RNA contact prediction by data efficient deep learning.

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Understanding RNA's structure-function relationship is crucial for its design and application.
• Predicting RNA 3D structure is a key challenge, often limited by sparse labeled training data for deep learning models.
• Contact map prediction serves as a proxy for 3D RNA structure determination.

Purpose of the Study:

• To develop an effective RNA structure prediction model that addresses the challenge of limited labeled data.
• To improve the prediction of spatial adjacencies (contact maps) in RNA molecules.
• To demonstrate the generalizability of the proposed approach to related prediction tasks.

Main Methods:

• Utilized self-supervised pre-training on unlabeled data to leverage available information.
• Employed an XGBoost classifier for efficient use of sparse labeled data.
• Developed the BARNACLE model integrating self-supervised pre-training and supervised classification.

Main Results:

• The BARNACLE model achieved considerable improvement in RNA contact map prediction compared to classical and deep neural network baselines.
• Demonstrated enhanced performance in predicting spatial adjacencies, a proxy for 3D structure.
• Showcased the generalizability of the approach to accessible surface area prediction, indicating applicability to similar data-constrained problems.

Conclusions:

• The BARNACLE model offers a significant advancement in RNA structure prediction, particularly under data scarcity.
• The combination of self-supervised pre-training and XGBoost provides an effective strategy for leveraging limited biological data.
• The approach is adaptable to related prediction tasks, highlighting its potential impact across different areas of bioinformatics.