just saw this; I'm part of the computational modeling team that worked on this -- can try to field any questions or find more qualified people to do so.
1. Would you advertise this tool as a visualization to help with future research and understanding cells OR a possible diagnostic aid?
2. Is there any project that aims to apply these tools to find changes in cells of an aging organism? Do you think that would be useful?
3. Is it possible to figure out for any given class of cell how much of its volume is understood? e.g. "there's this little part and we have no idea what's going on there" or "this protein is everywhere and we can't figure out what it does".
4. How can you evaluate the correctness of your probabilistic model? Neural nets and auto-encoders are known to produce bad results. as an exaggeration, you wouldn't want to have this as your model of human face: https://zo7.github.io/img/2016-09-25-generating-faces/random...
And thanks for publishing the source code for training!
1. All of the above. The label free tool in particular gives you such a big free lunch at the microscope that the combination of it and good visualization has the potential to massively impact research workflows.
2. We are very interested in how cells change as they divide, differentiate, age, are perturbed by their environment, etc. We study cells in culture right now -- getting good images of in vitro cells from multicellular organisms is way harder. So yes it would absolutely be useful. I don't know if we're going to tackle it ourselves, but one of our core missions is to lay the groundwork for the community to take our tools and run with them -- it's a big win for us if we can bring previously unfeasible research within the realm of the possible.
3. I am a Bayesian at heart, so modeling uncertainty is something that I'm always thinking about. It's high on my list of priorities to do something along these lines.
4. Image similarity is a hard problem. At the end of the day, metrics only get you so far and the proof is in the pudding. Unfortunately there is no ground-truth data to test against -- the probabilistic model was constructed exactly because we can't measure where everything is all at the same time. Some things we do to convince ourselves that we're on track is to see that the variation in the imputed predictions and the actual data are statistically similar, and to see if experts are confounded in differentiating the outputs of our models from actual data. You can read more here https://www.biorxiv.org/content/early/2017/12/21/238378
That's at least close to true. It might be better to think of it as a smoothed "segmentation" of a 3D image, i.e. some algorithm decided what pixels are officially part of e.g. the mitochondria set, and outlined them. That could be based on level sets or seeded watershed or whatever else works well.
There are some alternate visualizations here http://www.allencell.org/3d-cell-viewer.html of data that came off of our microscopes that we also use to visualize our models in house but wasn't;t included in the video. It's hard to visualize varying density 3D data in 2D -- there's no one good way to do it, especially on the fly over the web -- but if you have any feedback about what would be more informative / easier to understand, let us know.
Characterizing cellular variation is exactly what we're interested in, e.g when and why are the mitts clustered around the nuclear vs not. Lots of images in our data have them packed around the nucleus -- you can look at the localizations from our microscopes here (select the Tom20 tag): http://www.allencell.org/3d-cell-viewer.html here. You can also grab the raw data (including bright field images e.g. what you would "see" in the microscope) here http://www.allencell.org/data-downloading.html#DownloadImage...
