RNA-MSM is the first unsupervised MSA RNA language model based on aligned homologous sequences that outputs both embedding and attention map to match different types of downstream tasks.

The resulting RNA-MSM model produced attention maps and embeddings that have direct correlations to RNA secondary structure and solvent accessibility without supervised training. Further supervised training led to predicted secondary structure and solvent accessibility that are **significantly more accurate than current state-of-the-art techniques**. Unlike many previous studies, we would like to emphasize that we were **extremely careful in avoiding over training**, a significant problem in applying deep learning to RNA by **choosing validation and test sets structurally different from the training set**.