CoCoNet-boosting RNA contact prediction by convolutional neural networks

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Co-evolutionary models and deep neural networks accurately predict protein contact maps.
• RNA structure prediction is limited by smaller available structural datasets compared to proteins.
• Existing machine learning methods are not readily applicable to RNA due to data scarcity.

Purpose of the Study:

• To develop a method for improving RNA contact map prediction using limited RNA data.
• To introduce CoCoNet, an algorithm combining coevolutionary models and shallow convolutional neural networks.
• To evaluate the performance of CoCoNet in predicting RNA contact maps.

Main Methods:

• Developed CoCoNet, integrating a coevolutionary model with a shallow Convolutional Neural Network (CNN).
• Applied CoCoNet to predict RNA contact maps using available RNA structural data.
• Validated CoCoNet's performance through cross-validation on approximately eighty RNA structures.

Main Results:

• CoCoNet significantly boosts the positive predictive value (PPV) of predicted RNA contacts by approximately 70% compared to Direct Coupling Analysis (DCA).
• The method demonstrates effectiveness despite its simplicity and a small number of trained parameters.
• Direct integration of CoCoNet contacts into 3D modeling tools did not yield a proportional increase in 3D RNA structure prediction accuracy.

Conclusions:

• CoCoNet offers a significant improvement in predicting RNA contact maps from limited data.
• The study highlights the need for new metrics to better assess the impact of contact prediction on 3D RNA structure modeling.
• CoCoNet provides a valuable tool for RNA structure prediction research and is publicly available.