Protein Remote Homology Detection by Combining Profile-based Protein Representation with Local Alignment Kernel

Bin Liu, Xiaolong Wang, Ruifeng Xu, Buzhou TangJournal of Medical and Bioengineering 2013 · DOI

TL;DR

This paper introduces a novel method for detecting remote protein homologies by transforming evolutionary information from protein sequence alignments into "profile-based proteins." By combining these new representations with a Local Alignment kernel and Support Vector Machines, the authors achieve significant improvements in prediction accuracy, outperforming existing methods.

What Problem Does This Paper Address?

The vast number of protein sequences available far outstrips the number of experimentally determined protein structures. To understand protein function and evolution, it's crucial to classify these sequences, often by identifying remote homologies – evolutionary relationships between proteins that share little sequence similarity. Traditional methods like pairwise sequence alignment struggle with this due to low sequence identity. Profile-based methods, which leverage evolutionary information from multiple sequence alignments, have shown great promise. However, effectively extracting and utilizing this evolutionary information from profiles remains a key challenge. This paper aims to bridge this gap by proposing a new way to represent this profile information that can be more readily used by existing machine learning techniques.

Methods

The core innovation of this paper is the profile-based protein representation. Here's how it works:

  1. Frequency Profiles: Multiple sequence alignments (MSAs) generated by PSI-BLAST are used to create frequency profiles. For each position in a protein sequence, a profile represents the probability of each of the 20 standard amino acids occurring there, reflecting evolutionary conservation.
  2. Profile-Based Proteins: The authors propose converting these frequency profiles into a set of "profile-based proteins." For each position in the profile, the amino acids are sorted by their frequency. Then, the top n most frequent amino acids are combined to form a "profile-based protein." This process is repeated to generate multiple such proteins, each representing a different level of evolutionary information. The intuition is that the most frequent amino acids at a position are the most informative about evolutionary constraints.
  3. Local Alignment Kernel (LA) with SVM: The generated profile-based proteins are then fed into a Local Alignment kernel (LA) within a Support Vector Machine (SVM) framework. SVMs are powerful discriminative classifiers, and the LA kernel specifically measures similarity between sequences by considering local alignments with gaps.

Datasets

The study evaluates its method on a well-established benchmark dataset for protein remote homology detection.

  • SCOP 1.53 Benchmark: This dataset, commonly used in the field, consists of 54 protein families and 4352 proteins. Crucially, it's designed to ensure that proteins within the same family have sequence similarities no higher than an E-value of 10⁻²⁵, making it suitable for testing remote homology detection. Proteins within the same family are treated as positive training/testing samples, while proteins outside the superfamily are negative samples.

Key Insights and Ideas

  • Evolutionary Information is Key: The paper reinforces the idea that evolutionary information, captured in protein profiles, is more informative for remote homology detection than raw sequence information alone.
  • Ranking Amino Acid Frequencies Matters: Not all evolutionary information in a profile is equally important. The authors hypothesize and demonstrate that the most frequent amino acids at a given position carry the most significant evolutionary signal.
  • Bridging the Representation Gap: The core insight is that by transforming frequency profiles into a sequence-like representation (profile-based proteins), existing powerful sequence-based kernels (like LA) can be directly applied, overcoming a major hurdle in utilizing profile information.
  • Top-k Proteins are Sufficient: The experiments show that using just the top few profile-based proteins (specifically p1, p2, and p3) is sufficient to achieve high performance, suggesting a parsimonious approach to feature extraction.

Key Results

The proposed method, combining profile-based protein representation with the SVM-LA kernel, demonstrated significant improvements:

  • Improved Performance: The method using the top profile-based protein (p1) achieved the best performance, improving the average ROC score by 3.7% and the average ROC50 score by 13.6% compared to using the raw protein sequence with the same SVM-LA kernel.
  • Outperforming Baselines: The proposed approach outperformed other state-of-the-art methods like SVM-Top-n-gram-combine-LSA and SVM-PDT-Profile, indicating its effectiveness in capturing evolutionary signals.
  • Importance of Top Proteins: The performance decreased as fewer informative profile-based proteins were used (p1 > p2 > p3), aligning with the intuition that the most frequent amino acids are the most discriminative.

Limitations and Future Work

  • Parameter Sensitivity: While the paper mentions optimal parameters for the LA-ekm kernel, the sensitivity of the proposed representation to these parameters isn't deeply explored.
  • Number of Profile-Based Proteins: The study focuses on using the top 3 profile-based proteins. Further investigation into the optimal number of profile-based proteins to use for different datasets or tasks could be beneficial.
  • Future Directions: The authors suggest that this profile-based protein representation provides a general framework that can be combined with other sequence-based kernels beyond LA. Future work will focus on exploring these combinations to further enhance performance.

Summary updated

Abstract

Protein remote homology detection has attracted a great deal of interest as it is one of the most important problems in bioinformatics. Profile-based methods recently achieve the state-of-the-art performance. A key step to improve the performance of these methods is to find a suitable approach to use the evolutionary information in the profiles. In this study, we propose the profile-based protein representation to extract the evolutionary information from frequency profiles. In this approach, the frequency profiles calculated from the multiple sequence alignments outputted by PSI-BLAST are converted into several profile-based proteins and then the local alignment kernel (LA) is combined with these profile-based proteins for the prediction. Our experiments on a well-known benchmark show that the proposed approach can significantly improve the predictive performance.
BibTeX
@article{liu2013protein,
  author = {Bin Liu and Xiaolong Wang and Ruifeng Xu and Buzhou Tang},
  title = {{Protein Remote Homology Detection by Combining Profile-based Protein Representation with Local Alignment Kernel}},
  journal = {Journal of Medical and Bioengineering},
  year = {2013},
  doi = {10.12720/jomb.3.1.17-22},
}
Figure