 work of the adult mouse brain for data integration and discovery. Whoa. Sorry. OK. So at the Allen Institute for Brain Science, our mission is to accelerate the understanding of brain using the big science approach and generating useful public resources. In our current 10-year plan, our aim is to multiply and understand how the brain works with an initial focus on the visual system. So we are trying to find out how the visual information is represented and analyzed by the neurons and how we perceive this information to make the decisions. So to support this effort, we are generating various type of data sets. So how to integrate this multi-scale and a multi-model data sets, it's a challenge by itself. So one of the good ways to organize this kind of data is that we can define a standard coordinates of space and we register or attach every data to that space. So by doing this, first, we are going to have the standard consistent representation of the metadata. And secondly and more importantly, once this is done, we're going to be able to compile and analysis the difference from different subjects or even from different modalities. So this is the reason why in 2005, we build our first version common coordinate framework. That version was based on one single specimen using the nasal staining. And the second major revision was in 2011. At that time, we met the atlas symmetric and added multiple level annotation structures to it. And the major motivation that drive us to do the version 3 is because, as you can see, that nasal-based atlas is not really smooth in 3D due to the distortion introduced during the sectioning, staining, and the tissue transportation. So this time, we decided to move to a new modality. That is the serial two-photon imaging. It's a block-face image, a block-face system, which means that the block-face is going to be imaged instead of the sinc section cut it from the brain. So this will assure us a really smooth 3D volume by just standing together the 2D chrono sections. However, other than the 3D smoothness, there is another one property we really want the average template to have. That is, we want the template to be as generic as possible. However, each brain is unique in its shape. And also due to the imaging process, the artifacts or even the tissue damage will introduce even more uniqueness to that brain. So if you just choose one brain as a template and register all other brains to that template, that will introduce a strong bias toward that brain you just selected. So in order to avoid this, this time, instead of using just one brain, we build our average template based on a population of 1,675 brains. So we start our atlas construction by using the version 2 atlas as the initial template. So then we register all the other brains to that template using a fine registration first. And after each iteration, after the registration is done, we do the average of the waft image to get the average template for the next iteration. And we repeat doing this until there is a little change of that template. And then we introduce the deformable registration. So when the whole algorithm converges, what we have is such an average template with the average shape and average intensity. So if you compare this template to any of the brain in the population on the left side, you can see that first. For some unique features, it's no longer readable in the average template, which is good, actually. And on the other hand, for other features, which is a common feature shared by most of the brain in the population, they have actually been greatly enhanced. For example, the barrel fields and the barrel aids in the brain step. OK, so these two videos shows you the process of how this algorithm converges. So another very interesting example is that we also apply the same algorithm to the batch photo at the institute. So as you can see on both sides, the image starts from the very funny initial image and becomes more and more clear. And for that phase, actually, it's a unisac phase. Blends the man and the woman, so I cannot say it's handsome or it's beautiful, but it's a charming phase. And these two videos show you more details of the template we obtained after this algorithm. So as you can see, the average template we have is really clean and beautiful and full of anatomical detail information. So another infrastructure we build on top of the CCF is so-called cortical coordinates because a lot of effort will be focused on the cortex. So the ability to integrate information from different layers is very crucial. This also means we have to take into account the curvature of the cortex. So we use Laplacian equation to generate the equi-potential field between the pie and the wet surface, which is shown as the color rainbow in figure A. And the streamline you see in figure B and C is generated by going along the gradient direction of that potential field. So with these streamlines, we now have the ability to carry information from different layers to the surface of the CCF. For example, if we do a maximum intensity projection, so from the top of the CCF, we can clearly see the different patch of the somatosensory as well as primary visual and auditory areas. So the data annotation is another very important aspect of the CCF. Currently, we are in the first year of the four-year plan to do the annotation of the whole artist. For each structure, the annotation is led by our neural anatomists. They're using not only just the information from the average template, but also they use some other supplementary information, such as the gene expression and the connectivity information to determine the boundary. And till now, we have 178 structures annotated. And when it's completed in 2017, we are going to have 300-gram matter structures and the cortical layers plus another 80 fiber tracks. So next, I'm going to use the island mouse brain connectivity atlas as an example to show you how we use the CCF as an information hub to do the data integration and representation. So the aim of the connection of atlas is trying to provide a useful public resources to understand how the brain, how the neurons are connected in the brain. So this is achieved by the stereotactic injections of certain values into the brain. And then the population of the neurons get infected to that virus. We'll express a green fluorescent protein. So we then sacrifice a mouse and do the whole brain imaging to get the volume information about that injection site. So this is a standard attribute from one experiment. It consists of 140 chrono sections. Each of that is the 40k by 30k. So the total atlas consists of about 2,000 experiments that is 1,400,000 images and one petabets of data. So we built an automatic pipeline to convert each of the RGB pixels into biological meaningful informations. So the backbone of this pipeline is a series of informatics modules. The pre-processing module gets the image ready for analysis and display. And the detection module segments the signal from the background and the registration module transform each data into the CCF. And the grading module, we grade the data into different resolutions. And the last module is the unionization module, since all the data has been registered as a CCF. So we can sum up the experiment result with regard to each anatomical regions. And also because we have ready all the data into the CCF, we can display all the injection side in such one single image and use it as the navigation of our web app. Here, each dot represents one experiment. The size and the location of that dot is decided by the size and location of the injection. So the user can click any dots. And that will bring the user the 2D original image and a 3D summary of that experiment. And if the user decided to dive into more details of one experiment, we also provide these quantitative read-ups of the projection signal to each of the anatomical regions. And because we have ready all the data to the CCF, we also provide another very useful tool. It's called the spatial search to the users. For the spatial search, the users can click anywhere on the left side of Atlas to select a target walk-through. Now once that walk-through is selected, then all the experiment has signal projected to that target walk-through will be listed on the right side of the page. And the streamlines you see here is actually the link between the injection side and that target walk-through. So this is effectively a virtual-rature grade map. So using this spatial search, we can do something even more interesting. For example, this is the topography of the cortical to sub-cortical projections. The top is a visualization of about 80 injection sites through the brain. And we color-coded it according to its location. And the middle and the bottom image is generated by doing the spatial search repeatedly within the isocortics, striatum, and salamuses. And for each walk-through, we only keep the search, retain the highest retain. That is one of the 80 injection sites. And then we color that walk-through together with the streamlines to the color of that injection site. So this picture basically shows you the primary path between the cortical to sub-cortical connections. And instead of just keep the highest retain, on the update, we also provide a user and another viewer called a Composite Projection Viewer. So in this viewer, the user is able to use .org-gram to see the projection signal from different experiment at the same time. So at the end of the introduction to our Connectional Athletes, I want to show you this scientific art generated by the BrainExplorer, our interactive 3D viewer. So we selected 21 well-typed injections, which spanning the whole cortex. And beside the astonishing beauty, we can also find some very interesting scientific findings. For example, as you can see, all the signal injected to the right side of the brain projected to the other side in a very precise way, except in the primary visual area. You see the spouseness here, right? So if we increase the number of the injections, this spouseness can actually show you the way one area. So this is quite similar to the colossal labeling method. So this can be called as a virtual colossal labeling. Another example to show of using CCApp to do the data integration is the Allen cell type database. It's a database of a single cells characterized by its morphology and electrophysiology properties. So this is the portal page of that database. And the backdrop image of the map is the top surface view of the CCF with the outline showing the primary visual area. And here, each dot represents one cell. And the two strips to the bottom and the right side is can be viewed as a flattened cortex. So which shows the depth of the cell. So when the user moves their mouse around this navigation map, the cell will highlight itself. If the user click that cell, the cell will be bring to the top of the list on the right side. And from there, user can see more detailed information about that cell. So how we managed to map each cell into the CCF. So this is a diagram of this pipeline. So for each brain, we generated three type of the image. So first, for each cell around its location, we have about 12 coronal sections to form a partial brain, which bring us the anatomical context of that cell. And then we got the morphology information from the 60X3 image. And then we also use the 20X image as a bridge. Oh, I think here we have some problem. We cannot see the 20X image here. Okay. So 20X image is something between the BI5 image and the 60X3 image. It's just to show the half of the brain. So we use that as a bridge to link between these two kind of image. And we first connect these two images using the stage coordinates. And then the information loop is completed by two image registration. So we finally bring the cell and their investigation into the CCF. So with this map, we can register each of the cell into the CCF like this image shows 76 cells mapped to CCF. And as we know that the laminar information or the location of the cell plays a very important role in understanding the different types of cells. So now with this transform, we have a way to quantitatively compare them and have an intuitively way to seeing them. Okay. At last, I would like to thank all the contributors to this project because this is really a big science approach. Actually this all kinds of the image goes everybody in this image. Yeah. Thank you. Thank you very much. Exciting news on the addressing from here. So questions. Yes, please. So at some point you mentioned that the boundaries needs to be distinct but based on the gene expression patterns. So what do you think it would reveal something new because there could be a similar regional expression, regional regions that would have different expression patterns, right? I think so, you know, the gene expression and connectivity data is just used as a supplementary information to decide the boundary. So I think that part is doing by a group of the new anatomists but I think their idea is trying to merge all this kind of information to make the final decision. So it's correct, probably they can have different expressions but I think they will use their experience to make the final decisions. That's my understanding but if you want more details about this information, I think you're more than welcome to send just an email to our institutes. Yeah, somebody probably will help you. Thanks a lot. My question goes into a little bit the same direction. So different modalities may result in different segregational principles. So do you have developed tools to compare different parcelation schemes and at the end how do you define what a cortical areas is? I mean this cannot be simply a majority of the decision at the end but what are the underlying principles? So are you going into this direction or do you simply provide hydroarchitectonic parcelations, gene expression patterns, X, Y, Z and so on or do you really try also to come up with a new parcelation of the mouse? Even in the mouse we have different competing parcelation schemes with respect to different groups in the world. So how do you handle it? So, sorry, can you clarify your question? So your question is about using this CCF as a way to integrate data then we can compare them or your question is more about how we do the annotation about these average templates. It includes both. So the first question is, do you have developed also tools that allow you to superimpose connectivity-based parcelations with those obtained in gene expression, with those obtained in nissus staining? So this is the first. We have a way to visualize different parcelation schemes in your template space and if you would be able, how do you decide then how to label? Yeah, I think currently within our institute we already have some tools to allow the neural anatomists to overlay this kind of data otherwise they cannot do their annotation work. For the other part, in the future whether we can use this data set to map all the, use this CCF to map all the other modelled data into it and have a way, have overlapping to provide it to more users probably. I don't know. Maybe another way of asking the same question or could be relevant is the sharing of the template that you are preparing because the post-generation template was fully shared which was very helpful. Will this happen again with this new probabilistic version that you have shown us? Will it be shared fully with the community? Yes, currently in our latest release, data release in the, that's the May release, this year, this average template has been released. So if you want to use that you can just go to our website, that very section called API, then you can download this. You can download this, yes. Because then you can start dealing with the questions that Katrina is asking which probably will have happened many places with different scientists and then we need some other infrastructures to solve the problem perhaps. Yeah, there are several questions over there. Does Alan Brain have a plan to handle the breakdown of homology? So my question is around at some stage, coordinate systems just won't work as we go smaller and smaller because an organisation of neurons won't match function. So at the MRI scale, we assume structure organisation equals function and we can't assume that once we get to low histology level. So what do you think you're gonna do with coordinate systems then? Moving to the lower level? Yes, so I mean we can keep making these atlases finer and finer and finer and you showed overlay of neurons from different animals but their location with respect to them being next to each other doesn't mean much at that scale but at a larger scale it does mean something. So how do you think coordinate systems will handle that? I think this is a really good question because just as a map, sometimes we really need a detailed map in some certain areas. So for example, because our research is going to focus on the visual system. So in our plan, we are going to have a more detailed map that means a more detailed annotation of up to the higher visual area, in particular. Yeah, that's in our plan. So the final question if there was one more and then we'll have to move on. Yeah, sorry, just a basic question. So when you are building up your template, do you need to make it just on one, for example, one gender and one age group or are these characteristics not that relevant to have different brain templates? Because otherwise, don't you have the risk of making up a template that has got some average characteristics that don't fit any subject? I mean, just to make a very simple example, if you've got people that are 160 centimeters high and 108, then you've got an average of 170 but no one is, I mean, if you've got male and females, so no one is 170, but you've got different characteristics. Yes, I think if you want a more specific artist, just as you mentioned, you want different sex, you want even different disease models, that's possible. So you need to change your population, you use to build your artist. Okay, so then, okay, thank you very much again. Okay, thanks.