RIBOSPAN: A Long-Context RNA Foundation Model for Versatile RNA Modeling
The model is built to read complete long RNAs at single-nucleotide resolution, with native context up to 10,240 nucleotides.
RIBOSPAN is a 1.61B-parameter bidirectional RNA model aimed at transcripts that exceed the context windows of earlier RNA foundation models. The paper says native 10K pretraining preserves reconstruction quality and keeps perturbation effects localized better than short-context extrapolation. It reports state-of-the-art results for RNA representation learning, full-transcript property prediction, and zero-shot mutation-fitness modeling. The same backbone is also used for full-length mRNA generation and redesign, including synonymous-codon optimization that preserves protein sequence. HF Daily Papers' note
RIBOSPAN is a 1.61B-parameter bidirectional RNA model aimed at transcripts that exceed the context windows of earlier RNA foundation models. The paper says native 10K pretraining preserves reconstruction quality and keeps perturbation effects localized better than short-context extrapolation. It reports state-of-the-art results for RNA representation learning, full-transcript property prediction, and zero-shot mutation-fitness modeling. The same backbone is also used for full-length mRNA generation and redesign, including synonymous-codon optimization that preserves protein sequence. HF Daily Papers' note
score 5