Adaptive 2.5D Base-Pairing Subgraph Search Detects RNA Small-Molecule Binding Sites 1 Ribo-LENS is a geometric deep-learning framework that predicts RNA small-molecule binding sites by reasoning directly over the RNA “2.5D” base-pairing graph (Leontis–Westhof base-pair types + backbone), rather than relying on full all-atom geometry. 2 The key methodological shift is subgraph-level pocket construction: instead of scoring nucleotides independently and clustering later, Ribo-LENS uses a learned edge-contraction (adaptive pooling) process to iteratively merge high-scoring interactions into connected “supernodes,” yielding coherent pocket subgraphs and naturally allowing multiple sites per RNA. 3 Architecture summary: an RGCN embeds nucleotides on a typed relational graph (20 directed edge types), produces node embeddings + binding scores, then an edge-scoring MLP ranks interactions; greedy contraction merges the highest-scoring edges above a threshold (P=0.5), expanding a pocket outward from a high-confidence interaction core until signal is exhausted. 4 Despite using coarse-grained inputs, Ribo-LENS is competitive with or better than multiple baselines on the GerNA-Bind benchmark (22 test RNAs), including specialized predictors (GerNA-Bind, RNAsite) and all-atom co-folding models (AlphaFold3, Chai-1). Reported mean MCC: 0.380 for Ribo-LENS-3 vs 0.321 for GerNA-Bind. 5 The contraction module improves practical pocket predictions by trading an RGCN’s high-recall/low-precision residue scoring for more balanced, connected pockets (e.g., at 3 pooling steps: precision ~0.390, recall ~0.474), boosting MCC and producing site-shaped outputs rather than scattered residues. 6 Generalization is less tied to training-set homology than a fine-tuned sequence/LM approach: per-structure MCC increases more weakly with max sequence identity to training for Ribo-LENS (β=0.78, R2=0.21) than for GerNA-Bind (β=1.08, R2=0.39), which matters for transcriptome-scale deployment where homology is often low. 7 Robustness to apo→holo rearrangement is explained in graph space: on curated apo/holo pairs (RMSD up to 10.72 Å), early high-confidence cores transfer well from holo to apo, and the holo–apo performance gap correlates strongly with binding-site graph edit distance (Spearman ρ=0.82) but not with backbone RMSD (ρ=-0.15), suggesting sensitivity to base-pair rewiring rather than generic 3D displacement. 8 End-to-end virtual screening: using only sequence as the starting point (RhoFold+ structure prediction → Ribo-LENS site localization → rDock within a 10 Å sphere around predicted pocket), screening on the ROBIN assay (~25k compounds) achieves enrichment statistically indistinguishable from blind docking cavity search, while restricting docking to a small size-independent region and avoiding the steep size-scaling cost of blind cavity mapping. 9 Practical screening observations: pocket confidence scores correlate with overall ranking quality (ROC AUC; Spearman ρ=0.57) but not specifically with top-1% enrichment; and Ribo-LENS-guided docking surfaces largely complementary chemotypes to blind docking (only 13 of 57 Bemis–Murcko scaffolds shared among top-1% actives across targets). 10 Prospective relevance: on two low-homology SARS-CoV-2 RNA elements (frameshift pseudoknot 6XRZ and SL1 9EOW), predicted sites align with experimentally reported NMR chemical-shift perturbations, supporting use as a rapid pocket nomination step before more expensive 3D modeling and docking. 📜Paper: biorxiv.org/content/10.64898… #RNA #DrugDiscovery #GeometricDeepLearning #GraphNeuralNetworks #VirtualScreening #ComputationalBiology #StructuralBiology #Bioinformatics