Eterna100-V2: designing RNA sequences for 100 target folds

Eterna100-V2 is a set of 100 RNA secondary structures, made by players of the online game Eterna, for testing RNA design methods: for each one, find a sequence whose predicted lowest-energy fold is exactly that structure. Players have solved all 100 under the ViennaRNA 2 folding engine. The best program, DesiRNA, solves 92 on one CPU core within 24 hours per puzzle, and its authors report 97 running replicas in parallel.

Task

Each puzzle gives a target structure $y^\star$ of 12 to 400 nucleotides in dot-bracket notation, where matching parentheses are base pairs and dots are unpaired nucleotides. A method outputs a sequence $x$ over A, C, G, U of the same length, and the puzzle is solved when the minimum-free-energy (MFE) structure of $x$ is the target:

\mathrm{MFE}(x) = \arg\min_{y} \Delta G(x, y) = y^\star,

where $y$ ranges over pseudoknot-free structures with A–U, G–C and G–U pairs and $\Delta G$ is the Turner 2004 nearest-neighbour free energy at 37 °C computed by ViennaRNA 2.6.4. In practice, RNAfold -T 37 must print exactly $y^\star$ ; on a tie, the printed structure counts. The score is the number of puzzles solved out of 100, with at most 24 hours per puzzle.

Background

Finding a sequence for a given fold, called inverse folding, underlies the design of synthetic RNAs that regulate genes, sense molecules or guide CRISPR editing. Anderson-Lee et al. compiled the original Eterna100 in 2016 from player puzzles spanning a wide range of difficulty, using ViennaRNA 1.8.5. Under ViennaRNA 2 parameters, players found 19 of those targets unsolvable, so Koodli et al. had players redesign them with minimal changes. Eterna100-V2 is the version for the Vienna 2 engine that current tools use.

Data and evaluation

Data: data/eterna100_puzzles.tsv in the eterna100-benchmarking repository (MIT licence), with names, V1 and V2 targets, Eterna puzzle IDs and sample player solutions.
Split: none. All 100 targets are the test set, with no training data and no seeds; in the benchmark's runs, each method attempted each puzzle once.
Engine: a stock ViennaRNA 2.6.4 build run with no option other than the temperature (Turner 2004 parameters, dangles = 2). The repository notes that 2.1.9 gives identical MFE structures on Eterna100 designs.
Limit: 24 hours per puzzle on one CPU core, as in the benchmark authors' runs. Results with more cores are a separate comparison.
Code: scripts/util/fold.py checks a fold, scripts/benchmark.py runs a method on one puzzle and scripts/stats.py counts solves; the authors' runs are in data/results.tsv.

What is known

Eterna players: 100 of 100 with Vienna 2 (Koodli et al., 2025).
With ViennaRNA 2.6.4, one core and 24 hours per puzzle: NEMO 86, LEARNA 69, SentRNA 68, EternaBrain 49 and RNAinverse 20 (Koodli et al.), and DesiRNA 92 in the repository's own run.
DesiRNA as reported by its authors: 97 within 24 hours with Turner 2004 parameters, of which 95 within an hour and 85 within a minute (Wirecki et al., 2025). DesiRNA runs its replicas, ten by default, as parallel processes.
None of the six methods solved puzzles 68 (Taraxacum officinale), 99 (Shooting Star) or 100 (Teslagon) in the benchmark's runs, and DesiRNA's published solutions omit them. Four more (71, 79, 86 and 90) went unsolved by all six in those runs but are among DesiRNA's 97. Each of the three is a separate problem here.

What would count as progress

More than 92 targets solved under the benchmark's rule (stock ViennaRNA 2.6.4, one core, 24 hours per puzzle), the best in its runs, published with code, settings and the sequences. Each sequence checks with RNAfold -T 37 in seconds.
98 or more targets with more hardware, reported with the cores and time used; DesiRNA's authors reach 97 with parallel replicas.
An automated solution to puzzle 68, 99 or 100.

References

R. V. Koodli et al. Redesigning the Eterna100 for the Vienna 2 folding engine. bioRxiv (2021, revised 2025). doi:10.1101/2021.08.26.457839
J. Anderson-Lee et al. Principles for predicting RNA secondary structure design difficulty. J. Mol. Biol. 428, 748–757 (2016). doi:10.1016/j.jmb.2015.11.013
T. K. Wirecki et al. DesiRNA: structure-based design of RNA sequences with a replica exchange Monte Carlo approach. Nucleic Acids Res. 53, gkae1306 (2025). doi:10.1093/nar/gkae1306
R. Lorenz et al. ViennaRNA Package 2.0. Algorithms Mol. Biol. 6, 26 (2011). doi:10.1186/1748-7188-6-26

Source

Imported from Eterna100 V2, curated by Eterna / Das Lab, Stanford. Open according to the source as of 2026-10-20. The statement above is written in our own words; see the source for the original wording, history and discussion.