Augmented base pairing networks encode RNA-small molecule binding preferences.

Area of Science:

• Computational Biology
• Drug Discovery
• Structural Biology

Background:

• RNA-small molecule interactions are crucial for regulating RNA structure and function.
• Discovering novel RNA-targeting compounds is a significant area for therapeutic development.
• Current methods for RNA drug discovery lack scalability and generalization.

Purpose of the Study:

• To apply machine learning to RNA drug discovery for enhanced scalability and generalization.
• To develop a tool, RNAmigos, for predicting RNA-binding ligands and understanding binding specificity.
• To improve the prediction of ligands for novel RNA binding sites.

Main Methods:

• Developed RNAmigos, a tool that encodes RNA structures into network representations.
• Utilized virtual screening to evaluate ligand predictions made by RNAmigos.
• Investigated the impact of non-canonical base pairing data and graph representation learning on prediction performance.

Main Results:

• RNAmigos successfully placed true ligands in the 71st-73rd percentile in decoy libraries, outperforming baselines and state-of-the-art methods.
• Augmenting structural networks with non-canonical base pairing data was essential for identifying binding specificity.
• Pre-training with graph representation learning significantly improved ligand prediction performance.

Conclusions:

• RNAmigos demonstrates the potential of machine learning in RNA drug discovery by identifying structural patterns in RNA binding data.
• Non-canonical base pairings are critical determinants of RNA-ligand binding specificity.
• Graph representation learning offers a valuable strategy for RNA structure-function prediction, especially with limited data.