Practical Analysis of Macromolecule Identity from Cryo-electron Tomography Images using Deep Learning

Mostofa Rafid Uddin, Ajmain Yasar Ahmed, Kafi Khan, Md Shahrar Fatemi, Xiangrui Zeng, Min Xu

Abstract:

Cellular electron cryo-tomography (cryo-ET) has
made possible the systematic 3D visualization of the near-native
structures and spatial-organizations of large macromolecules
(represented as subtomograms) and their interactions with or-
ganelles inside single cells. It has emerged as a major tool for
in situ structural biology. However, the systematic identification
of such macromolecules from cryo-ET images is very difficult
due to structural complexity and imaging limits. In particular,
conventional methods are too slow to process millions of highly
structurally heterogeneous macromolecules fastly imaged using
cryo-ET. Since 2017, supervised deep learning has become an
important tool for facilitating high-throughput analysis of cryo-
ET data. However, supervised learning based approaches depends
on manual data annotation by biologists, which is an extremely
time-consuming and burdensome process. Therefore, none of
these methods are practical to use. In order to facilitate deep
learning for practical identification of macromolecules from cryo-
ET images, in this paper, we demonstrate the pathway towards
unsupervised learning for fast and high-throughput identifica-
tion of macromolecules from cryo-ET images. To this end, we
demonstrate the use of three selected recent macromolecule
identification methods on several commonly used benchmark
cryo-ET datasets.