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

Figures

• 

  Figure 1:. A circRNA system enables multiplexed IRES activity measurement. (A) Bicistronic dual reporter assay. (B) circRNA design with barcoded full-length IRESes circularized via Group I Intron ribozyme. (C) Library composition by host (outer) and IRES type (inner). (D,E) Polysome fractionation trace and resolved fractions. (F) Representative polysome distributions for EMCV, HRV27, and BKoV. (G) Polysome heatmap for all 96 IRESes. All data from HEK293T cells.

• 

  Figure 2:. Type V IRESes double EMCV activity and most IRESes exhibit tissue tropism. (A) IRES activity (ribosomes/RNA) ranked across 96 IRESes in HEK293T cells. Select IRESes are labeled. EMCV activity relative to the top-performing Caprine kobuvirus 12Q108 is indicated. p-values from Student’s t-test, n = 4. (B) Ribosomes/RNA grouped by IRES structural type. p-values from Kruskal-Wallis H test with Dunn posthoc analysis and Bonferroni correction (n = 4). (C) Heatmap of ribosomes/RNA for all 96 IRESes across six cell types, with hierarchical clustering. Color scale indicates ribosomes/RNA. (n = 4/cell type). (D) Density distributions of IRES activity by cell type. (E) Tropism significance ( p-value) versus effect size (Gini Index) for each IRES. Dashed line indicates significance threshold after Bonferroni correction. 38 IRESes with moderate to strong tropism. p-values from ANOVA. (F) ROC curves for DMS-MaPseq constrained models of Enterovirus A71, EMCV, and Salivirus.

• 

  Figure 3:. Albatross predicts IRES secondary structure from sequence alone with high precision. (A) Schematic of the dependency mapping approach. IRES sequences are used to fine-tune an RNA language model. For each position, the effect of every possible point mutation on nucleotide predictions is measured, generating a dependency map from which structure is inferred. (B) Dependency map of the EMCV IRES using the pre-trained model (no fine-tuning), showing no discernible structural signal. (C) Dependency map of the EMCV IRES after four epochs of fine-tuning. Antidiagonal features (dashed boxes) correspond to base-paired stems. (D) Albatross-predicted structure of the EMCV IRES overlaid on the experimentally validated secondary structure. Colors of the stems correspond to colored squares in C. Inset shows the dependency map for a stem region with corresponding sequence. (E) Precision per structure across training sets. (F) F1 score per structure across training sets. (G) MicroF1 over training steps for the 200k, 500k, and 50k training sets. Dashed line, pre-trained baseline.

• 

  Figure 4:. Albatross outperforms covariation analysis and uses motif recognition to predict structure. (A) Albatross dependency map of the Coxsackievirus B3 eIF4G-binding Domain V. (B) Albatross-predicted base pairs compared to the crystal structure, with predicted interactions in black and missed interactions in red. Albatross correctly identifies (I) a non-canonical A·C pair and (II) an isolated single G·C stabilizing interaction. (C) RNAstructure Fold thermodynamic prediction for the same region. (D,E) F1 and precision for Albatross versus CaCoFold across 96 IRESes p-values from a Mann-Whitney U test (n = 96). (F,G) F1 and precision binned by number of similar sequences in training set. p-values from a Mann-Whitney U test (n = 47, 44, 5, respectively). (H) BLAST alignment of the Gallivirus Pf-CHK1/GV GNRA region to the training data, showing only four related sequences. (I) Signal-to-noise ratio for GNRA stem in dependency maps of Gallivirus IRES with GNRA mutations. (J,K) Dependency maps of the Gallivirus IRES with GNRA mutations: wild-type (GUAA) and CCCC.

• 

  Figure 5:. Albatross reveals structural variation within IRES types and generalizes across them. (A) Length distribution of Type II IRESes grouped into three bins. (B) UMAP of spectral analysis scores for 671–751 nt Type II IRESes, showing five clusters (IIa–IIe). “Canonical fingerprint” indicated. (C,E) Dependency maps for Sicinivirus MW684796 and EMCV, with structural features boxed. (D,F) Predicted structures for Sicinivirus and EMCV Domain I, highlighting GNRA stem, RAAA stem, and the extended AAG stem in the Sicinivirus variant. (G) Sequence logo of the extended AAG region across related sequences (n = 49). (H) Predicted base pairs scale with model size (33M to 650M parameters) on Type I IRESes. p-value calculated with Page’s trend test (n = 100/model). (I) Cross-type generalization: predicted base pairs by IRES type across model sizes. Models trained only on Type I sequences. p-values from Page’s trend test (n = 20/type/model).