Hybrid representation learning for human m⁶A modifications with chromosome-level generalizability.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• N6-methyladenosine (m6A) is a crucial mRNA modification regulating gene expression post-transcriptionally.
• Existing deep learning models for m6A site prediction often lack chromosome-level generalizability due to dataset splitting methods.

Purpose of the Study:

• To develop and evaluate novel hybrid deep learning models for accurate and generalizable m6A site prediction.
• To assess model performance using both random and chromosome-out cross-validation strategies.

Main Methods:

• Proposed two hybrid deep learning models integrating k-mer sequence features and contextual embeddings using CNNs.
• Evaluated models with Random-Split and Leave-One-Chromosome-Out strategies for robust assessment.
• Compared performance against the state-of-the-art m6A-TCPred model.

Main Results:

• Both proposed models outperformed m6A-TCPred across key metrics.
• Hybrid Deep Model achieved highest accuracy in Random-Split validation.
• Hybrid Model demonstrated superior generalization in Leave-One-Chromosome-Out validation, suggesting potential overfitting of deep global representations in chromosome-independent settings.

Conclusions:

• The developed hybrid models offer improved m6A site prediction accuracy and generalization.
• Leave-One-Chromosome-Out validation is critical for assessing true robustness of m6A predictors.
• Findings provide insights for designing more reliable m6A prediction tools.