Protein Remote Homology Detection Based on an Ensemble Learning Approach

Junjie Chen, Bingquan Liu, Dong HuangBioMed 2016 · DOI

TL;DR

This paper introduces SVM-Ensemble, a novel approach for detecting remote homology in protein sequences by combining three distinct feature representations using a weighted voting strategy. This ensemble method significantly improves prediction accuracy, outperforming existing state-of-the-art techniques on a standard benchmark dataset.

What Problem Does This Paper Address?

Identifying remote homology in proteins is a fundamental challenge in bioinformatics. Proteins can share similar structures and functions even when their amino acid sequences have very low similarity (below 35%). This makes traditional sequence alignment methods unreliable. Detecting these distant relationships is crucial for understanding protein function, evolution, and for applications like drug discovery. Existing methods, while improving, still struggle with this "twilight zone" of sequence similarity.

Methods

The core of this work is the SVM-Ensemble classifier. It leverages a weighted voting strategy to combine the predictions of three individual Support Vector Machine (SVM) classifiers, each trained on a different protein feature representation:

  • SVM-Kmer: Represents proteins based on the frequency of short (k-length) contiguous amino acid subsequences. This captures local sequence composition.
  • SVM-ACC (Auto-Cross Covariance): This method builds feature vectors by calculating correlations between physicochemical properties (like hydrophobicity, hydrophilicity, and mass) of amino acids at different distances (lags) along the protein sequence. It captures both local and sequence-order information.
  • SVM-SC-PseAAC (Series Correlation Pseudo Amino Acid Composition): This approach incorporates both global sequence composition and local sequence-order information into a feature vector. It uses correlation factors to represent the sequence order.

The ensemble approach aims to harness the complementary strengths of these diverse feature representations, leading to a more robust and accurate prediction.

Datasets

The study uses a widely used superfamily benchmark dataset derived from the SCOP database (version 1.53). This dataset consists of 54 families and 4352 proteins, with sequence similarities between any two sequences less than an E-value of 10⁻²⁵. The dataset is structured such that proteins within a family are positive training samples, while proteins outside the family but within the same superfamily are positive test samples. Negative samples are drawn from outside the fold. This benchmark has been used in numerous prior studies, allowing for direct performance comparisons.

Key Insights and Ideas

  • Diversity is Key: Combining different feature representations (Kmer, ACC, SC-PseAAC) is more effective than relying on any single one. Each feature type captures distinct aspects of protein information.
  • Sequence Order Matters: Features that incorporate sequence-order effects (like ACC and SC-PseAAC) are crucial for improving performance beyond simple amino acid composition.
  • Ensemble Power: A weighted voting ensemble, where weights are determined by the performance of individual classifiers (e.g., ROC scores), can significantly boost predictive accuracy by mitigating the weaknesses of individual models.
  • Amino Acid 'M' is Important: Analysis of discriminative features revealed that the amino acid 'M' (Methionine) frequently appears in top features for Kmer and SC-PseAAC, suggesting its significance in distinguishing protein families.

Key Results

The proposed SVM-Ensemble achieved a top ROC score of 0.943, outperforming all three individual basic predictors (SVM-Kmer: 0.912, SVM-ACC: 0.787, SVM-SC-PseAAC: 0.911) and significantly surpassing other state-of-the-art methods on the benchmark dataset. For instance, it outperformed SVM-Pairwise (0.896) and SVM-Top-n-gram-combine-LSA (0.939). The ensemble also achieved a superior ROC50 score of 0.744.

Limitations and Future Work

  • Parameter Optimization: The performance of the basic predictors is sensitive to parameter choices (e.g., k for Kmer, LAG for ACC, λ and ω for SC-PseAAC). While grid search was used, further exploration of parameter spaces could potentially yield even better results.
  • Feature Engineering: While the chosen features are effective, the authors suggest exploring more advanced features and machine learning techniques for future work.
  • Broader Applications: The authors propose applying ensemble learning techniques and neural-like computing models to other bioinformatics problems, such as gene-disease relationship prediction and DNA motif identification.

Summary updated

Abstract

Protein remote homology detection is one of the central problems in bioinformatics. Although some computational methods have been proposed, the problem is still far from being solved. In this paper, an ensemble classifier for protein remote homology detection, called SVM-Ensemble, was proposed with a weighted voting strategy. SVM-Ensemble combined three basic classifiers based on different feature spaces, including Kmer, ACC, and SC-PseAAC. These features consider the characteristics of proteins from various perspectives, incorporating both the sequence composition and the sequence-order information along the protein sequences. Experimental results on a widely used benchmark dataset showed that the proposed SVM-Ensemble can obviously improve the predictive performance for the protein remote homology detection. Moreover, it achieved the best performance and outperformed other state-of-the-art methods.
BibTeX
@inproceedings{chen2016protein,
  author = {Junjie Chen and Bingquan Liu and Dong Huang},
  title = {{Protein Remote Homology Detection Based on an Ensemble Learning Approach}},
  booktitle = {BioMed},
  year = {2016},
  doi = {10.1155/2016/5813645},
}
Figure