Back

ATOM-1: A Foundation Model for RNA Structure and Function Built on Chemical Mapping Data

Boyd, N.; Anderson, B. M.; Townshend, B.; Chow, R.; Stephens, C. J.; Rangan, R.; Kaplan, M.; Corley, M.; Tambe, A.; Ido, Y.; Yukich, J.; Tcheau, T.; Abdeldayem, A.; Ferns, G.; Patel, H.; Barman, S.; Schleck, A.; Sanborn, A. L.; Eismann, S.; Townshend, R. J. L.

2023-12-14 bioinformatics
10.1101/2023.12.13.571579 bioRxiv
Show abstract

RNA-based medicines and RNA-targeting drugs are emerging as promising new approaches for treating disease. Optimizing these therapeutics by naive experimental screening is a time-consuming and expensive process, while rational design requires an accurate understanding of the structure and function of RNA. To address this design challenge, we present ATOM-1, the first RNA foundation model trained on chemical mapping data, enabled by data collection strategies purposely developed for machine learning training. Using small probe neural networks on top of ATOM-1 embeddings, we demonstrate that this model has developed rich internal representations of RNA. Trained on limited amounts of additional data, these small networks achieve state-of-the-art accuracy on key RNA prediction tasks, suggesting that this approach can enable the design of therapies across the RNA landscape.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.