orthologous Score (OS): This score reflects the evolutionary conservation of a protein, based on the presence of its orthologs in other species. The method draws inspiration from the ION algorithm. The orthologous score is calculated as the number of reference organisms in which a protein has orthologs, and then normalized to NOS (Normalized Orthologous Score).

The final sorting score for a protein pr(i) is a linear combination of these three normalized scores:

pr(i) = (1-α) * NOS(i) + α[(1−β) * NSL(i) + β * NNC(i)]

Here, α and β are tunable parameters that control the relative importance of orthology versus the combined contribution of subcellular localization and network centrality.

Datasets

The study utilizes several publicly available datasets for Saccharomyces cerevisiae (yeast):

  • PPI Network Dataset: Obtained from the DIP database.
  • Essential Protein Dataset: Compiled from MIPS, SGD, DEG, and SGDP.
  • Subcellular Localization Dataset: Sourced from the COMPARTMENTS database.
  • Orthologous Proteins Dataset: Derived from Version 7 of the InParanoid database, comparing yeast with 99 other organisms.

The authors mention using UNIPROT to unify protein identifiers across these datasets. No new benchmark datasets were introduced; instead, existing, well-established datasets were used for evaluation.

Key Insights and Ideas

  • Synergy of Multiple Data Sources: The paper strongly demonstrates that integrating diverse biological information (network topology, cellular location, evolutionary history) leads to significantly improved prediction accuracy compared to methods relying on a single data type.
  • Importance of Low-Connectivity Proteins: A critical finding is SON's superior ability to identify essential proteins that have few interactions in the PPI network. These proteins are often overlooked by traditional centrality-based methods, highlighting a significant gap in existing approaches.
  • Subcellular Localization as a Strong Predictor: The analysis reveals that certain subcellular locations are strongly correlated with essentiality, providing valuable clues for prediction.
  • Parameter Tuning is Key: The performance of SON is sensitive to the parameters α and β, which balance the contributions of the three integrated features. The authors identify optimal ranges for these parameters.

Key Results

  • Superior Overall Performance: SON consistently outperformed nine other existing methods (DC, BC, CC, SC, EC, IC, NC, PeC, and ION) across various evaluation metrics, including precision-recall curves and jackknife analysis.
  • Significant Improvement in Low-Connectivity Protein Prediction: In predictions of the top 20% ranked proteins, SON identified a substantial number of essential proteins with a degree of 10 or less, where many other methods (like DC and IC) predicted zero. For example, when predicting the top 20% proteins, SON identified 193 low-connectivity essential proteins, compared to 0 for DC and IC.
  • Parameter Optimization: The study found that optimal performance for SON was achieved with α = 0.7 and β = 0.3, indicating a strong reliance on the combined subcellular localization and network centrality scores, with a significant contribution from orthology.

Limitations and Future Work

  • Parameter Sensitivity: While optimal parameters were identified, the performance is dependent on these choices. Future work could explore more adaptive parameter selection methods.
  • Data Quality: The accuracy of PPI networks can be affected by false positives and negatives, which the authors acknowledge as a general limitation for topology-based methods.
  • Species Specificity: The current study focuses on S. cerevisiae. Extending SON to other species or developing species-agnostic versions would be a valuable future direction.
  • Integration of More Features: The authors suggest that incorporating additional biological information could further enhance prediction accuracy.

Summary updated

BibTeX
@article{li2016predicting,
  author = {Gaoshi Li and Min Li and Jianxin Wang and Jingli Wu and Fang-Xiang Wu and Yi Pan},
  title = {{Predicting essential proteins based on subcellular localization, orthology and PPI networks}},
  journal = {BMC Bioinformatics},
  year = {2016},
  doi = {10.1186/s12859-016-1115-5},
}
Figure