Improving RNA secondary structure prediction through expanded training data

In recent years, deep learning has revolutionized protein structure prediction, achieving remarkable speed and accuracy. RNA structure prediction, however, has lagged behind. Although several methods have shown moderate success in predicting RNA secondary and tertiary structures, none have reached the accuracy observed with contemporary protein models. The lack of success of these RNA structure prediction models has been proposed to be due to limited high-quality structural information that can be used as traini…

By the Truvace deskOctober 20, 20262 min read1 sourcePeer-reviewedTRV-2026-1980· cite & verify

Improving RNA secondary structure prediction through expanded training data — "R-chie double structure arc diagram by Daniel Lai, Jeff Proctor, Jing Yun and Irmtraud Meyer" by dullhunk is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.

In brief

Researchers built a large, diverse data set of paired RNA sequences and their secondary structures, called RNASSTR, to test whether scarce high-quality training data explains weak RNA structure prediction. They retrained the deep-learning model SincFold on it and also implemented a second model, Lyra-TransPred.

The work matters because RNA structure prediction has trailed protein structure prediction, and the results suggest that data limits are a real bottleneck. What remains open is how far the gains extend to tertiary structure and to accuracy comparable with protein models, which the abstract does not report.

Main points

RNA structure prediction has lagged behind deep-learning protein structure prediction in accuracy.
The proposed cause was limited high-quality structural training data, which the authors set out to test.
RNASSTR pairs RNA sequences with their corresponding secondary structures.
Retrained SincFold performed better on RNA families it had not seen before.
Lyra-TransPred achieved the highest mean F1 and MCC among evaluated models with substantially less training time per epoch.

The gain

Retraining the SincFold deep-learning model on the expanded RNASSTR data set improved its de novo RNA secondary structure prediction on previously unseen RNA families.

The rundown

The study starts from the observation that, although deep learning has transformed protein structure prediction, none of the RNA secondary and tertiary structure methods have reached comparable accuracy.

To probe the idea that training data is the constraint, the authors assembled RNASSTR and used it to retrain SincFold. Evaluation on unseen RNA families tested whether the model could predict de novo structures rather than recall familiar ones.

The authors also implemented Lyra-TransPred, which led the evaluated models on mean F1 and MCC while needing less training time per epoch, and they describe RNASSTR as a foundation for future prediction algorithms.