Classification of electron sub-tomograms with neural networks and its application to template-matching

Classification of electron sub-tomograms is a challenging task, due the missing-wedge and the low signal-to-noise ratio of the data. Classification algorithms tend to classify data according to their orientation to the missing-wedge, rather than to the underlying signal. Here we use a neural network approach, called the Kernel Density Estimator Self-Organizing Map (KerDenSOM3D), which we have implemented in three-dimensions (3D), also having compensated for the missing-wedge, and we comprehensively compare it to other classification methods. For this purpose, we use various simulated macromolecules, as well as tomographically reconstructed in vitro GroEL and GroEL/GroES molecules. We show that the performance of this classification method is superior to previously used algorithms. Furthermore, we show how this algorithm can be used to provide an initial cross-validation of template-matching approaches. For the example of sub-tomogram classification extracted from cellular tomograms of Mycoplasma pneumonia and Spiroplasma melliferum cells, we show the bias of template-matching, and by using differing search and classification areas, we demonstrate how the bias can be significantly reduced.