National Cancer Institute researchers have developed a method called HORNET for characterizing 3D topological structures of large and flexible RNA molecules. Scientists used atomic force microscopy (AFM) with deep neural networks and unsupervised machine learning to capture individual conformers under physiological conditions.

Human RNAs are transcribed with structural elements crucial for biological functions. Understanding these structures with conventional methods, such as cryo-electron microscopy, depends on highly homogeneous samples and signal averaging. Large, flexible, heterogeneous RNAs often remain difficult to analyze because they adopt multiple conformations once in solution.

No large RNA structure database exists that correlates sequence with 3D topology. Successful protein-centric methods like AlphaFold remain unavailable for RNA, creating a critical gap in structural biology. The general absence of RNA-specific deep-learning approaches likely reflects the challenges in capturing reliable structural models.

In the study “Determining structures of RNA conformers using AFM and deep neural networks,” published in Nature, scientists introduce HORNET and detail its groundbreaking capabilities for detecting previously hidden large and flexible RNA structural features.

Researchers collected single-molecule AFM images of benchmark RNAs in distinct conformations. Unsupervised machine learning and deep neural networks were then applied to correlate molecular topographies and energy distributions.

The system was trained on a pseudo-structure database covering a broad range of RNA folds and tested on multiple RNAs that exceeded 200 nucleotides in length (RNase P RNA, a cobalamin riboswitch, a group II intron, and the HIV-1 Rev response element RNA). Different initial models were used, including predicted structures and conformers derived from small-angle X-ray scattering data.

Test cases demonstrated that HORNET accurately reconstructed individual RNA conformations, with root-mean-square deviations (a measure of how closely the calculated structure aligns with a reference) frequently falling under the 7 Å threshold widely used to confirm major structural features in large RNAs.

Benchmark experiments with simulated and experimental AFM images confirmed the reliability of combining previously established constraints and AFM pseudo-potentials.

Validations showed that diverse RNase P RNA and HIV-1 Rev response element RNA conformations could be visualized at the single-molecule level. Estimated accuracies from the deep neural networks aligned with actual distances from known structures.

HORNET addresses a significant challenge in RNA structural biology by providing a holistic, direct method for examining previously elusive RNA structures, with profound implications for future research across multiple clinical, pharmaceutical and biotechnology applications.

© 2024 Science X Network