Feature Dimension Reduction Algorithm Based Prediction Method for Protein Quaternary Structure

Tong Wang, Xia Tian, Xiaoxia CaoInternational Journal of Wireless and Microwave Technologies 2012 · DOI

TL;DR

This paper introduces dimensionality reduction techniques, specifically Kernel LDA (KLDA), to predict protein quaternary structure from sequence data. By reducing complex biological data to a more manageable form, they achieve higher prediction accuracy than methods without dimension reduction, demonstrating the power of these techniques for biological problems.

What Problem Does This Paper Address?

Proteins are the workhorses of biology, and their function often depends not just on their individual 3D shape (tertiary structure) but also on how multiple protein chains assemble into larger complexes (quaternary structure). Predicting this quaternary structure from the protein's amino acid sequence is a significant challenge in bioinformatics. The amino acid sequence contains a wealth of information, but it's often buried within complex, high-dimensional data, making direct prediction difficult. This paper aims to tackle this challenge by using dimensionality reduction to simplify the data while preserving crucial information for predicting quaternary structure.

Methods

The authors explore several dimensionality reduction (DR) techniques:

  • Principal Component Analysis (PCA): A standard unsupervised method that finds a lower-dimensional representation by maximizing data variance.
  • Linear Discriminant Analysis (LDA): A supervised method that aims to maximize the separability between different classes (in this case, different quaternary structures like monomer, dimer, trimer, etc.).
  • Kernel PCA (KPCA): An extension of PCA that uses kernel functions to perform dimensionality reduction in a high-dimensional feature space, allowing for the capture of non-linear relationships.
  • Kernel LDA (KLDA): An extension of LDA that leverages kernel functions to perform supervised dimensionality reduction in a non-linear feature space.

The core idea is to represent each protein sequence using a Position-Specific Scoring Matrix (PSSM), which encodes evolutionary information. This PSSM is then transformed into a numerical vector. The DR methods are applied to these vectors to find a more compact representation before a K-Nearest Neighbors (KNN) classifier is used to predict the quaternary structure.

Datasets

The study utilizes a dataset previously established by Chou and Cai [3]. This dataset comprises 3,174 protein sequences with annotations for various quaternary structures:

  • Monomer: 382
  • Dimer: 817
  • Trimer: 593
  • Tetramer: 884
  • Pentamer: 54
  • Hexamer: 287
  • Octamer: 157

Key Insights and Ideas

  • Dimensionality Reduction is Key: The paper strongly suggests that simplifying the high-dimensional PSSM data using DR techniques is crucial for improving quaternary structure prediction accuracy.
  • Supervised vs. Unsupervised: Supervised DR methods (LDA and KLDA) generally outperform unsupervised methods (PCA and KPCA) for this task, indicating that leveraging the known quaternary structure labels during the reduction process is beneficial.
  • Non-linear Methods Shine: Kernel-based methods (KPCA and KLDA) show better performance than their linear counterparts (PCA and LDA), highlighting the presence of non-linear relationships in protein sequence data that are important for predicting quaternary structure.
  • KLDA is the Champion: Kernel LDA (KLDA) emerged as the top performer, achieving the highest jackknife success rate among all tested methods.

Key Results

The study evaluated the performance of the DR methods using a jackknife test. The results, summarized in Table 1, show a clear improvement in prediction accuracy when using DR methods compared to using KNN directly on the PSSM data:

  • K-NN (direct PSSM): 82%
  • PCA & K-NN: 84.65%
  • KPCA & K-NN: 85.73%
  • LDA & K-NN: 89.28%
  • KLDA & K-NN: 93.19%

This demonstrates that KLDA, a supervised non-linear dimensionality reduction technique, significantly boosts the accuracy of protein quaternary structure prediction.

Limitations and Future Work

Limitations: * The paper focuses on a specific dataset and encoding scheme (PSSM). The generalizability of these findings to other datasets or encoding methods is not explicitly discussed. * While KLDA performs best, the computational cost of kernel methods can be higher than linear methods, though this is not elaborated upon in the paper.

Future Work: The authors suggest that the demonstrated effectiveness of DR approaches for protein quaternary structure prediction indicates their potential applicability to other complex biological systems. This opens avenues for exploring DR in areas like protein function prediction, disease association studies, and drug discovery.

Summary updated

Abstract

Knowing the quaternary structure of an uncharacterized protein often provides useful clues for finding its biological function and interaction process with other molecules in a biological system. Here, dimensionality reduction algorithm is introduced to predict the quaternary structure of proteins. Our jackknife test results indicate that it is very promising to use the dimensionality reduction approaches to cope with complicated problems in biological systems, such as predicting the quaternary structure of proteins.
BibTeX
@article{wang2012feature,
  author = {Tong Wang and Xia Tian and Xiaoxia Cao},
  title = {{Feature Dimension Reduction Algorithm Based Prediction Method for Protein Quaternary Structure}},
  journal = {International Journal of Wireless and Microwave Technologies},
  year = {2012},
  doi = {10.5815/ijwmt.2012.05.04},
}
Figure