Fig. 1: Holistic RNA structure determination using HORNET.
From: Determining structures of RNA conformers using AFM and deep neural networks
a, Overall workflow for HORNET. The input comprises AFM topography data (x, y and z dimensions) from an experimental AFM image and an initial model. Dynamic fitting is driven by the AFM and structure-based potentials (Methods). Coarse-grained models are generated from the trajectory of dynamic fitting in the form of energy and topology information containing the complete list of all energy values and the overall fit to AFM topography (CCAFM) associated with each trajectory model. This information is passed to the unsupervised machine learning (UML) or/and DNN for clusterization and estimation of the accuracy of each model in terms of r.m.s.d. relative to the ground-truth structure. Convergence is defined as a distribution of estimated accuracy with a population of models with r.m.s.d. below 7 Å. If the DNN process has converged, the top 10 models are converted to all-atom coordinates. If convergence is not reached, the dynamic fitting is performed a second time with a different initial model or a longer time for the dynamic fitting. b, Schematic of the main steps of UML and DNN (Methods). PCA, principal components analysis. c, Representation of the psDatabase composed of 3.5 million continuous trajectory structures of the RPR catalytic domain used to train and optimize the machine learning algorithm. The trained DNN architecture has the capability to estimate the accuracy of the structural model underneath the AFM topography. d, The approximately 56 million trajectory models used for HORNET benchmarking, grouped according to usage: training–training validation–testing (BM0, BM1 and BM2), validation–testing (BM5) and fully blind testing (BM3 and BM4).
