RAM-ESVM

RAM-ESVM predicts N6-methyladenosine (m6A) sites in Saccharomyces cerevisiae RNA sequences to support detection and analysis of m6A-mediated post-transcriptional regulation.

Key Features:

• Benchmark performance: Outperformed nine publicly available m6A prediction tools on the independent M6Atest6540 dataset.
• Independent evaluation: Performance was assessed on the M6Atest6540 independent benchmark compiled for cross-tool comparison.
• Dataset analysis: Identified positional bias in the Met2614 training dataset compared against RNA segments from RMBase using a position bias index.
• Bias-mitigated dataset: Introduced newMet2614 by randomly selecting RNA segments from non-redundant RMBase data to reduce positional bias.
• Feature comparison: Compared models trained on Met2614 and newMet2614 using three different types of features extracted from the datasets.
• Position-specific propensity features: Found position-specific propensity-based features achieved the best predictive accuracy but exhibited overfitting on biased datasets like Met2614.
• Generalization improvement: Models trained on newMet2614 demonstrated improved generalization relative to those trained on Met2614.

Scientific Applications:

• S. cerevisiae m6A site prediction: Predicting N6-methyladenosine (m6A) sites in Saccharomyces cerevisiae RNA sequences.
• Benchmarking predictors: Providing an independent benchmark (M6Atest6540) for comparative evaluation of m6A prediction tools.
• Dataset bias assessment and correction: Assessing positional bias in training data and guiding construction of less biased datasets using RMBase.
• Feature evaluation: Evaluating and selecting feature sets, including position-specific propensity-based features, for predictive power and overfitting risk.

Methodology:

Compared nine publicly available predictors using the independent M6Atest6540 benchmark; computed a position bias index to compare Met2614 with RMBase; created newMet2614 by random sampling of non-redundant RMBase RNA segments; trained and compared models on Met2614 and newMet2614 using three feature types and evaluated performance, generalization, and overfitting of position-specific propensity-based features.