Impact of training data on deep learning-based RNA 3D structure quality assessment.

Training data composition strongly shapes cross-dataset generalization of deep learning RNA 3D structure quality predictors, with RNA3DQA giving the best overall accuracy and ranking stability for ARES.

• Why it matters: Reference-free quality assessment is needed for large-scale RNA modeling, but how training sets affect deep learning predictors’ robustness across benchmarks is unclear.
• What they did: The study systematically cross-trained three RNA 3D quality predictors—retraining each on alternative nonnative datasets, including a new RNA3DQA set, then testing on independent RNA-Puzzles benchmarks.
• The result: Cross-dataset performance differed by model: predictors often generalized better from nonnative training, RNA3DQA improved overall accuracy and ranking stability for ARES, while native sets were better for RNA3DCNN and lociPARSE.

AI summary of the abstract · may contain errors · not medical advice · read the paper.