RNA sequence design model that generates protein-binding RNAs from a target structure alone, growing sequences outward from an anchored seed.
No providers recorded yet. Browse all providers
RNA-BAnG is a generative model that designs RNA sequences predicted to bind a given target protein, taking only the protein's 3D structure as input. It was developed by Roman Klypa and Sergei Grudinin at Université Grenoble Alpes and CNRS, with Alberto Bietti at the Flatiron Institute, and presented at ICML 2025. The paper's title names the decoding strategy, BAnG (Bidirectional Anchored Generation); RNA-BAnG is the RNA design model built on it.
Designing RNA that binds a chosen protein is central to aptamer development and to probing RNA-protein regulation, but existing computational approaches carry prohibitive requirements. Most need a substantial set of already-known binders for the specific target protein, which is exactly what is missing for a novel target, or they need detailed knowledge of the RNA's structure, which is the thing being designed. RNA-BAnG requires neither: it conditions on the protein structure directly, and structures predicted by folding models are explicitly supported.
The technical insight behind BAnG is that protein-binding RNA sequences are not uniformly functional. A short binding motif sits embedded within a broader sequence context that positions and presents it. Conventional left-to-right autoregressive generation handles this badly, because the motif's position within the generated sequence is not something the decoder controls. BAnG instead anchors generation at a seed position and grows the sequence outward in both directions simultaneously, so the functional motif lands in the middle of realistic flanking context by construction.
RNA-BAnG has 14.5 million parameters split across a protein module and a nucleotide module, each 10 layers deep, with latent dimension 128 and head dimension 64. The blocks combine embedding, self-attention with rotary position embeddings, geometric attention over the protein structure, and cross-attention linking the two modules.
Training proceeded in two stages. The model first learned RNA sequence statistics from roughly 3 million standalone RNA sequences drawn from RNAcentral and deduplicated at 90% similarity, then trained on 123,043 protein-nucleotide pairs from the Protein Data Bank, spanning 3,580 protein sequence clusters.
Evaluation covered both synthetic and biological settings. On DoubleBind, a synthetic task constructed to mimic localized motifs embedded in context, BAnG reached 0.97 accuracy against 0.53 for standard autoregressive generation — the clearest isolation of the decoding strategy's contribution. On biological targets, RNA-BAnG achieved an area under the threshold curve of 0.57, compared with 0.88 for the experimentally validated positive set, and 33 of 71 test proteins yielded generated sequences above the 50% affinity threshold. Generated sequences were diverse (0.93 ± 0.13) and novel relative to natural binders (0.99 ± 0.01). RNA-BAnG outperformed GenerRNA and RNAFlow on the tested samples.
The immediate use case is aptamer discovery: generating a starting library of candidate RNA sequences against a protein target for which no binders are yet known, then triaging those candidates experimentally. Because the model accepts predicted structures, targets without crystallographic data are in scope, which widens the practical set of proteins considerably. Researchers studying RNA-protein regulation can also use it to probe which sequence features a model associates with a given binding surface. Model weights and a runnable inference script are released, so a single command on a CIF file yields candidate sequences without any setup beyond installing the environment.
RNA-BAnG's contribution is a decoding strategy rather than a scaling result — at 14.5 million parameters it is small by contemporary standards, and its advantage comes from matching the generation order to the structure of the biological problem. That framing is transferable: any sequence design task where a short functional element must sit inside a longer context faces the same mismatch with left-to-right generation. The honest limits are visible in the numbers: the biological area-under-threshold-curve of 0.57 trails the 0.88 of experimentally validated binders by a wide margin, and generated sequences are computational predictions that have not been tested in the lab. The repository carries no license file, which constrains reuse.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.