4 ms·

This is fascinating to watch. I understand that they get the proteins from PDB/ChimeraX, but how much manual process is involved to map and place the individua

This is fascinating to watch. I understand that they get the proteins from PDB/ChimeraX, but how much manual process is involved to map and place the individual proteins? The paper says it gets the protein locations from CryoET tomograms, but I'd be surprised if these let you automatically identify which proteins are where, and exactly how to place them, and even less so how the ligomers bind together to form larger structures - for example, in the video the membrane surfaces are very smooth, and look almost textbook picture perfect, which suggests they come from a hardcoded model or are smoothed in some way. One part of the paper mentions subatomic averaging from the tomogram, but another mentions: > From a tomogram (Figure3a), we select the particles and determine the orientation of the crown of the spike, as well as the stalk that connects the spike to the membrane Is this a manual process, where the researcher is using his mental model of how the proteins are fit and placing/rotating individual proteins? Or do the tools they developed let this be automated. Both are impressive! If the former, I'm blown away by the effort it must take to make these kinds of videos.

COGlory 3y ago
It's a guided, but automated process. EMAN2 (the software used/partially written by the author), for instance, has a convolutional neural network particle picker. So you can pick a few particles by hand, pick some noise, pick some bad particles, train a neural network, and then inference that network to pick the remaining particles throughout the tomogram. There are a variety of other methods too. There is simple 6 dimensional real-space cross correlation. You can place points by hand, or according to a model. For instance, if you are trying to identify viral spike proteins, and the virus is spherical, and the spike proteins are always on the surface of the sphere, you have a great starting point. So you can say "place points at n interval along the surface of this sphere" and then oversample the spherical surface. You then take a reference volume (can be generated a number of ways), and check each point you placed to see how well it matches the reference volume. You can allow for rotations and translations of the reference volume at each point, and if you find points that are too close together, you can merge them automatically. If you have a high contrast, relatively static protein (such as a ribosome), you can do 2D template matching in the tomogram, where you use a central slice (or maybe a collapsed projection) to do cross correlation in 3 dimensions instead of six (X translation, Y, translation, Z rotation). Or you can beef that up with more neural network/YOLO type stuff. EDIT: To expand on this, continuous density like membranes can be roughly modeled just with techniques like thresholding, watershedding, etc. There are some neural nets such as Membrain [0] and tomoseg [1] (also by the author of this paper), but membranes certainly are trickier. I typically segment membranes by hand (and do so rarely). [0] https://www.sciencedirect.com/science/article/pii/S0169260722003728 https://www.sciencedirect.com/science/article/pii/S016926072... [1] https://blake.bcm.edu/emanwiki/EMAN2/Programs/tomoseg https://blake.bcm.edu/emanwiki/EMAN2/Programs/tomoseg

COGlory 3y ago

It's a guided, but automated process. EMAN2 (the software used/partially written by the author), for instance, has a convolutional neural network particle picker. So you can pick a few particles by hand, pick some noise, pick some bad particles, train a neural network, and then inference that network to pick the remaining particles throughout the tomogram. There are a variety of other methods too. There is simple 6 dimensional real-space cross correlation. You can place points by hand, or according to a model. For instance, if you are trying to identify viral spike proteins, and the virus is spherical, and the spike proteins are always on the surface of the sphere, you have a great starting point. So you can say "place points at n interval along the surface of this sphere" and then oversample the spherical surface. You then take a reference volume (can be generated a number of ways), and check each point you placed to see how well it matches the reference volume. You can allow for rotations and translations of the reference volume at each point, and if you find points that are too close together, you can merge them automatically. If you have a high contrast, relatively static protein (such as a ribosome), you can do 2D template matching in the tomogram, where you use a central slice (or maybe a collapsed projection) to do cross correlation in 3 dimensions instead of six (X translation, Y, translation, Z rotation). Or you can beef that up with more neural network/YOLO type stuff. EDIT: To expand on this, continuous density like membranes can be roughly modeled just with techniques like thresholding, watershedding, etc. There are some neural nets such as Membrain [0] and tomoseg [1] (also by the author of this paper), but membranes certainly are trickier. I typically segment membranes by hand (and do so rarely). [0] https://www.sciencedirect.com/science/article/pii/S0169260722003728 https://www.sciencedirect.com/science/article/pii/S016926072... [1] https://blake.bcm.edu/emanwiki/EMAN2/Programs/tomoseg https://blake.bcm.edu/emanwiki/EMAN2/Programs/tomoseg