Classification of polyhedral shapes from individual anisotropically resolved cryo-electron tomography reconstructions

Sukantadev Bag, Michael B Prentice, Mingzhi Liang, Martin J Warren, Kingshuk Roy Choudhury

Cryo-electron tomography (cryo-ET) enables 3D imaging of macromolecular structures. Reconstructed cryo-ET images have a "missing wedge" of data loss due to limitations in rotation of the mounting stage. Most current approaches for structure determination improve cryo-ET resolution either by some form of sub-tomogram averaging or template matching, respectively precluding detection of shapes that vary across objects or are a priori unknown. Various macromolecular structures possess polyhedral structure. We propose a classification method for polyhedral shapes from incomplete individual cryo-ET reconstructions, based on topological features of an extracted polyhedral graph (PG). We outline a pipeline for extracting PG from 3-D cryo-ET reconstructions. For classification, we construct a reference library of regular polyhedra. Using geometric simulation, we construct a non-parametric estimate of the distribution of possible incomplete PGs. In studies with simulated data, a Bayes classifier constructed using these distributions has an average test set misclassification error of < 5 % with upto 30 % of the object missing, suggesting accurate polyhedral shape classification is possible from individual incomplete cryo-ET reconstructions. We also demonstrate how the method can be made robust to mis-specification of the PG using an SVM based classifier. The methodology is applied to cryo-ET reconstructions of 30 micro-compartments isolated from E. coli bacteria. The predicted shapes aren't unique, but all belong to the non-symmetric Johnson solid family, illustrating the potential of this approach to study variation in polyhedral macromolecular structures.