In Silico Prediction of the Peroxisomal Proteome in Fungi, Plants and Animals

Olof Emanuelsson, Arne Elofsson, Gunnar von Heijne, Susana CristóbalJournal of Molecular Biology 2003 · DOI

TL;DR

This paper presents "PeroxiP," a novel computational method for predicting peroxisomal proteins across fungi, plants, and animals. By combining machine learning with cross-species comparisons and domain analysis, PeroxiP significantly improves prediction accuracy and identifies a substantial number of previously unannotated peroxisomal proteins, offering a more complete picture of this vital organelle's proteome.

What Problem Does This Paper Address?

Peroxisomes are crucial organelles involved in a wide range of metabolic processes, from fatty acid breakdown to lipid synthesis. Understanding the full complement of proteins within peroxisomes (the "proteome") is essential for comprehending their functions and the mechanisms behind peroxisomal diseases. However, accurately identifying these proteins computationally has been challenging. Existing methods often struggle with a high rate of false positives and negatives, particularly due to the conserved nature of peroxisomal targeting signals (PTS1) which can be found in non-peroxisomal proteins. This paper aims to overcome these limitations by developing a more robust and specific prediction system.

Methods

The core of the PeroxiP predictor is a multi-stage approach:

  • Preprocessing: Proteins predicted to be secreted or transmembrane by existing tools (TargetP and TMHMM) are filtered out, as these are unlikely to be peroxisomal matrix proteins.
  • Motif Identification: This module checks for the presence of peroxisomal targeting signal 1 (PTS1) at the C-terminus of a protein. Two versions are used: a "restrictive" set based on experimentally verified PTS1 motifs and a "permissive" set using a relaxed consensus.
  • Pattern Recognition: For proteins passing the motif check, a machine learning module analyzes the sequence region adjacent to the PTS1. This module uses both neural networks (NN) and support vector machines (SVMs) to learn patterns indicative of peroxisomal localization. The final prediction is based on the union of NN and SVM results, improving overall performance.
  • Cross-Species Clustering: To enhance reliability, predicted peroxisomal proteins from eight different eukaryotic genomes were clustered based on shared Pfam domains (protein families) and sequence homology (using BLAST). This assumes that conserved functions and localization are correlated across species.
  • Domain Analysis: Pfam domain content was used to further refine predictions and identify orthologous proteins.

Datasets

  • Training/Test Data: The study utilized the SWISS-PROT protein sequence database.
  • Positive Set: 152 experimentally verified peroxisomal proteins with a PTS1 motif.
  • Negative Set: 308 non-peroxisomal proteins with a PTS1-like C-terminal tripeptide.
  • Redundancy reduction was applied to create smaller, non-redundant sets for training and testing.
  • Genome-Wide Prediction: Eight eukaryotic genomes were scanned: Saccharomyces cerevisiae, Schizosaccharomyces pombe, Arabidopsis thaliana, Oryza sativa, Caenorhabditis elegans, Drosophila melanogaster, Mus musculus, and Homo sapiens.

Key Insights and Ideas

  • PTS1 is Necessary but Not Sufficient: The presence of a PTS1 motif is a strong indicator, but its effectiveness is heavily influenced by its accessibility on the folded protein. Many non-peroxisomal proteins possess PTS1-like motifs that are not exposed for receptor binding.
  • The Importance of the Adjacent Region: The sequence region immediately following the PTS1 plays a crucial role in receptor binding and thus peroxisomal targeting. Machine learning models trained on this region significantly improve prediction accuracy.
  • Cross-Species Homology Reinforces Predictions: Clustering predicted peroxisomal proteins across multiple genomes based on shared domains and sequence homology provides a powerful way to increase confidence in predictions, especially for proteins with less clear annotations.
  • Organelle-Specific Proteomes Vary: The study revealed significant differences in the predicted peroxisomal proteome size across different kingdoms, with plants generally having the largest and fungi the smallest, aligning with known differences in metabolic pathways and microbody diversity.

Key Results

  • PeroxiP Performance: The PeroxiP predictor (method 1, most restrictive) achieved a Matthews correlation coefficient (MCC) of 0.50 on a test set, outperforming PSORT (MCC 0.44) after accounting for preprocessing. Specificity was also improved.
  • Genome-Wide Predictions: The integrated approach identified a total of 430 predicted peroxisomal proteins across the eight scanned genomes.
  • Identification of Novel Proteins: A significant portion of these 430 proteins lacked prior localization annotations, suggesting they are newly identified candidates for peroxisomal localization.
  • Functional Classification: These predicted proteins were grouped into 23 families based on known peroxisomal pathways, covering major metabolic functions.
  • Species-Specific Differences: Plants showed the largest predicted peroxisomal proteome, followed by mammals and then fungi, reflecting known differences in metabolic roles and microbody diversity.

Limitations and Future Work

  • PTS2 Pathway Not Addressed: The current predictor focuses solely on PTS1-mediated targeting. Proteins relying on the PTS2 pathway or alternative import mechanisms were not captured.
  • Annotation Incompleteness: The accuracy of predictions is still limited by the quality and completeness of existing genome annotations.
  • Species-Specific Motifs: While the study pooled data, it acknowledges that PTS1 motifs might have subtle organelle specificity across different species.
  • Future Work: The authors suggest expanding the predictor to include PTS2 signals and exploring alternative import mechanisms. Further refinement of cross-species clustering and domain analysis could also improve accuracy. The identified set of 430 proteins provides a valuable resource for future experimental validation and functional studies.

Summary updated

BibTeX
@article{emanuelsson2003silico,
  author = {Olof Emanuelsson and Arne Elofsson and Gunnar von Heijne and Susana Cristóbal},
  title = {{In Silico Prediction of the Peroxisomal Proteome in Fungi, Plants and Animals}},
  journal = {Journal of Molecular Biology},
  year = {2003},
  doi = {10.1016/s0022-2836(03)00553-9},
}
Figure