DevoWorm Tutorials, 9 Apr 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: DevoLearn Tutorial (13 minutes)

Description

A tutorial on the DevoLearn platform, April 2021. Presented at the INCF Assembly (virtual, 2021). Contributors: Bradly Alicea, Mayukh Deb, Ujjwal Singh, and DevoWorm group, especially our Google Summer of Code students and applicants.

A

Hello, my name is bradley alicea and I'm the head of the diva worm project, and today, I'm going to tell you about diva, learn engaging learners with computational developmental biology.

A

uh This is a project that I've worked on along with mayuk deb lujwal singh in the diva worm group mayo can usual were jisock students in 2020, and so this project grew out of their efforts, but also the efforts of previous gsoc students. In previous years. You can visit our website at divaworm.weebly.com and there you'll find a list of our other collaborators, and people contributed in different ways.

A

You'll find github.com divaworm, which is our github repository, and that has a lot of uh our code that for different projects that we have and finally github.com divo learn, is the place where you'll find this not only the diva learning software, but this entire platform, and so what is uh diva worm.

A

So one of the things we do is we're involved with the open worm foundation and we're sort of the developmental arm of the open norm foundation and the open room foundation's mission is to build the world's first digital c elegans in an open source fashion. So it's one of our uh commitments.

A

The other is to explore development, and this is very broad, I say developing the nematode and other creatures. We really do focus on c elegans, but we're also focused on a number of other model organisms, and we take a very broad approach to development.

A

We use secondary data. We use other types of approaches to build simulations, analog data, analysis and theory building to understand development. Better often, we can talk about ourselves as a computational developmental biology group or data science biology group. Finally, we are very invested in open data, our entire.

A

We have some collaborators, but a lot of our activities are based around open data, and so we have this platform devo zoo, where you can find openly available. Primary secondary tertiary data sets all annotated and put together in a way for people to come to the group and use so this includes data such as segmented, microscopy images, public repositories of different types of data and literature mining. Although we have a broader array of data that are available to us.

A

Finally, we've contributed to the building the neuro commons special issue of neuroinformatics. There are a number of incf people contributing to that volume, but if you want to read our article, this is the doi identifier for it.

A

So this is the diva learn platform and you'll find it at github.com divalern, and it consists of three parts. So the first part is this: diva learn 0.3.0, and this is the standalone diva learn program. It's a pre-trained model for machine learning for deep learning, so it takes data and it does a sort of it pre-trains the model so that you don't have to train it with labels.

A

This is a nice sort of advance. We worked on this for a couple years kind of putting this together and getting this ready to go in a gsock project and uh myok deb is the maintainer of this project. So we started out. I think it's 0.2.0 and he's been maintaining it up to 0.3.0.

A

You can find it on pip.org that project, evil, learn and we've actually had. You know he's not only maintained it, but we've had a lot of contributors trying to apply for gsoc 2021, who have also made contributions to the code base. So you see up at the top.

A

You have the number of downloads of the package and you can see that it's increased since january when we've been getting people involved and trying to build it up in into something that's more substantial, a package, so the other part of the one of the other parts of diva learn is diva worm, ai, which are species-specific machine learning or deep learning models, and so we're interested in in this case we're interested in not in the pre-trained models, but in regular machine learning models but specialized for different types of data.

A

So in this case, we have one model: that's specialized for c elegans embryogenesis. We have another model, that's specialized for diatoms, and yet there are other models where we're going to be. You know using different techniques to get a different biological questions.

A

One of the more exciting uh projects that we have going is the uh sort of a neural cellular automata model, and so this is where you combine neural networks or deep learning, networks and cellular automata and of course, cellular automata are a good way to model morphogenesis, and so this is yet another thing that we're going to be putting into this uh platform. So we're constantly uh building up this platform, you can find divorm ai at evwarm.github.io divormai.

A

So then, the final part of the stevenlearn platform is the devozu, and so the devozu is this collection of secondary data. It's organized by species, and we give you uh publicly available data sets, we annotate everything and we give you a list of things to download, and so anyone can come into our group and conduct an analysis and it might be a comparative analysis. It could be an analysis of a certain species and you know they're whatever they would like to do.

A

It could be machine learning, but it also could be just a regular, more straightforward biological question.

A

So in our devo zoo we have access to zebrafish data, drosophila data, c elegans data and even spiders, and so spider embryos are wonderful models and we have some of those data too.

A

So we have this uh further goal of looking at quantitative morphology, using deep learning, um and we we're doing this in a number of ways, as I said before, there's the devo zoo, but we also have these different target models that we're going to be looking at it now and in the next couple years. The first is a model of axolotl brain development, so we have a collaborator who's. Collecting data on axolotl embryos and the thing about axolotl embryos is that they have a transparent surface.

A

So you can see the brain as it develops in these embryos, and so this is the embryo at the top, and this is the acts the fully grown axolotl at the bottom, and you can actually observe this process and our collaborator is able to acquire these images of uh axolotl embryos using a number of really innovative techniques. For example, she has a ball microscope which takes images from different angles of the surface.

A

She also has a microscope that inverts the embryo using a water column, and so all these things are available for people to analyze. Further, you can see the c elegans embryogenesis example on the bottom. This is a screenshot from the devo learn 0.3.0 software. This is where you have you have you can build these animations?

A

You can segment the cells in the build animation, so you get these time lapse images or these time series of an embryo, or you know, looking at different uh frames of the acquisition of the embryo, and so you and you can also put it into a coordinate system, much as like you might do with cell tracking. So you can reconstitute cell tracking in this way, uh you can also look at uh diatoms and the species of dye or the genus that diatom we're interested in is basil area.

A

So this is a diatom that live that has this rod-like phenotype and lives in colonies and these colonies move around and we're able to use uh deep learning to segment the images and look at it as it's moving. So these images are all you know these. These colonies are always moving around.

A

um We've developed a technique to pull apart some of these images and look at how they sort of move relative to one another in the colony and we're getting these images from another collaborator. Who is able to do some actually pretty interesting microscopy with respect to this organism? So our sort of our end goal is to conduct theory building or maybe, like you know, straightforward analyses, testing, hypotheses to explain developmental processes, and you can see we're interested in this process where you go from a sphere to this differentiated organism.

A

So I mentioned before that. This is google summer code is an integral part of this platform, developing it. So we've been preparing for google summer of code. This is where we provide education and evaluation for students in deep learning and machine learning, but also people who are interested in developmental biology may not have the background to join a lab, and so we offer this sort of bridge, and so by you know, getting involved in the diva owner platform is an excellent way to do that.

A

We've made calls for involvement and some of those have involved g suck and some of those have involved just regular calls for involvement. uh People will contribute to a github issue during the application period or whenever they want. We've had a lot of contributors so far, this gsox season where people are trying to apply for their projects and in the process, they issue a number of pull requests, and my oak has been good enough to keep on top of that and maintain these pull requests.

A

And, uh finally, they get to join our weekly meetings. We have a meeting once a week and this helps them to get involved in bigger projects to gain a better perspective on what they're doing in the code base. So if they're just issuing a pull request, you know you can see the code and you can see the net the deep learning uh problem, but you don't actually know why and so we're sort of bridging that gap for students um and so this divalerm package.

A

Of course, this is, uh you know, maintained by mayok and ujwal, and myself and we've had a lot of contributors who have done not necessarily developed whole repositories, but they've developed, they've they've committed a number of made a number of pull requests, and so um myo can usual, though, have been the main maintainers and developers of this, and so a lot of things we're doing developing things aside from divorm, ai and evil learn involve data science, demos, so we've had a number of people create jupiter, notebooks and methodological tutorials on different topics.

A

This is in an area we call sort of data science, or you know some sort of collection of demos that people would need if they want to say, for example, take the next step from the deep learning part which is like you know, segmenting images and finding pattern broad patterns in the data to something like doing a data analysis, and they can learn that in this platform we also have educational resources. So we look at teaching computational students about developmental biology, so we have in some of our devo zoo materials.

A

We have educational stubs, where you know we allow people to read about a model organism find out why it's important, but we also have these epistemological directories for lifelong learners, and this is an idea that we've come up with.

A

um Actually, I've helped come up with in another context where you're building these things on these version controlled sort of dictionaries on github, where you have a number of different topics that are essential to learn so for lifelong learners, who don't have any background developmental biology, what's important to learn about developmental biology or what's important, to learn if you're, a developmental biologist about machine learning, if you can do that, if you know you can provide students with those kind of cross fertilizing opportunities, then it will improve their ability to contribute.

A

Finally, we do a number of things with theory building we help to build explanatory frameworks, hypothesis generation, so we're doing this big data type stuff, but we're also doing it with an eye towards hypotheses and theories, and things like that.

A

So thank you. If you want to contact me, this is my email address at openworm.org. You're welcome to join our slack via this launch pass and look forward to seeing you around. Thank.

A