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

Ribonucleic acid (RNA) has emerged as a crucial therapeutic target owing to its central roles in gene regulation, catalysis, and disease progression. Small molecules can modulate RNA function by binding specific structural motifs; therefore, accurately predicting RNA-ligand binding free energies (ΔG) is vital for structure-guided RNA drug discovery. However, experimental determination of RNA-ligand affinities remains costly and low-throughput. We present RNALig, a machine learning (ML)-driven, structure-informed scoring function that quantitatively predicts RNA-ligand binding affinities using three-dimensional (3D) structural and physicochemical descriptors. The model employs a Random Forest Regressor trained on 164 experimentally resolved RNA-ligand complexes and validated on an independent test set of 70 complexes. RNALig integrates detailed RNA-specific, ligand-specific, and complex-level interaction features, achieving R2 = 0.81 and RMSE = 0.64 kcal/mol, thereby outperforming state-of-the-art methods such as RSAPred (R2 = 0.52) and DeepRSMA (R2 = 0.67). By combining ML interpretability with structure-based descriptors, RNALig offers a transparent, quantitative, and generalizable framework for modeling RNA-ligand binding thermodynamics, advancing ML-driven RNA-targeted drug discovery. The complete pipeline and dataset are available at https://github.com/Computational-biolab/RNALig.