Abstract
Recently, several machine-learning-based DNA N-4-methycytosine (4mC) predictors have been developed to provide deeper insight into the biological functions and mechanisms of 4mC. However, the performance of the existing classifiers for identification of Escherichia coli DNA 4mC sites is inadequate. Here, we present a new support vector machine 4mC predictor, named iEC4mC-SVM, for Escherichia coli (E.coli) DNA 4mC site identification, optimized using light gradient boosting machine feature selection technology. The iEC4mC-SVM predictor had a 10-fold cross-validation accuracy of 85.4% and Jackknife cross-validation accuracy of 84.9%. The 83.2% independent testing accuracy of iEC4mC-SVM was 1.0-6.5% higher than those of state-of-the-art E. coli DNA 4mC site predictors. A t-distributed stochastic neighbor embedding analysis confirmed that the prediction performance enhancement of iEC4mC-SVM was due to the light gradient boosting machine feature selection.
Introduction
DNA methylation serves as a critical epigenetic mechanism, acting as a "switch" for various biological processes, including embryonic development and disease progression. While 5-methylcytosine (5mC) and N6-methyladenine (6mA) are widely studied, N-4-methycytosine (4mC) is a vital modification primarily found in prokaryotes like Escherichia coli. Accurate identification of 4mC sites is essential for understanding genomic regulation. However, experimental methods like SMRT sequencing are costly and labor-intensive. Existing computational predictors for E. coli have shown limited accuracy, necessitating the development of more robust tools like iEC4mC-SVM.
Methodology
The proposed iEC4mC-SVM predictor utilizes a three-stage pipeline to identify 4mC sites from 41-base DNA sequences.
- Feature Extraction: Sequences are encoded into a high-dimensional (10,060D) vector. This includes binary profiles, composition frequencies (e.g., X-gap-Y patterns), and various physicochemical properties such as stacking energy and twist.
- Feature Selection: To mitigate the "curse of dimensionality" and prevent overfitting, a Light Gradient Boosting Machine (LightGBM) is employed to rank features by importance.
- Classification: The top-ranked features are processed by a Support Vector Machine (SVM) with a Gaussian radial basis function to make the final prediction.
Results and Discussion
The performance of iEC4mC-SVM was evaluated using 10-fold cross-validation, Jackknife cross-validation, and independent testing. A key finding was the identification of an optimal feature subset. By incrementally adding features based on LightGBM importance rankings, the researchers determined that accuracy peaks at 187 features. Adding features beyond this point caused performance to fluctuate or decline due to noise.
The model achieved an independent testing accuracy of 83.2%, which is 1.0% to 6.5% higher than existing state-of-the-art predictors such as 4mCPred, 4mCPred-IFL, and meta-4mCPred. The area under the ROC curve (auROC) reached 0.925 in Jackknife testing, demonstrating high reliability. Furthermore, t-distributed stochastic neighbor embedding (t-SNE) analysis confirmed that the LightGBM-selected features significantly improved the separability of 4mC and non-4mC sites compared to the original high-dimensional set.
Conclusion
iEC4mC-SVM represents a significant advancement in the computational identification of 4mC sites in E. coli. By combining comprehensive feature extraction with efficient LightGBM-based selection, the model achieves superior accuracy and robustness. This framework provides a scalable approach that could potentially be adapted for other species and different types of DNA/RNA modifications, such as 6mA or 5mC.
Summary updated