Geometric Deep Learning Reveals Ligandable and Cryptic RNA Binding Small Molecule Pockets (SMARTPocket).

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• RNA molecules are crucial therapeutic targets, but identifying small-molecule binding pockets is a significant challenge in drug discovery.
• Existing methods for predicting RNA binding pockets are limited, hindering the development of RNA-targeted therapeutics.

Purpose of the Study:

• To introduce SMARTPocket, a novel deep learning framework for predicting RNA-small molecule binding pockets.
• To enhance the identification of ligandable pockets on RNA structures for drug discovery.

Main Methods:

• SMARTPocket utilizes an atomic-level geometric deep learning approach, representing RNA as full-atom point clouds.
• Transfer learning from extensive protein binding interface data (over 110,000 structures) is employed to address the scarcity of RNA-ligand complex data.
• The framework was evaluated on multiple established and curated benchmarks.

Main Results:

• SMARTPocket demonstrated superior performance compared to existing RNA pocket predictors and general biomolecular modeling tools.
• The model successfully generalized to apo RNA structures and identified cryptic ligandable pockets.
• SMARTPocket accurately predicted known binding sites, including those in SARS-CoV-2 and an RNA aptamer.
• SMARTPocket-guided docking improved pose recovery and computational efficiency over blind docking.

Conclusions:

• SMARTPocket provides a generalizable framework for structure-based identification of ligandable RNA pockets.
• The developed tool significantly accelerates the discovery of small molecules targeting RNA therapeutics.