RNA Motif Detection, Structure Prediction and Analysis in Cryo-EM Density Maps: A Deep Learning Approach

Abstract

RNA function is closely linked to recurring secondary structural motifs, yet automated RNA three-dimensional structure determination from cryo-EM density remains limited by the lack of large, motif-resolved training datasets. This thesis develops a motif-aware pipeline for RNA structure prediction from cryo-EM density maps through three contributions. First, it introduces \dataset{}, an open-source dataset of over 100,000 motif-resolved density segments spanning 25 RNA secondary-structure classes and 1.5--34.0~\angstrom{} resolutions, with voxel-level backbone, ribose, and nucleobase labels validated using cross-correlation and Q-score measures. Second, it presents Motif 3D CNN, which classifies five consolidated motif classes with 96.8\% peak validation accuracy and 0.948 specificity on high-resolution maps. Third, it develops an end-to-end pipeline that localises motifs using FFT template matching, classifies candidates with Motif 3D CNN, segments the phosphate-sugar backbone using a MONAI UNETR transformer, and derives phosphorus-atom coordinates through skeletonisation and chain ordering. The resulting backbone probability maps achieve correlations of 0.1219--0.996 across hairpin and bulge motifs at varying resolutions. Together, these contributions provide the datasets, models, and pipeline required for motif-driven RNA modelling from cryo-EM density.