Classification of short human exons and introns based on statistical features

Yonghui Wu, Alan Wee‐Chung Liew, Hong Yan, Mengsu YangPhysical 2003 · DOI

TL;DR

This paper introduces a new, more efficient set of statistical features (SZ features) for classifying short human DNA sequences into exons and introns. These SZ features achieve comparable accuracy to existing methods but with a significantly reduced feature space, leading to better recognition efficiency. Combining SZ features with a stop codon feature further boosts accuracy to over 90%.

What Problem Does This Paper Address?

Identifying exons (coding regions) and introns (non-coding regions) within DNA sequences is a fundamental challenge in bioinformatics. While statistical methods have shown success in other organisms, human DNA presents a particular difficulty due to the short average length of exons and introns. This makes it harder to distinguish them based on sequence properties alone. Existing methods often rely on a larger set of features, which can increase computational complexity. The authors aim to find a more efficient set of features that can maintain high accuracy for human sequences.

Methods

The paper builds upon the "Z-curve" features, which represent differences in nucleotide frequencies at three codon positions. They propose a new set of "simple Z (SZ) features" derived from these Z-curve features. The core idea behind the SZ features is to use a max operation across the three codon positions for specific nucleotide combinations. This is hypothesized to better capture information at the correct reading frame.

To evaluate these features, the authors employ:

  • Mutual Information: An information-theoretic approach to quantify how much information each feature provides about the class label (exon or intron).
  • K-Nearest Neighbors (KNN) Classifier: A machine learning algorithm used for classification. They explore different values of 'K' and use a sixfold cross-validation for robust evaluation.
  • Fisher Discriminant Algorithm: Another classification algorithm used for comparison.

Datasets

The study utilizes human exon and intron datasets. Specifically, they extracted:

  • 1500 human exons
  • 1500 human introns

These sequences were chosen to be short (less than 140 bp) to specifically address the challenge of classifying short human sequences. The datasets were randomly divided into training and testing sets for cross-validation.

Key Insights and Ideas

  • SZ Features Capture Essential Information Efficiently: The SZ features, despite being only three in number compared to the original nine Z-curve features, retain a high recognition rate. This suggests they are more parsimonious and capture the most discriminative information.
  • Reduced Dimensionality, Improved Efficiency: By using fewer features, the SZ features significantly reduce the dimensionality of the feature space. This leads to better recognition efficiency, meaning faster processing and potentially less computational resources needed.
  • Max Operation is Key: The max operation in the SZ features is crucial for capturing information at the correct reading frame, which is vital for distinguishing coding from non-coding regions.
  • Synergy with Stop Codons: Combining the SZ features with a "stop codon feature" (which leverages the unique occurrence of stop codons in coding regions) leads to a substantial improvement in accuracy, reaching up to 92%. This highlights the benefit of integrating different types of biological information.
  • SZ Features are Less Redundant: Correlation analysis shows that the SZ features are less correlated with each other than the original Z-curve features, indicating they provide more independent information.

Key Results

  • High Recognition Rates:
    • The nine Z-curve features achieved an accuracy of up to 88.4% using KNN.
  • The three SZ features achieved an accuracy of up to 86.3% using KNN, demonstrating comparable performance with fewer features.
  • When SZ features were combined with the stop codon feature, the accuracy reached up to 91.9% (based on sixfold cross-validation).
  • SZ Features Outperform Fisher Discriminant: Using the SZ features with the Fisher discriminant algorithm yielded an accuracy of around 80%, suggesting that the linear decision boundary assumption of Fisher's method is not ideal for this problem, unlike the more flexible KNN.
  • SZ Features are Less Correlated: Correlation coefficients between SZ features were generally small, indicating less redundancy.

Limitations and Future Work

  • Focus on Short Sequences: The study specifically focuses on short human exons and introns (<140 bp). The performance of these features on longer sequences is not explicitly discussed.
  • Comparison with Other Algorithms: While KNN and Fisher's discriminant were used, the paper mentions that the SZ features, when combined with the stop codon feature, perform comparably or better than other algorithms like "lengthen shuffling fast Fourier transform" and "Markov chain model" for short sequences. However, a direct, comprehensive comparison across a wider range of state-of-the-art algorithms might be beneficial.
  • Future Work: The authors suggest that combining SZ features with other types of features (like the stop codon feature) can significantly improve recognition rates, indicating a promising direction for future research in developing more comprehensive gene prediction models.

Summary updated

Abstract

The classification of human gene sequences into exons and introns is a difficult problem in DNA sequence analysis. In this paper, we define a set of features, called the simple Z (SZ) features, which is derived from the Z-curve features for the recognition of human exons and introns. The classification results show that SZ features, while fewer in numbers (three in total), can preserve the high recognition rate of the original nine Z-curve features. Since the size of SZ features is one-third of the Z-curve features, the dimensionality of the feature space is much smaller, and better recognition efficiency is achieved. If the stop codon feature is used together with the three SZ features, a recognition rate of up to 92% for short sequences of length $&lt;140\mathrm{bp}$ can be obtained.
BibTeX
@inproceedings{wu2003classification,
  author = {Yonghui Wu and Alan Wee‐Chung Liew and Hong Yan and Mengsu Yang},
  title = {{Classification of short human exons and introns based on statistical features}},
  booktitle = {Physical},
  year = {2003},
  doi = {10.1103/physreve.67.061916},
}
Figure