AnnData Conversion Design Document Part 2
#1078
Assignees
Labels
design_doc
Detailed implementation plan
Comments
|
See #1079 for the notebook where you can run it locally. |
|
@alex-l-kong @camisowers @jranek @ngreenwald Take a look, run the notebook and let me know your guys' thoughts! |
|
Great, let's discuss today to get everyone's thoughts |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
AnnDataConversion Design Document Part 2Relevant Background
The purpose of this design document is to provide guidelines and generate ideas for an in-depth implementation of
AnnData. It differs to #1073, in where that was a high level overview on why. Here we will discuss how we will transition to this format.Design Overview / Code Walkthrough
Currently, we have decided to transition
AnnDataafter Notebook 4, as it's a good point to get our feet wet without rewriting a large chunk ofArk.Let's consider all the data associated with single FOV from the Example Dataset, FOV 0.
FOV 0 is$512px \times 512px$ image, consisting of $669$ individual cells.
We will walk through converting a majority of the data associated with FOV 0 to an
AnnDatatable, as it's easiest to understand the conversion process by walking through a concrete example.You will need to run the following notebooks beforehand:
Let's get the following packages installed:
AnnDataDaskRayoptional, only used here to provide a superior scheduler forDaskImport packages, set up a local
Daskcluster and get the example data.Cell Table (Post Notebook 4)
The
cell_table_size_normalized_cell_labels.csvhouses the following informationFor each cell we have the following types of metrics:
label: The unique segmentation label for each cell, (integer)markers: The intensity of each marker for each cell, (float), specifically the following markers:CD68, CD8, CK17, Collagen1, ECAD, Fibronectin, GLUT1,
H3K27me3, H3K9ac, HLADR, IDO, Ki67, PD1, SMA, Vim
region properties: This consists of physical metadata or measurements of each cell, (int,float), specifically the following:cell_size,area,eccentricity,major_axis_length,minor_axis_length,perimeter,convex_area,equivalent_diameter,centroid-0,centroid-1,major_minor_axis_ratio,perim_square_over_area,major_axis_equiv_diam_ratio,convex_hull_resid,centroid_dif,num_concavitiesfov: A unique identifier for the FOV, (str)"fov0".cell_meta_cluster: A classification for the cell (str), in the example dataset we have:To convert these to
AnnDatacomponents we have:X: The values for themarkersobs: Housesregion properties,fov,cell_meta_cluster,labelsvar_names: The name of themarkers(channel names)obs_names: The celllabelIDsSince we also have transformations ($\tt{arcsinh}$ ) of the cell-by-markers matrix, we should add those in as an additional
layerforX.We will use
DaskDataFramesto build up a computation graph and only execute it when needed.Subset the cell tables to only extract the channels for
var.Subset the cell tables to only extract the properties for
obs.Make sure to use Categorical
Dtypesas they provide very efficient indexing / sub-setting for well, categorical properties.Here we create the$\tt arcsinh$ transformation of
AnnDataobject withX,obsand theX. You might see the following warningImplicitModificationWarning, where it converts the current index (usually an integer) to a string.The
layersparameter is a mapping from a string key to a matrix of the same shape asX, representing the same data just transformed by an operation, linear or otherwise. You can have multiple layers as well.Xgets saved as a NumPyNDArray.We can view
Xas aDataFrameby using theto_df()method.Since we created that layer in the initialization of
fov0_adata, we can also view it'sDataFramerepresentation ofX.Additional Components for
SquidPySupportIf we want to make use of
SquidPy's Spatial Analysis tools, we need to add the following components to ourAnnDataobject:spatial_key: Several functions inSquidPymake use of this key to identify the spatial coordinates of each cell (obsmand is a 2D matrix of shapegr.spatial_neighborsWe will also rename the
Centroid 0andCentroid 1columns to something more meaningful to demonstrate modifying anAnnDataobject.Spatial Analysis Components
Here we will discuss how to convert many (not all) of the spatial analysis results to
AnnDatacomponents.Distance Matrices
The distance matrices get saved as a
netcdf3file, which can be loaded in withxarray.fov0_distances_xrThe shape of the distance matrix for FOV 0 is$669 \times 669$ , which we can generalize to of $\times$
n_obsn_obs.We can extract the underlying NumPy array from the
xarray.DataArraywith the.dataattribute.A quick sidenote on
.data, it preserves the underlying array type, be it a Dask Array, in memory NumPy array, SciPy sparse array, or a CuPy GPU Array.obspis a container to store pairwise annotations of observations, so, the pairwise distances should be stored here. We'll use the key"distance"for now to identify the distance matrix.We can view the current
AnnDataobject with the newly added"distance"attribute forobsp.fov0_adataThe neighborhood counts and frequencies are stored in the following files:
neighborhood_counts-cell_meta_cluster_radius50.csvneighborhood_freqs-cell_meta_cluster_radius50.csvWe subset on those associated with FOV 0 and extract only the unique
cell_meta_clusterlabels. This gives us the following:The shapes would be$\times$
n_cell_meta_clustern_cell_meta_clusterwhich is square, but doesn't fit in with any of the otherAnnDataspecs for storing pairwise information. These would be stored inuns.fov0_adataCell Neighbors Analysis
The neighborhood diversity would be ideal for
obssince there is a value per cell for each FOV. This would be an additional column. We can perform an outer merge withobsDataFrame and theneighborhood_diversity_cell_meta_cluster_radius50.csvDataFrame per FOV, on thelabelcolumn.Let's just select the diversity for FOV 0.
neighborhood_diversity_fov0unsis a catch-all container (dictionary) for unstructured annotations of the object. It's a good place to store parameters, metadata, and other information that doesn't fit into the other components ofAnnData. In this case, we'll store the radius value for the neighborhood analysis.The cell distances in$k$ the closest cells. The shape of this is $\times$
cell_meta_cluster_avg_dists-nearest_5.csvcontains the average distance of then_obsn_cell_meta_cluster. This shape fits withobsm, so we can store it there or inuns.While in genomic analyses, users generally place embeddings of their
Xmatrix inobsm, we can also place the cell distances here as well, anything goes as long as the shapes fit, we just have to make sure to document it.I have placed the
cell_meta_cluster_avg_dists-nearest_5inobsm["cell_distances].Mixing Scores
The mixing scores are stored in the following files:
{popultation 1}_{population 2}-{mixing type}_mixing_score.csv, and with the example dataset they areCD4_CD8-homogeneous_mixing_score.csv,Just like the others, we will subset on just FOV 0.
This file has the columns:
fov,mixing_score,cell_count,cell_ratio, and a row per FOV.There are a few ways to add this to the
AnnDataobject, the first is to add themixing_score,cell_count, thecell_ratio,population 1andpopulation 2into theunssuch as shown below:cd4_cd8_homogeneous_mixing_score_fov0Another approach would be to transform the mixing scores into Matrices, one for each FOV,
For example, here is the mixing score matrix for FOV 0, where (CD4T, CD8T) has been filled.
And fill out all the values for each FOV. This would be stored in
unsas well. We would also store matrices like these forcell_countandcell_ratiofields.For now, I've stored it in
unsusing the first, nested dictionary approach.Fiber Segmentation
The fiber segmentation generates two files of interest:
fiber_object_table.csvfov,label,centroid-0,centroid-1,major_axis_length,minor_axis_length,orientation,area,eccentricity,euler_number,alignment_scorefiber_stats_table.csvfov,pixel_density,fiber_density,avg_major_axis_length,avg_minor_axis_length,avg_orientation,avg_area,avg_eccentricity,avg_euler_number,avg_alignment_scoreThe fiber segmentation notebook does not make use of the cell level segmentations, and instead creates computes the fiber level segmentations from the raw image.
This one I'm not too sure about, as it doesn't look like there is a
Xequivalent here, so creating a uniqueAnnDataobject may not be ideal. If we could,MuDatacould be useful as it's a way to represent multi-modalAnnDatatables.Perhaps place it in
uns? Let's place inunsfor this example.Neighborhood Analysis
The neighborhood analysis notebook generates three files of interest:
cell_table_size_normalized_cell_labels_kmeans_nh.csvkmeansneighborhood columnneighborhood_marker.csvkmeans_neighborhood,*var_names(all the markers)neighborhood_cell_type.csvkmeans_neighborhood,*unique_clusters(all the pixie clusters)The
neighborhood_marker.csvcould be stored invarm. Here we've transposed it to fit theAnnDataspec.Similar to
Xwe can viewvarmas aDataFrameby using theto_df()method.The
neighborhood_cell_type.csvcould be stored inunsdue to it's shape.fov0_adataFinal
AnnDataTableOur final
AnnDatatable for FOV 0 now looks like this:Some other things to go over would be the Spatial LDA notebook, and I'm sure I've missed a few other metrics as well. But I think this gets an idea of the general structure and how we can convert our data to
AnnDatacomponents. We can discuss the specifics of which parameters get placed where of course.One
AnnDataTable per Cohort or OneAnnDataTable per FOV?One of the main design decisions we have to make is whether to store all the data in one
AnnDataobject, or to split it up into manyAnnDataobjects.I am in favor of a single
AnnDataobject per FOV. This is mainly because the majority ofScanPyandSquidPyfunctions tend to work best on an individual "image" by "image" basis.If we have a single$i$ and cell $j$ for all cells in the FOV. If we have a single
AnnDataobject per FOV, the square shaped components might not make sense. For example, the takedistancecomponent ofobspwhich is a 2D array consisting of the distances between cellAnnDataobject, this matrix would take up a much larger footprint, and would require some rather obtuse indexing to get the distances for a single FOV.This will quickly take up$\mathcal{O} (n^2)$ space which is not feasible. The majority of it will be empty though and can be represented as a Block Diagonal Matrix.
The SPAIN cohort has about 400 FOVs. While taking advantage of sparsity and$\mathcal{O} (n\log{n})$ . This is completely feasible, but it can make working with the data more difficult.
Daskcan make the overhead of loading the data into memory more manageable. This kind of setup would (generously) upper bound the storage toIn addition,
SpatialDatais converting to a manyAnnDatatable approach, where a table can be associated to 0,1 or many images / coordinate systems. See scverse/spatialdata#298 for the discussion on this.Does this mean we need to run a
forloop every time we want to run a particular function across all FOV-levelAnnDatatables (such asscanpy.pp.filter_cells)?No, we can use
AnnCollectionto do this. We can lazily concatenate theAnnDataobjects and perform operations on them as if it was one bigAnnDataobject, internally it'll load eachAnnDataobject into memory as needed withDask(done automatically for us) and map a function to it.Internally, the majority of functions in
Arkload the whole cell table, but end up iterating by FOV anyway, so this can simplify some functions.As far as I know the benefits with one table per cohort start and end with having a single file stored to disk. I'm sure there are more, but I wasn't able to find them. Let me know if there are any others.
Storage: Saving and Loading
Let's take a quick detour to discuss storage.
For persistent storage we should save the
AnnDatatable as aZarrstore.Zarrprovides a nice interface for storingAnnDataobjects, as it's a hierarchical key-value store, which is exactly whatAnnDatais.Unfortunately, we will have to write to the
Zarrstore once we are done modifying theAnnDatatable.SpatialDataprovides a psuedo-"backed" feature to write modifications to disk immediately.Where
backedis defined as the property where in memory changes are written to disk immediately.ZarrbackedAnnDatastores are sort of slowly in the works. See scverse/anndata#219.We can read
ZarrbackedAnnDataTables usingAnnData.read_zarr().Lets' take a look at what the saved
Zarrfile saved at various depths:Depth 1 Hierarchy
. ├── layers ├── obs ├── obsm ├── obsp ├── uns ├── var ├── varm ├── varp └── XDepth 2 Hierarchy
. ├── layers │ └── arcsinh ├── obs │ ├── _index │ ├── area │ ├── cell_meta_cluster │ ├── cell_size │ ├── Centroid X │ ├── Centroid Y │ ├── centroid_dif │ ├── convex_area │ ├── convex_hull_resid │ ├── diversity_cell_meta_cluster │ ├── eccentricity │ ├── equivalent_diameter │ ├── fov │ ├── kmeans_neighborhood │ ├── label │ ├── major_axis_equiv_diam_ratio │ ├── major_axis_length │ ├── major_minor_axis_ratio │ ├── minor_axis_length │ ├── num_concavities │ ├── perim_square_over_area │ └── perimeter ├── obsm │ ├── cell_distances │ └── coordinates ├── obsp │ └── distance ├── uns │ ├── fiber_object │ ├── fiber_stats │ ├── mixing_scores │ ├── neighborhood_cell_type │ ├── neighborhood_counts │ ├── neighborhood_diversity │ └── neighborhood_freqs ├── var │ └── _index ├── varm │ └── kmeans ├── varp └── X └── 0.0Depth 3 Hierarchy
. ├── layers │ └── arcsinh │ └── 0.0 ├── obs │ ├── _index │ │ └── 0 │ ├── area │ │ └── 0 │ ├── cell_meta_cluster │ │ ├── categories │ │ └── codes │ ├── cell_size │ │ └── 0 │ ├── Centroid X │ │ └── 0 │ ├── Centroid Y │ │ └── 0 │ ├── centroid_dif │ │ └── 0 │ ├── convex_area │ │ └── 0 │ ├── convex_hull_resid │ │ └── 0 │ ├── diversity_cell_meta_cluster │ │ └── 0 │ ├── eccentricity │ │ └── 0 │ ├── equivalent_diameter │ │ └── 0 │ ├── fov │ │ ├── categories │ │ └── codes │ ├── kmeans_neighborhood │ │ └── 0 │ ├── label │ │ └── 0 │ ├── major_axis_equiv_diam_ratio │ │ └── 0 │ ├── major_axis_length │ │ └── 0 │ ├── major_minor_axis_ratio │ │ └── 0 │ ├── minor_axis_length │ │ └── 0 │ ├── num_concavities │ │ └── 0 │ ├── perim_square_over_area │ │ └── 0 │ └── perimeter │ └── 0 ├── obsm │ ├── cell_distances │ │ ├── _index │ │ ├── APC │ │ ├── Bcell │ │ ├── CD14_monocyte │ │ ├── CD4T │ │ ├── CD8T │ │ ├── endothelium │ │ ├── immune_other │ │ ├── M1_macrophage │ │ ├── M2_macrophage │ │ ├── Myofibroblast │ │ ├── other │ │ ├── stroma │ │ ├── tumor_ck17 │ │ └── tumor_ecad │ └── coordinates │ └── 0.0 ├── obsp │ └── distance │ ├── 0.0 │ ├── 0.1 │ ├── 1.0 │ └── 1.1 ├── uns │ ├── fiber_object │ │ ├── _index │ │ ├── alignment_score │ │ ├── area │ │ ├── centroid-0 │ │ ├── centroid-1 │ │ ├── eccentricity │ │ ├── euler_number │ │ ├── label │ │ ├── major_axis_length │ │ ├── minor_axis_length │ │ └── orientation │ ├── fiber_stats │ │ ├── avg_alignment_score │ │ ├── avg_area │ │ ├── avg_eccentricity │ │ ├── avg_euler_number │ │ ├── avg_major_axis_length │ │ ├── avg_minor_axis_length │ │ ├── avg_orientation │ │ ├── fiber_density │ │ └── pixel_density │ ├── mixing_scores │ │ └── 0 │ ├── neighborhood_cell_type │ │ ├── APC │ │ ├── Bcell │ │ ├── CD14_monocyte │ │ ├── CD4T │ │ ├── CD8T │ │ ├── endothelium │ │ ├── immune_other │ │ ├── kmeans_neighborhood │ │ ├── M1_macrophage │ │ ├── M2_macrophage │ │ ├── Myofibroblast │ │ ├── other │ │ ├── stroma │ │ ├── tumor_ck17 │ │ └── tumor_ecad │ ├── neighborhood_counts │ │ ├── _index │ │ ├── APC │ │ ├── Bcell │ │ ├── CD14_monocyte │ │ ├── CD4T │ │ ├── CD8T │ │ ├── endothelium │ │ ├── immune_other │ │ ├── M1_macrophage │ │ ├── M2_macrophage │ │ ├── Myofibroblast │ │ ├── other │ │ ├── stroma │ │ ├── tumor_ck17 │ │ └── tumor_ecad │ ├── neighborhood_diversity │ │ └── cell_meta_cluster │ └── neighborhood_freqs │ ├── _index │ ├── APC │ ├── Bcell │ ├── CD14_monocyte │ ├── CD4T │ ├── CD8T │ ├── endothelium │ ├── immune_other │ ├── M1_macrophage │ ├── M2_macrophage │ ├── Myofibroblast │ ├── other │ ├── stroma │ ├── tumor_ck17 │ └── tumor_ecad ├── var │ └── _index │ └── 0 ├── varm │ └── kmeans │ └── 0.0 ├── varp └── X └── 0.0Misc Notes
The actual data, matrices,
DataFrames, dictionaries and all are stored in the leaf nodes of the hierarchy as binary data. The rest of the nodes are just metadata, providing column names (if applicable) and other information, like the encoding type and encoding versioning.Not sure where we can save the
AnnDataobjects, perhaps under a directory calledtablesin the root of the cohort's dataset?It's also very much worth considering making use of
SquidPy's tool set for single cell spatial analysis. We should read through their code and learn how they interface withAnnDatawhen designing future components / redesigning current ones.The text was updated successfully, but these errors were encountered: