A Novel Computational Method for Detecting DNA Methylation Sites with DNA Sequence Information and Physicochemical Properties

Gaofeng Pan, Limin Jiang, Jijun Tang, Fei GuoInternational Journal of Molecular Sciences 2018 · DOI

TL;DR

This paper introduces a novel computational method for accurately identifying DNA methylation sites by combining four distinct feature extraction techniques (k-gram, MMI, DWT, PseAAC) with a sparse Bayesian learning classifier. The proposed method demonstrates superior performance compared to existing approaches, offering a more robust tool for understanding epigenetic regulation and its role in diseases like cancer.

What Problem Does This Paper Address?

DNA methylation, the addition of a methyl group to cytosine bases, is a crucial epigenetic mechanism that influences gene expression and plays a significant role in various biological processes, including development and disease, particularly cancer. Accurately pinpointing these methylation sites within DNA sequences is essential for understanding these regulatory mechanisms and for disease sub-classification and biomarker identification. While many computational methods have been developed, especially with the advent of high-throughput sequencing, there's a continuous need for improved accuracy and robustness in predicting these sites, especially considering the complex interplay of sequence context and physicochemical properties.

Methods

The core of this work lies in a novel feature extraction pipeline and a robust classification model. The authors propose combining four distinct feature extraction methods:

  • k-gram: Captures the frequency of nucleotide combinations (e.g., "AA", "ACG") within a sequence, providing basic compositional information.
  • Multivariate Mutual Information (MMI): Analyzes the correlation between nucleotides, going beyond simple frequencies to understand dependencies within the sequence.
  • Discrete Wavelet Transform (DWT): Decomposes physicochemical properties of the DNA sequence into coefficients at different resolutions, capturing both frequency and location information while also acting as a noise filter.
  • Pseudo Amino Acid Composition (PseAAC): Adapts a method typically used for protein analysis to DNA by incorporating physicochemical properties and correlations between nucleotide pairs.

These diverse features are then fed into a Sparse Bayesian Learning (SBL) model. SBL is chosen for its ability to produce sparse solutions, offering advantages in terms of computational complexity and probabilistic predictions compared to traditional methods like Support Vector Machines (SVMs). The authors also explored feature selection to further optimize the model by identifying the most informative features.

Datasets

The proposed method was evaluated on three datasets:

  • Benchmark Dataset: Extracted from the MethDB database, this dataset was used for initial testing and comparison with existing methods. It was preprocessed to reduce redundancy and address class imbalance using CD-HIT and SMOTE (Synthetic Minority Over-sampling Technique) respectively.
  • Two scBS-seq Datasets: These datasets, derived from mouse embryonic stem cells (mESCs), were used to assess the method's performance on larger, more complex datasets. These were the 2i-cultured mESCs and serum-cultured mESCs datasets.

Key Insights and Ideas

  • Synergy of Diverse Features: The combination of sequence-based features (k-gram, MMI) and physicochemical property-based features (DWT, PseAAC) proved more powerful than any single feature type alone. This highlights that both the sequence composition and the underlying physical structure of DNA are critical for accurate methylation site prediction.
  • DWT's Prominence: Discrete Wavelet Transform emerged as a particularly important feature extraction method, contributing significantly to the model's accuracy, especially in capturing subtle patterns and filtering noise.
  • Sparse Bayesian Learning's Efficacy: The choice of SBL as the classifier provided a robust and efficient framework for handling the high-dimensional feature space and achieving accurate predictions.
  • Feature Selection's Value: Applying feature selection further refined the model, identifying a subset of highly informative features that boosted performance, demonstrating that not all extracted features are equally valuable.

Key Results

The proposed method achieved impressive results across all tested datasets:

  • Benchmark Dataset:

    • AUC: 0.8632
    • Accuracy (ACC): 0.8017
    • Matthew's Correlation Coefficient (MCC): 0.5558
    • Sensitivity (SN): 0.7268 These metrics surpassed those of existing methods like iDNA-Methyl, Methylator, and MethCGI.
  • scBS-seq Datasets:

    • 2i-cultured mESCs: AUC of 0.8896
    • Serum-cultured mESCs: AUC of 0.9511 These results were also competitive with or superior to other state-of-the-art methods like DeepCpG and RF_Zhang.
  • Feature Importance: DWT and PseAAC features were found to contribute more significantly to the classifier's performance than k-gram and MMI.

  • Feature Selection Impact: Using a selected set of 114 features (out of 612) resulted in an AUC of 0.8632, an improvement of 0.03 over the combined feature set without selection.

Limitations and Future Work

  • Computational Cost of Feature Extraction: While the classifier (SBL) is fast, some feature extraction methods (MMI and DWT) can be computationally intensive, especially for very large datasets. The authors measured the running times for each feature extraction method, indicating that k-gram is the fastest, followed by PseAAC, with MMI and DWT being the slowest.
  • Optimization of DWT: The paper suggests that further optimization of DWT levels and feature selection within DWT could lead to even better performance.
  • Generalizability: While tested on benchmark and mESC datasets, further validation on diverse biological contexts and species would be beneficial.
  • Feature Selection Complexity: Feature selection, while beneficial, can be time-consuming. The authors note that it is optional.

The authors have made their code publicly available, facilitating further research and application of their method.

Summary updated

Abstract

DNA methylation is an important biochemical process, and it has a close connection with many types of cancer. Research about DNA methylation can help us to understand the regulation mechanism and epigenetic reprogramming. Therefore, it becomes very important to recognize the methylation sites in the DNA sequence. In the past several decades, many computational methods-especially machine learning methods-have been developed since the high-throughout sequencing technology became widely used in research and industry. In order to accurately identify whether or not a nucleotide residue is methylated under the specific DNA sequence context, we propose a novel method that overcomes the shortcomings of previous methods for predicting methylation sites. We use <i>k</i>-gram, multivariate mutual information, discrete wavelet transform, and pseudo amino acid composition to extract features, and train a sparse Bayesian learning model to do DNA methylation prediction. Five criteria-area under the receiver operating characteristic curve (AUC), Matthew's correlation coefficient (MCC), accuracy (ACC), sensitivity (SN), and specificity-are used to evaluate the prediction results of our method. On the benchmark dataset, we could reach 0.8632 on AUC, 0.8017 on ACC, 0.5558 on MCC, and 0.7268 on SN. Additionally, the best results on two scBS-seq profiled mouse embryonic stem cells datasets were 0.8896 and 0.9511 by AUC, respectively. When compared with other outstanding methods, our method surpassed them on the accuracy of prediction. The improvement of AUC by our method compared to other methods was at least 0.0399 . For the convenience of other researchers, our code has been uploaded to a file hosting service, and can be downloaded from: https://figshare.com/s/0697b692d802861282d3.
BibTeX
@article{pan2018novel,
  author = {Gaofeng Pan and Limin Jiang and Jijun Tang and Fei Guo},
  title = {{A Novel Computational Method for Detecting DNA Methylation Sites with DNA Sequence Information and Physicochemical Properties}},
  journal = {International Journal of Molecular Sciences},
  year = {2018},
  doi = {10.3390/ijms19020511},
}
Figure