bioRxiv · 10.64898/2026.11.12.766109
B-RANKO: An ncRNA Foundation Model with Bidirectional Decoder-Only Architecture
Abstract
Ribonucleic acid (RNA) plays central roles in gene expression, regulation and catalysis, and has emerged as an important target for therapeutic design. While recent RNA foundation models have achieved strong performance on representation learning tasks, relatively few approaches support both native sequence generation and transferable representation learning within a single framework. We present B-RANKO, a 150M-parameter bidirectional decoder-only ncRNA foundation model that introduces a novel bidirectional autoregressive attention mechanism, enabling a single architecture to support both de novo sequence generation and representation learning. Pretrained on a large corpus derived from RNAcentral, Rfam and GARNET, B-RANKO generates biologically plausible and thermodynamically realistic RNA sequences that exhibit strong novelty relative to their closest matches in the pretraining corpus. We further show that B-RANKO can be efficiently fine-tuned for aptamer generation and achieves competitive performance on BEACON, with particularly strong results on RNA engineering tasks. B-RANKO is publicly available on GitHub at filipvucenik/B-RANKO, together with usage examples and instructions for downloading model weights and datasets.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Vucenik, F., Penic, R. J., Sikic, M.. 2026-11-16. B-RANKO: An ncRNA Foundation Model with Bidirectional Decoder-Only Architecture. https://doi.org/10.64898/2026.11.12.766109
Cite the original work for its findings. Save a collection to share your selection of sources.