Loading…
Structure-based RNA virtual screening that scores small molecules against a binding site's base-pairing graph, around 10,000x faster than docking.
No providers recorded yet. Browse all providers
Docking a single compound into an RNA pocket with rDock costs roughly 65 CPU-seconds, so a one-core-day budget buys about 1,400 compounds — a rounding error against libraries that run to millions. The obvious remedy, a learned scoring function, collides with the reason RNA lags proteins in the first place: PDBBind holds tens of thousands of measured protein-ligand affinities and only around a hundred for RNA. RNAmigos2 manufactures the missing supervision instead of waiting for it, docking 500 drug-like ChEMBL compounds and the 264 native binders of the PDB into each of 1,740 curated RNA binding sites to produce over 1.3 million synthetic affinity values, and pretraining its RNA encoder on unlabeled structure.
The second ingredient is how a pocket is described. Rather than atomic coordinates, RNAmigos2 encodes a binding site as a 2.5D graph: nucleotides are nodes, and edges carry the Leontis-Westhof classification of canonical and non-canonical base pairs. That representation is coarse enough that base-pairing data alone is sufficient input, which matters for a molecule class where high-resolution structures are scarce. Where sequence-based RNA binder predictors such as SMARTBind drop structure entirely, RNAmigos2 keeps the tertiary information that survives coarse-graining.
The model comes from a collaboration centred on McGill University, jointly supervised by Carlos Oliver and Jérôme Waldispühl. Juan G. Carvajal-Patiño and Vincent Mallet share first authorship, with co-authors in Colombia, France, Germany and the United States. It succeeds the group's earlier RNAmigos, which introduced base-pairing graphs for RNA chemoinformatics but was limited by data scarcity; the redesigned encoder, directed graphs and synthetic augmentation lift native-ligand recovery by over 25% AuROC on that original task.
The RNA encoder is a three-layer relational graph convolutional network with 64 hidden dimensions over 20 edge relations covering Leontis-Westhof base-pair types and backbone connectivity; each node carries a one-hot nucleotide identity alongside a 640-dimensional RNA-FM embedding. Ligands are encoded by a graph convolutional network initialized from the OptiMol variational autoencoder, itself pretrained on about 1.5 million ZINC compounds. Pockets were clustered with RMalign at a 0.75 similarity threshold into 436 groups, 367 of which supply the training data against a held-out test set of 69 binding sites, so no structurally similar site straddles the split. On a drug-like ChEMBL decoy set the ensemble reaches 0.972 mean AuROC against rDock's 0.959, placing actives in the top 2.8% of the library. Blind screening of the ROBIN in-vitro microarray — 24,572 compounds against the TPP, ZTP, SAM-II and PreQ1 riboswitches, none seen in training — took two CPU minutes and gave a mean enrichment factor of 2.93 at 1%, with enrichment reaching 5.09.
RNAmigos2 suits early-stage RNA-targeted drug discovery: given a pocket defined by a residue list on an mmCIF file and a text file of SMILES, it returns a binding score per compound in about ten seconds on a laptop, via a command-line script or a Colab notebook. Because scoring is effectively free relative to docking, it is most useful as a prefilter — rank a large library first, then spend the docking budget on the top slice, a hybrid the authors show cuts docking error fourfold at a quarter of the compute.
RNAmigos2 is the first structure-based deep learning screen for RNA validated against an independent in-vitro binding assay rather than retrospective PDB recovery, and it retrieves chemically diverse actives that partly do not overlap with docking's hits. Its practical limits are stated plainly by the authors: a binding site must be supplied, since pocket detection is not part of the pipeline, and the model treats the site as rigid. One operational caveat matters for users — the shipped checkpoints were trained on x3dna-dssr annotations while inference on user structures annotates with fr3d-python, which can shift scores where the two disagree.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.