Overcoming Shortcut Learning in RNA-Small Molecule Modeling via Bias-Matched Decoys and Structure-Aware Network...

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Small molecule targeting of RNA is a promising drug discovery area.
• Current computational models face challenges due to physicochemical property misalignment in training data.
• Existing methods often rely on property filtering rather than genuine structural recognition for RNA binding prediction.

Purpose of the Study:

• To address the physicochemical property mismatch in RNA-targeted small molecule datasets.
• To develop a robust computational model for predicting RNA-binding ligands.
• To create a large, prioritized library of potential RNA-targeted drug candidates.

Main Methods:

• Introduced RNAdecoyDB, a property-matched decoy dataset using feature-matching.
• Developed the Target RNA Network (TRN), a ligand-centric model using 2D molecular graphs.
• Validated TRN on independent datasets (R-SIM, SM2miR) and applied it to the ZINC database for virtual screening.

Main Results:

• TRN demonstrates superior generalization and robustness compared to existing methods.
• Identified intrinsic ligand structural features governing RNA-binding propensity.
• Generated Zinc-TRN, a library of 6.96 million prioritized candidate molecules for RNA interaction.

Conclusions:

• The developed RNAdecoyDB and TRN model provide an unbiased and effective approach for RNA-targeted drug discovery.
• TRN's ligand-centric approach accurately identifies potential RNA binders based on structural motifs.
• Analysis of Zinc-TRN reveals key structural features, such as nitrogen-rich heterocycles and sp3 scaffolds, for designing novel RNA-targeted ligands.