AutoRNA: RNA tertiary structure prediction using variational autoencoder.

Abstract

Understanding the three-dimensional (3D) organization of RNA is essential for advancing therapeutic development and vaccine design. However, the limited availability of experimentally resolved RNA structures restricts the applicability of data-intensive machine learning approaches for tertiary structure prediction. This study aims to develop a data-efficient method for learning coarse-grained RNA structural organization directly from sequence information. We propose AutoRNA, a variational autoencoder-based model that learns sequence-conditioned structural priors in the form of inter-nucleotide distance matrices. The model was trained on RNA structures obtained from the Protein Data Bank and restricted to sequences up to 64 nucleotides. Predicted distance matrices were converted into 3D coordinates using multidimensional scaling, followed by template-based assembly and molecular dynamics refinement. Model performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), global distance test (GDT), and template modeling (TM) scores. On the test dataset, AutoRNA achieved an RMSE of 4.49~\AA{} and an MAE of 3.13~\AA{} for predicted inter-nucleotide centroid distances. However, the reconstructed three-dimensional structures showed limited global fold recovery, as indicated by moderate GDT scores and low TM-scores. Performance decreased for longer sequences, indicating limitations associated with data scarcity and increased structural complexity. Molecular dynamics refinement provided modest improvements, particularly for initially low-quality predictions. AutoRNA demonstrates that variational autoencoders can learn meaningful coarse-grained structural representations of RNA from limited data. While not suitable for near-native tertiary structure prediction, the method generates candidate coarse-grained inter-nucleotide distance restraints that could potentially be incorporated into downstream structure-reconstruction or physics-based refinement workflows. This work highlights the potential of generative models for RNA structure prediction in low-data regimes.