Which RNA secondary structure package best predicts experimental data? Hannah Wayment-Steele used high-throughput structural data from Eterna to answer this question, and found a suprising result: lesser-known packages developed by statistical learning performed notably better than more widely-used packages. Motivated by these results, she developed a multi-task learning algorithm trained on Eterna data, named EternaFold, which showed the best predictions on Eterna data as well as completely independent datasets of viral RNAs and synthetic mRNAs important for vaccine design. We hope that the EternaFold algorithm will be useful for improving RNA design, and that the EternaBench database will provide an independent platform to benchmark more improvements in secondary structure prediction.
