SIGNALD: Improving nucleic acid-ligand docking score using Graph Neural Network

The development of nucleic acid (NA)-targeting therapeutics offers potential for disease treatment, but their development is hindered by challenges in accurately predicting NA-ligand interactions. Conventional docking scoring functions often struggle to identify near-native poses and rank ligands across diverse targets. Here, we present SIGNALD, an Equivariant Graph Neural Network that predicts the root-mean-square deviation of NA-ligand docking poses, yielding a pose rescoring score that can be compared across ligands when normalized by ligand size. Ablation analysis identified intermolecular edges as the primary contributor to performance, with their removal reducing per-ligand SRCC from 0.415 to 0.119. Across four benchmark docking datasets, SIGNALD improved near-native pose identification relative to PLANTS and rDock. On 85 NA-ligand complexes from PDBbind, SIGNALD achieved the highest correlations with experimental K d among the evaluated methods (SRCC = 0.285 and 0.238 for native poses and docked poses, respectively). In virtual screening against HIV-1 TAR RNA, SIGNALD achieved competitive enrichment factors under an average-rank ensemble strategy (EF2% = 17.75), while SPRank remains superior under a best-score ensemble strategy. Evaluation on two dsDNA targets further demonstrated its ability to rank structurally related ligands by experimental K d (SRCC = 0.415 and 0.297), although unnormalized SIGNALD scores and RmsdXNA performed slightly better on some targets. Collectively, these findings highlight the potential of graph-based modeling of atomic interactions for structure-based NA-targeted drug discovery. Broader validation across additional target classes remains an important direction for future work. The code for SIGNALD is available at https://github.com/laiheng001/SIGNALD .