UFold: fast and accurate RNA secondary structure prediction with deep learning

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• RNA secondary structure is crucial for molecular function.
• Traditional RNA structure prediction methods face limitations in speed and accuracy.
• Current prediction performance has plateaued using thermodynamic models.

Purpose of the Study:

• To introduce UFold, a novel deep learning approach for RNA secondary structure prediction.
• To improve the accuracy and efficiency of RNA secondary structure prediction.
• To provide a fast and accurate tool for genomic research.

Main Methods:

• Developed UFold, a deep learning model utilizing Fully Convolutional Networks (FCNs).
• Introduced an image-like representation for RNA sequences to facilitate FCN processing.
• Trained the model on annotated data and base-pairing rules.

Main Results:

• UFold significantly outperforms existing methods on within-family RNA datasets.
• Achieved comparable performance to traditional methods on cross-family datasets.
• Demonstrated accurate prediction of pseudoknots.
• Exhibits fast inference times (approx. 160 ms for 1500 bp sequences).

Conclusions:

• UFold offers a significant advancement in RNA secondary structure prediction accuracy and speed.
• The deep learning approach overcomes limitations of traditional thermodynamic models.
• UFold provides a valuable, efficient tool for RNA research with an accessible web server and open-source code.