|
| 1 | +--- |
| 2 | +file_format: mystnb |
| 3 | +kernelspec: |
| 4 | + name: python |
| 5 | +jupytext: |
| 6 | + text_representation: |
| 7 | + extension: .md |
| 8 | + format_name: myst |
| 9 | + format_version: 0.13 |
| 10 | + jupytext_version: 1.16.2 |
| 11 | +--- |
| 12 | + |
| 13 | +# Usage |
| 14 | + |
| 15 | ++++ |
| 16 | + |
| 17 | +To demonstrate the usage of the `spatialleiden` package we are going to use a MERFISH data set from [Moffit _et al._ 2018](https://doi.org/10.1126/science.aau5324) that can be downloaded from [figshare](https://figshare.com/articles/dataset/MERFISH_datasets/22565170) and then loaded as {py:class}`anndata.AnnData` object. |
| 18 | + |
| 19 | +```{code-cell} ipython3 |
| 20 | +--- |
| 21 | +tags: [hide-cell] |
| 22 | +--- |
| 23 | +
|
| 24 | +from tempfile import NamedTemporaryFile |
| 25 | +from urllib.request import urlretrieve |
| 26 | +
|
| 27 | +import anndata as ad |
| 28 | +
|
| 29 | +with NamedTemporaryFile(suffix=".h5ad") as h5ad_file: |
| 30 | + urlretrieve("https://figshare.com/ndownloader/files/40038538", h5ad_file.name) |
| 31 | + adata = ad.read_h5ad(h5ad_file) |
| 32 | +``` |
| 33 | + |
| 34 | +First of all we are going to load the relevant packages that we will be working with as well as setting a random seed that we will use throughout this example to make the results reproducible. |
| 35 | + |
| 36 | +```{code-cell} ipython3 |
| 37 | +import scanpy as sc |
| 38 | +import spatialleiden as sl |
| 39 | +import squidpy as sq |
| 40 | +
|
| 41 | +seed = 42 |
| 42 | +``` |
| 43 | + |
| 44 | +The data set consists of 155 genes and ~5,500 cells including their annotation for cell type as well as domains. |
| 45 | + |
| 46 | ++++ |
| 47 | + |
| 48 | +## SpatialLeiden |
| 49 | + |
| 50 | +We will do some standard preprocessing by log-transforming the data and then using PCA for dimensionality reduction. The PCA will be used to build a kNN graph in the latent gene expression space. This graph is the basis for the Leiden clustering. |
| 51 | + |
| 52 | +```{code-cell} ipython3 |
| 53 | +sc.pp.log1p(adata) |
| 54 | +sc.pp.pca(adata, random_state=seed) |
| 55 | +
|
| 56 | +sc.pp.neighbors(adata, random_state=seed) |
| 57 | +``` |
| 58 | + |
| 59 | +For SpatialLeiden we need an additional graph representing the connectivities in the topological space. Here we will use a kNN graph with 10 neighbors that we generate with {py:func}`squidpy.gr.spatial_neighbors`. Alternatives are Delaunay triangulation or regular grids in case of e.g. Visium data. |
| 60 | + |
| 61 | +We can use the calculated distances between neighboring points and transform them into connectivities using the {py:func}`spatialleiden.distance2connectivity` function. |
| 62 | + |
| 63 | +```{code-cell} ipython3 |
| 64 | +sq.gr.spatial_neighbors(adata, coord_type="generic", n_neighs=10) |
| 65 | +
|
| 66 | +adata.obsp["spatial_connectivities"] = sl.distance2connectivity( |
| 67 | + adata.obsp["spatial_distances"] |
| 68 | +) |
| 69 | +``` |
| 70 | + |
| 71 | +Now, we can already run {py:func}`spatialleiden.spatialleiden` (which we will also compare to normal Leiden clustering). |
| 72 | + |
| 73 | +The `layer_ratio` determines the weighting between the gene expression and the topological layer and is influenced by the graph structures (i.e. how many connections exist, the edge weights, etc.); the lower the value is the closer SpatialLeiden will be to normal Leiden clustering, while higher values lead to more spatially homogeneous clusters. |
| 74 | + |
| 75 | +The resolution has the same effect as in Leiden clustering (higher resolution will lead to more clusters) and can be defined for each of the layers (but for now is left at its default value). |
| 76 | + |
| 77 | +```{code-cell} ipython3 |
| 78 | +sc.tl.leiden(adata, directed=False, random_state=seed) |
| 79 | +
|
| 80 | +sl.spatialleiden(adata, layer_ratio=1.8, directed=(False, True), seed=seed) |
| 81 | +
|
| 82 | +sc.pl.embedding(adata, basis="spatial", color=["leiden", "spatialleiden"]) |
| 83 | +``` |
| 84 | + |
| 85 | +We can see how Leiden clustering identifies cell types while SpatialLeiden defines domains of the tissue. |
| 86 | + |
| 87 | ++++ |
| 88 | + |
| 89 | +## Resolution search |
| 90 | + |
| 91 | +If you already know how many domains you expect in your sample you can use the {py:func}`spatialleiden.search_resolution` function to identify the resolutions needed to obtain the correct number of clusters. |
| 92 | + |
| 93 | +Conceptually, this function first runs Leiden clustering multiple times while changing the resolution to identify the value leading to the desired number of clusters. Next, this procedure is repeated by running SpatialLeiden, but now the resolution of the latent layer (gene expression) is kept fixed and the resolution of the spatial layer is varied. |
| 94 | + |
| 95 | +```{code-cell} ipython3 |
| 96 | +n_clusters = adata.obs["domain"].nunique() |
| 97 | +
|
| 98 | +latent_resolution, spatial_resolution = sl.search_resolution( |
| 99 | + adata, |
| 100 | + n_clusters, |
| 101 | + latent_kwargs={"seed": seed}, |
| 102 | + spatial_kwargs={"layer_ratio": 1.8, "seed": seed, "directed": (False, True)}, |
| 103 | +) |
| 104 | +
|
| 105 | +print(f"Latent resolution: {latent_resolution:.3f}") |
| 106 | +print(f"Spatial resolution: {spatial_resolution:.3f}") |
| 107 | +``` |
| 108 | + |
| 109 | +In our case we can compare the resulting clusters to the annotated ground truth regions. If we are not satisfied with the results, we can go back and tweak other parameters such as the underlying neighborhood graphs or the `layer_ratio` to achieve the desired granularity of our results. |
| 110 | + |
| 111 | +```{code-cell} ipython3 |
| 112 | +sc.pl.embedding(adata, basis="spatial", color=["spatialleiden", "Region"]) |
| 113 | +``` |
0 commit comments