RNALig: An ML-Driven Structure-Based Scoring Function for Estimating Binding Affinities of RNA-Ligand Complexes.

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Ribonucleic acid (RNA) plays critical roles in gene regulation, catalysis, and disease.
• Small molecules can target RNA function, making RNA a key therapeutic target.
• Predicting RNA-ligand binding is essential for structure-guided drug discovery, but experimental methods are slow and costly.

Purpose of the Study:

• To develop a machine learning (ML)-driven scoring function, RNALig, for accurate prediction of RNA-ligand binding free energies (ΔG).
• To provide a structure-informed computational tool to accelerate RNA-targeted drug discovery.
• To offer a transparent and generalizable framework for modeling RNA-ligand binding thermodynamics.

Main Methods:

• Developed RNALig, a Random Forest Regressor model trained on 164 experimentally resolved RNA-ligand complexes.
• Utilized three-dimensional (3D) structural and physicochemical descriptors, including RNA-specific, ligand-specific, and complex-level interaction features.
• Validated the model on an independent test set of 70 complexes.

Main Results:

• RNALig achieved high predictive performance with R² = 0.81 and RMSE = 0.64 kcal/mol.
• The model outperformed existing methods like RSAPred (R² = 0.52) and DeepRSMA (R² = 0.67).
• Demonstrated the integration of ML interpretability with structure-based descriptors for quantitative binding affinity prediction.

Conclusions:

• RNALig provides a transparent, quantitative, and generalizable ML-driven framework for predicting RNA-ligand binding affinities.
• The tool advances structure-guided RNA drug discovery by offering a more efficient alternative to experimental methods.
• The RNALig pipeline and dataset are publicly available to facilitate further research.