An RNA Language Model trained on sequence alone reveals the structural logic of Internal Ribosome Entry Sites

Viral RNA genomes are among the most information-dense codes in biology. In picornaviruses, translation depends entirely on Internal Ribosome Entry Sites (IRESes), yet their structures remain largely unresolved. Previous studies either screened short IRES fragments in high throughput or characterized full-length elements individually. Here, we profile 96 full-length IRESes across six cell types, revealing that recently described Type V IRESes double the activity of EMCV, the standard in bioengineering, and that most IRESes exhibit significant tissue tropism. We introduce Albatross, an RNA language model fine-tuned on 50,000 IRES sequences. Trained on sequence alone, Albatross predicts IRES structures with precision comparable to chemical probing, outperforming covariation anal-ysis. We generate structure maps for ∼75,000 full-length IRESes and show that structural discovery scales with model size.