OCm7G: An interpretable one-class predictor for m7G methylation sites trained with limit negative samples

Area of Science:

• Biochemistry
• Bioinformatics
• Computational Biology

Background:

• N7-methylguanosine (m7G) is a prevalent RNA modification implicated in various diseases.
• Accurate m7G site detection is crucial for understanding its biological functions.
• Existing detection methods are often resource-intensive, and AI models struggle with imbalanced datasets common in real-world scenarios.

Purpose of the Study:

• To develop a robust AI model for accurate m7G site detection, particularly addressing dataset imbalance.
• To benchmark AI model performance under realistic, imbalanced data conditions.
• To provide an interpretable AI solution for m7G site identification.

Main Methods:

• Reconstruction of independent test sets with low positive-to-negative ratios to simulate real-world data.
• Benchmarking of various AI models on both balanced and imbalanced datasets.
• Development and implementation of OCm7G, an ensemble of one-class classifiers featuring hierarchical thresholding.

Main Results:

• OCm7G demonstrates performance on par with state-of-the-art methods on balanced datasets.
• OCm7G significantly outperforms existing methods in highly imbalanced scenarios.
• The model achieves high accuracy while utilizing only 52.5% of the training data and provides interpretable predictions.

Conclusions:

• OCm7G offers a superior solution for m7G site detection, especially in imbalanced datasets.
• The interpretability of OCm7G aids researchers in understanding AI-driven biological insights.
• The developed tool and datasets are publicly available to advance RNA modification research.