Protein classification using texture descriptors extracted from the protein backbone image

Loris Nanni, Jian‐Yu Shi, Sheryl Brahnam, Alessandra LuminiJournal of Theoretical Biology 2010 · DOI

TL;DR

This paper introduces a novel method for protein classification by extracting texture descriptors from the 2D distance matrix of a protein's backbone. By combining these texture features with traditional amino acid sequence descriptors, the authors achieve improved classification performance across multiple benchmark datasets, particularly for tasks involving structural properties.

What Problem Does This Paper Address?

Classifying proteins based on their structure and function is a fundamental challenge in bioinformatics. Traditional methods often rely on analyzing amino acid sequences, but these can miss crucial information encoded in the 3D spatial arrangement of proteins. While some approaches analyze protein structure, extracting meaningful features from these complex 3D representations can be computationally intensive and challenging. This paper aims to bridge this gap by proposing a method that efficiently extracts informative features from a simplified representation of protein structure – the 2D distance matrix of its backbone – and combines them with sequence-based features for enhanced classification.

Methods

The core of this work lies in extracting "texture descriptors" from the 2D distance matrix (DM) of a protein's backbone. The DM is generated from the 3D coordinates of the protein's alpha-carbon atoms. This DM is treated as a grayscale image, and various texture analysis techniques are applied:

  • Dominant Local Binary Patterns (DLBP) and Local Ternary Patterns (LTP): These methods analyze local pixel neighborhoods to capture texture information. The authors combine DLBP with LTP to improve robustness to noise.
  • Radon Transform + Discrete Cosine Transform (Radon+DCT): The Radon transform projects the image along radial lines, and DCT is applied to the resulting coefficients to extract features.
  • Haralick Texture Features: These are derived from Gray Level Co-occurrence Matrices (GLCMs) and capture statistical properties of texture. The authors use two sets of Haralick features, one considering two angles and another considering four angles.
  • Chou's Pseudo Amino Acid Composition (PseAA): This is a standard sequence-based descriptor that incorporates information about amino acid order and physicochemical properties.

To improve classification performance, the authors employ an ensemble of Support Vector Machines (SVMs) using a random subspace method. They also explore extracting features not only from the entire DM but also from specific sub-windows (quadrants) of the matrix. Finally, they combine features from different descriptors using a sum rule, demonstrating that this fusion approach outperforms individual methods.

Datasets

The proposed method was evaluated on three benchmark datasets:

  • Protein Fold Recognition (FOLD): This dataset, derived from Ding and Dubchak (2001), contains 313 proteins for training and 385 for testing, with sequence similarities below 35-40% and 27 classes.
  • DNA-binding proteins (DNA): This dataset consists of 118 DNA-binding proteins and 231 non-DNA-binding proteins, with less than 35% sequence identity.
  • Gene Ontology (GO) dataset: This dataset was created by collecting proteins based on GO annotations, distinguishing between biological processes and molecular functions.

Key Insights and Ideas

  • Backbone Distance Matrix as a Rich Feature Source: The study demonstrates that the 2D distance matrix of a protein's backbone, despite being a simplified representation, contains sufficient information for effective protein classification, especially for structural properties.
  • Texture Descriptors Excel for Structural Tasks: Texture features extracted from the distance matrix show significant promise, particularly for classifying proteins based on their structural properties, outperforming sequence-based methods in these cases.
  • Fusion is Key: Combining texture descriptors from the distance matrix with traditional amino acid sequence descriptors (like PseAA) leads to a substantial improvement in classification performance, suggesting complementary information is captured.
  • Sub-window Analysis Enhances Performance: Extracting features from specific sub-windows (quadrants) of the distance matrix, in addition to the whole matrix, further boosts classification accuracy.
  • Dominant LTP/DLBP Outperforms Standard Variants: For texture descriptors, the "dominant" versions of Local Binary and Ternary Patterns (DLBP/DLTP) show better performance than their standard counterparts.

Key Results

The paper reports several key findings:

  • The proposed ensemble method, particularly the fusion of multiple texture descriptors and PseAA, consistently outperforms stand-alone approaches across the tested datasets.
  • For the FOLD dataset, the proposed system (FUS3) achieved an accuracy of 81.04%, outperforming other methods in the literature.
  • In the BINDING dataset, the fusion approach (FUS4) achieved a high accuracy of 92.16%.
  • The study highlights that texture descriptors extracted from the distance matrix are particularly effective for tasks where structural configuration is important, such as in the FOLD and BINDING datasets.
  • The performance of PseAA was notably low on the FOLD dataset, underscoring the importance of structural features for this type of classification.

Limitations and Future Work

  • Feature Selection Not Performed: The authors explicitly state they do not perform feature selection, preferring longer feature vectors to ensure generality across datasets. They acknowledge that feature selection could potentially improve performance for specific datasets.
  • Parameter Optimization: The paper explores parameter optimization for some descriptors (e.g., τ and K for dominant LTP, d for MI HARA3, and R for Radon+DCT), suggesting that optimal parameter choices are crucial.
  • Future Work: The authors plan to investigate more texture descriptors and explore other machine learning algorithms. They also express interest in developing user-friendly web servers for their predictors, following the trend in bioinformatics. Preliminary tests with neighborhood preserving embedding (NPE) and Gabor-based descriptors show promise.

Summary updated

BibTeX
@article{nanni2010protein,
  author = {Loris Nanni and Jian‐Yu Shi and Sheryl Brahnam and Alessandra Lumini},
  title = {{Protein classification using texture descriptors extracted from the protein backbone image}},
  journal = {Journal of Theoretical Biology},
  year = {2010},
  doi = {10.1016/j.jtbi.2010.03.020},
}
Figure