PickYOLO
The Marlovits group and collaborators at Helmholtz Imaging/DESY have developed PickYOLO, an ultrafast particle finder for cryo electron tomograms.
To understand mechanisms of complex molecular machines such as Salmonella’s Type three secretion system, scientist use an imaging method called cryo electron tomography (cryo-ET). Using a technique similar to that of a medial CT scan, cryo-ET generates multiple images of cryogenically frozen samples by rotating and titling them at different angels. All the images are then combined to generate 3-D reconstructions known as tomograms which can be further refined to reveal high resolution molecular structures.
Localizing individual molecular particles within a busy tomogram is known as picking. “Picking is a bit like finding a two-letter word in a large, crossword puzzle,” explains CSSB group leader Thomas Marlovits “it’s a laborious and time-intensive process and we needed to find a way to speed it up.”
The Marlovits group combined forces with the group of Philipp Heuser, an expert in AI based technology at DESY, and began to develop PickYOLO, an ultrafast particle finder for cryo EM tomograms. Based on the deep-learning real-time object recognition system YOLO (You Only Look Once) used in self-driving cars, the detector was trained with the coordinates of representative particles and tested on single particles, filamentous structures, and membrane-embedded particles. “Using PickYOLO all individual particles in tomograms can now be identified in just a few seconds,” notes co-first author Erik Genthe from the Heuser group “this compares to 20 to 30 minutes with other programs. PickYOLO has greatly increased picking efficiency – this is truly a game changer.”
The researchers are confident that PickYOLO can be used to help to optimize cryo ET workflows. “Picking particles was one of the bottle necks in developing a high throughput process for cryo ET,” explains co-first author Sean Miletic “with PickYOLO you always have an immediate view of your particles and get intermediate results quite quickly.”
PickYOLO is not only a great resource for the cryo-EM community but also an example of the novel innovations resulting from CSSB’s collaborative concept. “It was great to work closely with the DESY AI Unit,” notes Marlovits “This was a true community effort and it’s great to see CSSB’s integrative, multi-institute concept bearing fruit.”
PickYOLO can be downloaded here:
https://gitlab.desy.de/ric/helmholtz_imaging/pickyolo
Source:
Genthe E, Miletic S, Tekkali I, Hennell James R, Marlovits TC, Heuser P (2023) PickYOLO: Fast deep learning particle detector for annotation of cryo electron tomograms. J Struct Biol. 215(3):107990. doi: 10.1016/j.jsb.2023.107990.
