A validation approach for RNA structures based on deep-learning 3D-structure predictions

Understanding the molecular mechanisms of life depends on accurate macromolecular structures. RNA, in particular, poses a particular challenge for structural biology due to its inherent flexibility and dynamic nature. The challenging nature of structure determination experiments and their interpretation have led to the development of validation approaches to prevent or fix errors in solved structures.

Existing methods for RNA structure validation can, as with proteins, be broadly split into two types. Map-based validation tools seek to detect discrepancies between the experimental data and the determined structure. They are excellent for fixing all types of errors, but only when high-resolution experimental data are available. Model-only validation methods, which usually focus on molecular geometry, are excellent for fixing short-lengthscale issues even when experimental resolution is low. Conkit-validate is a mapless method for protein validation, which is not based on chemical plausibility, but rather on the similarity between experimental models and structure predictions for the same protein. By comparing contact patterns, this tool leverages evolutionary covariance information, learned by state-of-the-art structural prediction models, into a mapless validation method capable of detecting sequence assignment issues, with notable success in detecting register errors.

Here, we extend the conkit-validate method to work on RNA. The results show performance similar to that previously achieved for proteins. In a scan of the PDB, its detection rate does not depend on the local resolution of the density, so it remains applicable where map-based approaches are not, and case studies show it can identify register errors where the density alone cannot decide between models. Together, this highlights the complementary value of a contact-prediction based validation approach using information orthogonal to map- and geometry-based approaches to ensure the quality of deposited RNA structures against varying criteria. The performance of the method can be expected to track ongoing improvements in RNA structure prediction methods.