An 80M RNA model reuses one core four times and beats Evo 2 40B on 70 fitness assays

Inference depth in an RNA foundation model can now be chosen after training, without retraining or swapping checkpoints. Conventional RNA foundation models couple parameter capacity and inference depth inside a fixed architecture, so each checkpoint executes one prescribed computation graph. On the 70-assay RNAGym fitness benchmark, raising refinement from one to four iterations lifts aggregate absolute Spearman correlation from 0.242 to 0.286, an 18% relative gain, and the gain is retained through eight iterations.

The shared core receives the original sequence information at every iteration while retaining only a bounded fraction of the previous state, so repeated application does not discard the input. The method, Eco-RNA, is a learned iterative refinement: an input-conditioned Transformer core is applied repeatedly, and the iteration count is selected after training at fixed parameter count. Training combines masked reconstruction at sampled terminal depths with a relative terminal-displacement penalty and a one-sided reconstruction anchor that adds supervision only when a deeper endpoint's reconstruction loss exceeds a shallower reference on the same trajectory.

The benefit of depth tracks prediction granularity, with coarse classification saturating early and finer outputs still improving at larger depths. On 13-class nRC family classification, accuracy reaches 98.69% at four iterations and then stays essentially flat. On chemical-probing reactivity over 114,836 held-out sequences, all three probe means reach their best observed values only at eight iterations; on bpRNA TS0 secondary structure, recall rises 3.3 percentage points from one to eight iterations while precision stays around 83% to 84%.

The report covers Eco-RNA evaluations on RNAGym fitness prediction, nRC family classification, bpRNA TS0 secondary-structure prediction, and Ribonanza chemical-probing reactivity prediction, for readers working on RNA sequence modeling and biological foundation models. Fitness prediction scores the pretrained checkpoint directly without task-specific training; classification, structure, and reactivity each start from the same Stage-B pretrained backbone and are fine-tuned into separate copies, with the adapted checkpoint and head held fixed across the depth sweep. Model weights, inference code, and evaluation configurations are to be released publicly at a later date.

Saturation depth differs across tasks: splicing and aptamer assays improve most and start from the weakest first-iteration scores, while the coding assay improves more gradually. Within one assay, different metrics can prefer different depths, since splicing Spearman correlation and AUC peak at four iterations while absolute MCC keeps improving at eight. The authors therefore recommend choosing depth per assay and target metric using validation, rather than fixing a single depth.