Towards inferring atomic scale conformation landscape of biomolecules from cryo-electron tomography data

Understanding continuous conformational variability of biomolecular complexes at atomic resolution is essential for linking structure to function, but remains challenging for cryo-electron tomography (cryo-ET) data due to high noise and missing-wedge (MW) artifacts. Physics-based methods, such as MDTOMO (based on classical molecular dynamics simulations of a given atomic structure to flexibly fit subtomograms), provide accurate estimation of atomic coordinates from subtomograms, but their computational cost limits large-scale applications. We present DeepMDTOMO, a supervised deep learning (DL) framework that uses a given set of pairs of atomic coordinates and subtomograms to learn their relationships in order to predict unknown atomic coordinates from a previously unseen set of subtomograms. The proposed regressor encoder-decoder architecture combines 3D convolutional extraction of features from subtomograms with a multilayer perceptron to predict Cartesian all-atom coordinates. Experiments on synthetic datasets show that DeepMDTOMO achieves low errors of coordinate prediction in presence of noise, MW, and large continuous conformational variability. Additionally, fine-tuning to a new motion demonstrates that learned representations capture general structure-density relationships rather than specific patterns. The results presented are encouraging and motivate future studies on speeding up subtomogram flexible-fitting methods with DL for fast atomic-scale conformational landscape determination from cryo-ET data.