RNAmigos

Every structure-based screen against an RNA target has the same shape: take a pocket, take a compound, search for a pose, score it, repeat. The cost scales with the library, and the scoring functions are weak besides, because RNA has only a handful of measured small-molecule affinities to fit against. RNAmigos rearranges the problem so the pocket is processed once. Rather than scoring a pocket-ligand pair, it reads the pocket alone and predicts what a binder should look like chemically — a 166-bit MACCS fingerprint, one bit per chemical property. That prediction becomes a query: sort a library by distance to it, and the actives should rise.

What makes the prediction possible is the representation. RNAmigos does not see atomic coordinates. A binding site enters as an augmented base pairing network, a graph whose nodes are nucleotides and whose edges carry the Leontis-Westhof geometry of the base pair joining them. The ablations are the paper's sharpest result: encode the site as sequence plus backbone, or as canonical secondary structure, or as the full graph with edge types stripped away, and performance collapses to the randomized baselines. Signal appears only when the non-canonical geometries are retained and labelled as distinct edge relations — evidence that the specificity of an RNA pocket lives in exactly the interactions that secondary structure discards.

RNAmigos was built at McGill University with collaborators at Mila, UQAM, Institut Pasteur and MINES ParisTech, and published in Nucleic Acids Research in 2020. The group's later RNAmigos2 replaced fingerprint prediction with direct pocket-compound scoring and is the maintained successor; this first model is the origin of the base-pairing-graph approach both share, and which sequence-only RNA binder predictors such as SMARTBind forgo.

#Key Features

• Ligand-free pocket input: The model never sees a candidate compound. It takes only the RNA graph and emits a chemical descriptor, so screening cost is independent of library size.
• Non-canonical pairs as edge relations: Thirteen edge types — backbone plus the twelve Leontis-Westhof geometries — are modelled as separate relations in a relational graph convolutional network rather than collapsed into a single "paired" label.
• Unsupervised pretraining under data scarcity: The node encoder is first trained on the full set of RNA structures to give structurally similar nucleotide neighbourhoods similar embeddings, then fine-tuned on the much smaller labelled set.
• Fixed checkpoint for new structures: A 124 KB trained model ships in the repository and runs on user-supplied mmCIF files annotated into graphs, with no retraining step.

#Technical Details

The training set was assembled from 2,993 PDB entries containing RNA and at least one non-ion ligand, keeping residues within 10 Å of a ligand atom and discarding sites with fewer than five RNA residues or a large protein component, which left 773 binding sites covering 270 unique ligands, averaging 15.76 nodes per graph. Base pair and backbone annotations come from the BGSU RNA 3D Motif Atlas, computed by FR3D. The encoder is a three-layer relational graph convolutional network with 16-dimensional embeddings throughout, followed by a graph attention pooling layer and a fully connected layer producing the 166-dimensional fingerprint, trained with binary cross-entropy. Pretraining uses 3,972 full RNA structures under a struc2vec-derived node similarity that compares edge-type counts in each node's neighbourhood. Under 10-fold cross-validation, the pretrained model places the native ligand at mean normalized rank 0.715 against a decoy set of all 270 PDB RNA ligands and 0.735 against DecoyFinder decoys — the 71st-73rd percentile — versus 0.681 and 0.695 without pretraining, and 0.542 and 0.611 for a random predictor. The secondary-structure-only tool Inforna reached 0.43 on a reduced subset of the same benchmark.

#Applications

RNAmigos is an upstream filter for RNA-targeted drug discovery. Given an mmCIF file containing the residues a user considers the binding site, rnaglib converts it to an annotated graph and the shipped checkpoint returns one fingerprint per pocket, which can then be matched against a compound collection before committing docking time to the top slice. Recovery is strongest for aminoglycosides, whose large polysaccharide-like scaffolds make many contacts with well-characterized ribosomal sites, and weaker for sparsely represented ligand classes. Pocket detection is not part of the pipeline — the site must be supplied.

#Impact

RNAmigos established that augmented base pairing networks carry enough information to predict a functional property, not just to describe structure, and that unsupervised pretraining on unlabelled RNA graphs is a workable answer to the field's chronic shortage of labels, both of which carried into the group's later work. The authors are candid about the limits: the model was trained as a proof of concept on the PDB entries public at the time and has not been evaluated beyond them, enrichment to the 71st-73rd percentile is a screening prefilter rather than an identification, and the repository carries no license file. Users needing a maintained model are directed by the authors to RNAmigos2.