Multi-kernel feature extraction with dynamic fusion and downsampled residual feature embedding for predicting rice...

Area of Science:

• Agricultural Science
• Molecular Biology
• Bioinformatics

Background:

• RNA N$^{6}$-methyladenosine (m$^{6}$A) is a crucial epigenetic modification influencing rice growth, development, and stress responses.
• Accurate m$^{6}$A identification is vital for rice breeding and understanding phenotype regulation.
• Existing rice m$^{6}$A prediction methods face challenges with variable-length sequences, leading to information sparsity and reduced feature extraction accuracy.

Purpose of the Study:

• To develop an end-to-end deep learning framework, MFDm$^{6}$ARice, for precise prediction of m$^{6}$A sites in rice.
• To address information sparsity and improve feature extraction accuracy in m$^{6}$A prediction.
• To enhance the computational efficiency and generalization of m$^{6}$A prediction models.

Main Methods:

• Developed MFDm$^{6}$ARice, a deep learning framework incorporating a multi-kernel feature fusion module for enhanced information mining.
• Implemented a downsampling residual feature embedding module to optimize feature space and computational performance.
• Utilized multi-kernel feature extraction and global-local dynamic fusion for effective information transfer.

Main Results:

• MFDm$^{6}$ARice demonstrated superior performance compared to existing methods in cross-validation and independent test sets (same and cross-species).
• The model exhibited good robustness and generalization capabilities.
• Application to maize m$^{6}$A prediction indicated the framework's scalability.

Conclusions:

• MFDm$^{6}$ARice effectively overcomes limitations of traditional methods for rice m$^{6}$A site prediction.
• The framework's architecture, including multi-kernel fusion and residual downsampling, significantly improves feature representation and extraction.
• MFDm$^{6}$ARice offers a robust and scalable solution for m$^{6}$A site prediction, advancing precision rice breeding and molecular studies.