Self-configuring cryo-ET automation for determining structures inside cells

A preprint links molecule detection in cellular images with structure reconstruction

On August 28, the authors posted a preprint describing a cryo-electron tomography workflow that combines Octopi with py2rely to produce three-dimensional images of frozen cells. A researcher manually labels several copies of the target molecule in an image. Octopi finds similar copies, and py2rely combines them to reconstruct their shared structure.

In these images, distinguishing the target molecular complex, a group of molecules that work together, from membranes, internal cell structures, and noise is difficult. Resolving the shape of a single complex requires finding many copies, extracting matching regions, aligning them, and combining them into a common map. Averaging reduces random noise and reveals the shared structure. This method is called subtomogram averaging.

Previously, specialists configured the search for copies of the target molecule and the structure reconstruction separately for each dataset. Octopi learns from a small number of manual labels and selects a neural network and training method suited to the specific three-dimensional images. py2rely then uses the detected coordinates and tilt series acquired at different angles. It extracts the regions, aligns them, and progressively refines the common structural map.

The resulting map provides a check of the entire workflow because it shows whether enough copies were found to reconstruct the structure. On Phantom, a test set of three-dimensional images in which the particle locations are known in advance, Octopi achieved an Fβ score of 0,763, using a metric that gives greater weight to avoiding missed particles. TopCup, a model configured specifically for Phantom, scored 0,761, while the winning competition entry scored 0,788. After identical processing, all three coordinate sets produced structures for six targets at resolutions ranging from 3,5 to 11 angstroms. One angstrom is one tenth of a nanometer.

The authors then applied the combined workflow to 17 tasks involving purified molecules, internal cell structures, and whole cells. They obtained maps at resolutions ranging from 3 to 20 angstroms. A model trained on only three tomograms, which are three-dimensional images of a single sample, detected virus-like particles in three independent Escherichia coli datasets. The resulting maps had resolutions of 3,0–3,9 angstroms. For V-ATPase, a protein complex found in small vesicles in nerve cells, the researchers manually selected suitable groups of regions and specified a mask, which defines the part of the map included in the reconstruction.

The authors released the Octopi and py2rely code, along with the tomograms, tilt series, coordinates, and maps used in the study. Other laboratories can process available cellular tomograms through the same workflow and test whether a small number of manual labels is sufficient to reconstruct the structures of their target complexes.