Predicting Drosha and Dicer Cleavage Sites with DeepMirCut

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• MicroRNAs regulate gene expression post-transcriptionally, impacting development and disease.
• Existing computational tools primarily focus on identifying novel microRNAs, not their cleavage sites.
• Predicting Drosha cleavage sites from primary sequence and using deep neural networks were previously unexplored.

Purpose of the Study:

• To develop DeepMirCut, a recurrent neural network-based software for predicting Drosha and Dicer cleavage sites.
• To create a comprehensive microRNA primary sequence database with flanking genomic sequences.
• To evaluate different model inputs, including sequence, secondary structure, and annotated structures.

Main Methods:

• Development of a recurrent neural network model (DeepMirCut).
• Creation of a database comprising 34,713 microRNA annotations and flanking genomic sequences.
• Comparative analysis of models trained on diverse data types (sequence, structure).

Main Results:

• DeepMirCut accurately predicts both Drosha and Dicer cleavage sites.
• The best model achieved higher prediction accuracy (closer proximity) compared to existing methods.
• Specific nucleotide sequences (G before, U after Dicer sites) and structural bulges influenced prediction accuracy.

Conclusions:

• DeepMirCut offers a novel approach for predicting microRNA precursor cleavage sites.
• The findings provide insights into sequence and structural determinants of microRNA processing.
• The curated dataset and developed model are expected to facilitate future microRNA research.