SiteSeek: Post-translational modification analysis using adaptive locality-effective kernel methods and new profiles

Paul D. Yoo, Yung Shwen Ho, Bing Bing Zhou, Albert Y. ZomayaBMC Bioinformatics 2008 · DOI

Introduction

Protein phosphorylation is a fundamental post-translational modification (PTM) that regulates a vast array of cellular processes, including signal transduction, metabolism, and cell division. By adding a phosphate group to specific residues—typically Serine (S), Threonine (T), or Tyrosine (Y)—cells can effectively toggle the function of proteins. While experimental methods like mass spectrometry provide high-fidelity mapping of these sites, they are resource-intensive and often struggle with low-abundance proteins. Consequently, computational prediction tools have become essential in the bioinformatics toolkit.

The Challenge of High-Dimensional Sequence Data

Predicting phosphorylation sites from primary protein sequences is fraught with the "curse of dimensionality." Protein sequences are complex, and standard encoding methods often generate high-dimensional feature vectors. This complexity frequently leads to overfitting, where machine learning models perform exceptionally well on training data but fail to generalize to novel, unseen sequences. SiteSeek addresses this by combining a novel feature extraction profile with a robust, semi-parametric learning architecture.

Methodological Innovations

SiteSeek introduces two primary technical advancements: the CEH-Profile and the Ada-LEKM model.

1. The CEH-Profile (Compact Evolutionary and Hydrophobic Profile)

Traditional predictors often rely solely on evolutionary information via Position Specific Scoring Matrices (PSSM). SiteSeek enhances this by integrating the SARAH1 hydrophobicity scale. To combat redundancy and noise, the researchers employed an auto-associative neural network to compress the initial 20-dimensional feature space into a more efficient 17-dimensional representation. This "less is more" approach ensures that the model focuses on the most informative biochemical and evolutionary signals.

2. Ada-LEKM (Adaptive Locality-Effective Kernel Machine)

The core of SiteSeek is a semi-parametric model known as the Locality-Effective Kernel Machine (LEKM). By using Vector Quantization to partition the data space into local Voronoi regions, the model can capture complex local patterns (non-parametric) while maintaining a global structure (parametric). This architecture is further refined using the AdaBoost algorithm, which adaptively combines multiple learners to optimize performance on challenging phosphoprotein datasets.

Performance and Benchmarking

The effectiveness of SiteSeek was evaluated using the PS-Benchmark_1 dataset, containing over 1,600 experimentally verified phosphorylation sites. The model was compared against a suite of standard machine learning algorithms, including Support Vector Machines (SVM), k-Nearest Neighbors (kNN), and Multi-Layer Perceptrons (MLP).

Comparison of Ada-LEKM with other machine learning models.

As shown in the comparative analysis, the Ada-LEKM approach consistently outperformed other models in terms of accuracy (Ac), sensitivity (Sn), and specificity (Sp). Specifically, the model achieved an overall accuracy of approximately 86.6%, significantly surpassing established tools like DISPHOS and Scansite.

Kinase-Specific Analysis

A critical strength of SiteSeek is its performance across diverse kinase families. The model was rigorously tested on four major groups: CDK, CK2, PKA, and PKC.

Performance metrics across major kinase families.

The results indicate that SiteSeek is particularly adept at identifying sites for the CMGC kinase group (e.g., CDK and CK2), while maintaining strong performance across the AGC group (e.g., PKA and PKC). The high specificity (averaging over 92.5%) ensures a low false-positive rate, which is crucial for researchers prioritizing high-confidence predictions for experimental follow-up.

Conclusion and Future Directions

SiteSeek represents a significant step forward in PTM prediction by balancing evolutionary data with physical properties and utilizing an adaptive, semi-parametric learning framework. While the current model shows exceptional performance, future work will focus on addressing sampling biases in training data and exploring variable window sizes to better capture the unique structural contexts of different kinase-substrate interactions.

Summary updated

BibTeX
@article{yoo2008siteseek,
  author = {Paul D. Yoo and Yung Shwen Ho and Bing Bing Zhou and Albert Y. Zomaya},
  title = {{SiteSeek: Post-translational modification analysis using adaptive locality-effective kernel methods and new profiles}},
  journal = {BMC Bioinformatics},
  year = {2008},
  doi = {10.1186/1471-2105-9-272},
}
Figure