Fast subcellular localization by cascaded fusion of signal-based and homology-based methods

Man‐Wai Mak, Wei Wang, Sun‐Yuan KungProteome 2011 · DOI

TL;DR

This paper tackles the computational bottleneck in predicting where proteins reside within a cell. By cleverly combining fast signal-based methods with more accurate but slower homology-based approaches, and crucially, by truncating protein sequences at predicted cleavage sites before analysis, they achieve comparable accuracy to full-length methods while drastically reducing computation time. This makes large-scale protein annotation much more feasible.

What Problem Does This Paper Address?

Understanding a protein's subcellular localization is fundamental to deciphering its function. While experimental methods exist, they are time-consuming and laborious. Computational prediction methods have emerged, broadly categorized into:

  • Signal-based methods: These are fast but often lack accuracy. They rely on identifying specific "sorting signals" within a protein's sequence, typically at the N-terminus.
  • Homology-based methods: These are more accurate, leveraging known locations of similar proteins. However, they are computationally expensive, especially for large datasets, as they often involve comparing entire protein sequences against vast databases.

The challenge is to develop a method that is both fast and accurate, bridging the gap between these two approaches to enable efficient, large-scale protein annotation, particularly for the ever-growing number of uncharacterized proteins from genome projects.

Methods

The core innovation lies in a "cascaded fusion" approach that combines signal-based and homology-based techniques, with a key optimization:

  1. Cleavage Site Prediction: They first use predictors like TargetP or CSitePred (based on Conditional Random Fields) to identify potential "cleavage sites" in the N-terminal region of protein sequences. These sites are where sorting signals are typically located.
  2. Sequence Truncation: Instead of using the entire protein sequence for homology searches, they truncate the sequence at the predicted cleavage site. This significantly shortens the input for subsequent steps.
  3. Profile Creation (PSI-BLAST): The truncated sequences are then fed into PSI-BLAST to generate protein profiles (PSSMs). This step is much faster because the input sequences are shorter.
  4. Profile Alignment and Classification: These truncated profiles are then aligned. To further speed up training and recognition, they replace traditional Support Vector Machines (SVMs) with a novel kernel method called Perturbational Discriminant Analysis (PDA). PDA is shown to have a significantly shorter training time.

Two schemes were investigated: "Truncating Profiles" (Scheme I) where full-length profiles are truncated, and "Truncating Sequences" (Scheme II) where sequences are truncated before profile generation. Scheme II is favored for its greater computational advantage.

Datasets

The primary dataset used was derived from the Swiss-Prot Release 57.5. Entries were filtered to include only eukaryotic species with experimentally verified annotations, excluding ambiguous or "fragment" entries. For signal peptides, mitochondrial transit peptides, and chloroplast transit peptides, only sequences with experimentally annotated cleavage sites were included. The dataset was further filtered using BLAST-Clust to ensure sequence identity was less than 25%, creating a benchmark dataset for evaluation.

Key Insights and Ideas

  • Cleavage sites are critical anchors: The N-terminal sorting signals, though short, contain crucial information for localization. Identifying these sites accurately is key to effective truncation.
  • Truncation is a powerful speed-up: Reducing the length of sequences or profiles before homology-based analysis dramatically cuts down computation time without a significant loss in accuracy. This is because the most informative regions are often at the N-terminus.
  • Sequence truncation is more efficient than profile truncation: Truncating the sequence before generating the profile (Scheme II) offers a greater computational advantage by reducing the time spent in the profile creation step itself.
  • PDA offers faster training: While SVMs are powerful, PDA provides a substantial reduction in training time, making it more suitable for large-scale applications.
  • Signal and homology methods are complementary: The fusion approach effectively leverages the speed of signal-based methods and the accuracy of homology-based methods.

Key Results

  • Accuracy: The proposed method achieves accuracy comparable to full-length sequence methods. For example, in Table 4, Scheme II (sequence truncation) achieved 91.45% accuracy, very close to the full-length sequence accuracy of 91.69%.
  • Speed-up:
    • Profile Creation: Truncating sequences at cleavage sites reduced profile creation time by approximately 6 times (Table 4).
  • Profile Alignment: Truncating sequences led to a nearly 20-fold reduction in alignment time (Table 5).
  • Training Time: PDA classifiers had a training time that was about one-fifth of conventional SVMs (Table 6).
  • Comparison: The method outperformed TargetP and SubLoc in overall accuracy, achieving up to 92.6% accuracy when using TargetP(N) as the cleavage site predictor.

Limitations and Future Work

  • Cleavage Site Prediction Accuracy: The accuracy of the overall method is dependent on the accuracy of the cleavage site predictors. Errors in predicting cleavage sites can lead to information loss. The paper notes that Chlorop is weak in predicting chloroplast transit peptide cleavage sites, impacting chloroplast localization prediction accuracy when used.
  • Signal Strength: The paper suggests that mitochondrial and chloroplast transit peptides are "weaker" signals and more sensitive to errors in cleavage site prediction compared to secretory pathway signals.
  • Future Work: The authors advocate for the method's importance in large-scale protein annotation and preliminary investigations of new algorithms involving pairwise alignments. The paper also implicitly suggests further refinement of cleavage site prediction methods to improve overall accuracy.

Summary updated

Abstract

Experimental results on a new dataset based on Swiss-Prot Release 57.5 show that the method can make use of the best property of signal- and homology-based approaches and can attain an accuracy comparable to that achieved by using full-length sequences. Analysis of profile-alignment score matrices suggest that both profile creation time and profile alignment time can be reduced without significant reduction in subcellular localization accuracy. It was found that PDA enjoys a short training time as compared to the conventional SVM. We advocate that the method will be important for biologists to conduct large-scale protein annotation or for bioinformaticians to perform preliminary investigations on new algorithms that involve pairwise alignments.
BibTeX
@inproceedings{mak2011fast,
  author = {Man‐Wai Mak and Wei Wang and Sun‐Yuan Kung},
  title = {{Fast subcellular localization by cascaded fusion of signal-based and homology-based methods}},
  booktitle = {Proteome},
  year = {2011},
  doi = {10.1186/1477-5956-9-s1-s8},
}
Figure