Understanding how data is organized in Scarf¶
In this notebook, we provide a more detailed exploration of how the data is organized in Scarf. This can be useful for users who want to customize certain aspects of Scarf or want to extend its functionality.
[1]:
%load_ext autotime
%config InlineBackend.figure_format = 'retina'
import scarf
scarf.__version__
[1]:
'0.8.5'
time: 963 ms (started: 2021-08-22 18:49:03 +00:00)
We download the CITE-Seq dataset that we analyzed earlier
[2]:
scarf.fetch_dataset('tenx_8K_pbmc_citeseq', save_path='scarf_datasets', as_zarr=True)
INFO: Download started...
INFO: Download finished! File saved here: /home/docs/checkouts/readthedocs.org/user_builds/scarf/checkouts/0.8.5/docs/source/vignettes/scarf_datasets/tenx_8K_pbmc_citeseq/data.zarr.tar.gz
INFO: Extracting Zarr file for tenx_8K_pbmc_citeseq
2021-08-22 18:49:15 URL:https://storage.googleapis.com/cos-osf-prod-files-de-1/9d8e4be82c0b76dd8683a1f6659011c65277aa2d464853cfccd49e27835c4f88?response-content-disposition=attachment%3B%20filename%3D%22data.zarr.tar.gz%22%3B%20filename%2A%3DUTF-8%27%27data.zarr.tar.gz&GoogleAccessId=files-de-1%40cos-osf-prod.iam.gserviceaccount.com&Expires=1629658212&Signature=LO9XgLb908qHLUsHqoOltHPumvR1zEBjFkR1P57JZZrg9J3KjF8xu1ebxuWsh1aaV6lqiijeDgxKo2u1YnLx584GglvXispxX8R533bYrT%2FI2a44wJmV3iVMo0ZGEhPxbYKGAAZ3LtRab%2FRNIzC4UqIwzpOPwewCA%2BSVVnKA4Cy4uRiEu38L3D%2BleKXH5cslZwdC69xhqYUPQtYj%2BkdW%2FJana0tSUU2Nega2EX7Nh3eDDum%2F8LWjUBmKQo0%2F2RCcQyM8eu%2FyDmE%2FOH9fKZpjVJsnbOd%2BCs8gDY6FXVtSFC9vOIm5S3MD%2FT%2FEvXZjChtd7%2Bg4oCbTE%2Br8cIkikYByMA%3D%3D [68237599/68237599] -> "/home/docs/checkouts/readthedocs.org/user_builds/scarf/checkouts/0.8.5/docs/source/vignettes/scarf_datasets/tenx_8K_pbmc_citeseq/data.zarr.tar.gz" [1]
time: 12.4 s (started: 2021-08-22 18:49:04 +00:00)
[3]:
ds = scarf.DataStore('scarf_datasets/tenx_8K_pbmc_citeseq/data.zarr', nthreads=4)
time: 19.3 ms (started: 2021-08-22 18:49:17 +00:00)
1) Zarr trees¶
Scarf uses Zarr format to store raw counts as dense, chunked matrices. The Zarr file is in fact a directory that organizes the data in form of a tree. The count matrices, cell and features attributes, and all the data is stored in this Zarr hierarchy. Some of the key benefits of using Zarr over other formats like HDF5 are: - Parallel read and write access - Availability of compression algorithms like LZ4 that provides very high compression and decompression speeds. - Automatic storage of intermediate and processed data
In Scarf, the data is always synchronized between hard disk and RAM and as such there is no need to save data manually or synchronize the
We can inspect how the data is organized within the Zarr file using show_zarr_tree
method of the DataStore
.
[4]:
ds.show_zarr_tree(depth=1)
/
├── ADT
├── RNA
└── cellData
time: 2.18 ms (started: 2021-08-22 18:49:17 +00:00)
By setting depth=1
we get a look of the top-levels in the Zarr hierarchy: ‘RNA’, ‘ADT’ and ‘cellData’. The top levels here are hence composed of two assays and the cellData
level, which will be explained below. Scarf attempts to store most of the data on disk immediately after it is processed. Since, that data that we have loaded is already preprocessed, below we can see that the calculated cell attributes can now be found under the cellData
level. The cell-wise statistics that were
calculated using ‘RNA’ assay have ‘RNA’ preppended to the name of the column. Similarly the columns stating with ‘ADT’ were created using ‘ADT’ assay data.
[5]:
ds.show_zarr_tree(start='cellData')
cellData
├── ADT_UMAP1 (7865,) float32
├── ADT_UMAP2 (7865,) float32
├── ADT_leiden_cluster (7865,) int64
├── ADT_nCounts (7865,) float64
├── ADT_nFeatures (7865,) float64
├── I (7865,) bool
├── RNA_UMAP1 (7865,) float32
├── RNA_UMAP2 (7865,) float32
├── RNA_cluster (7865,) int64
├── RNA_leiden_cluster (7865,) int64
├── RNA_nCounts (7865,) float64
├── RNA_nFeatures (7865,) float64
├── RNA_percentMito (7865,) float64
├── RNA_percentRibo (7865,) float64
├── RNA_tSNE1 (7865,) float64
├── RNA_tSNE2 (7865,) float64
├── ids (7865,) <U18
└── names (7865,) <U18
time: 4.1 ms (started: 2021-08-22 18:49:17 +00:00)
The ``I`` column : This column is used to keep track of valid and invalid cells. The values in this column are boolean indicating which cells where filtered out (hence False) during filtering process. One of can think of column I
as the default subset of cells that will be used for analysis if any other subset is not specifially chosen explicitly by the users. Most methods of DataStore
object accept a parameter called cell_key
and the default value for this paramter is I
,
which means that only cells with True in I
column will be used.
If we inspect one of the assay, we will see the following zarr groups. The raw data matrix is stored under counts
group (more on it in the next section). featureData
is the feature level equivalent of cellData
. featureData
is explained in further detail in the next section. The markers
level will be explained further in the last section. summary_stats_I
contains statistics about features. These statistics are generally created during feature selection step.
‘normed__I__hvgs’ is explained in detail in the fourth section of this documentation.
[6]:
ds.show_zarr_tree(start='RNA', depth=1)
RNA
├── counts (7865, 33538) uint32
├── featureData
├── markers
├── normed__I__hvgs
└── summary_stats_I
time: 2.63 ms (started: 2021-08-22 18:49:17 +00:00)
The feature wise attributes for each assay are stored under the assays featureData
level. The output below shows some of the feature level statistics.
[7]:
ds.show_zarr_tree(start='RNA/featureData', depth=1)
featureData
├── I (33538,) bool
├── I__hvgs (33538,) bool
├── dropOuts (33538,) int64
├── ids (33538,) <U15
├── nCells (33538,) int64
└── names (33538,) <U16
time: 2.37 ms (started: 2021-08-22 18:49:17 +00:00)
2) Cell and feature attributes¶
The cell and feature level attributes can be accessed through DataStore
, for example, using ds.cells
and ds.RNA.feats
respectively (both are objects of Metadata
class). In this section we dive deeper into these attribute tables and try to perform CRUD operations on them.
head
provides a quick look at the attribute tables.
[8]:
ds.cells.head()
[8]:
I | ids | names | ADT_UMAP1 | ADT_UMAP2 | ADT_leiden_cluster | ADT_nCounts | ADT_nFeatures | RNA_UMAP1 | RNA_UMAP2 | RNA_cluster | RNA_leiden_cluster | RNA_nCounts | RNA_nFeatures | RNA_percentMito | RNA_percentRibo | RNA_tSNE1 | RNA_tSNE2 | |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
0 | True | AAACCCAAGATTGTGA-1 | AAACCCAAGATTGTGA-1 | -13.884989 | 9.510007 | 16 | 981.0 | 17.0 | 29.421942 | -7.417822 | 4 | 3 | 6160.0 | 2194.0 | 8.668831 | 15.259740 | -4.26865 | -18.08830 |
1 | True | AAACCCACATCGGTTA-1 | AAACCCACATCGGTTA-1 | -5.097915 | 10.784672 | 15 | 1475.0 | 17.0 | 30.801842 | -12.649998 | 4 | 3 | 6713.0 | 2093.0 | 6.316103 | 19.037688 | 5.02671 | -14.88920 |
2 | True | AAACCCAGTACCGCGT-1 | AAACCCAGTACCGCGT-1 | -7.885502 | 12.961405 | 9 | 7149.0 | 17.0 | 29.397541 | 4.256729 | 5 | 5 | 3637.0 | 1518.0 | 8.056090 | 16.002200 | -14.91320 | -18.83030 |
3 | True | AAACCCAGTATCGAAA-1 | AAACCCAGTATCGAAA-1 | 12.072053 | 11.531997 | 2 | 6831.0 | 17.0 | 5.441914 | -32.227364 | 3 | 4 | 1244.0 | 737.0 | 9.003215 | 18.729904 | 16.69630 | -3.35003 |
4 | True | AAACCCAGTCGTCATA-1 | AAACCCAGTCGTCATA-1 | 13.928782 | 12.290624 | 2 | 6839.0 | 17.0 | 7.166490 | -30.397615 | 3 | 4 | 2611.0 | 1240.0 | 6.204519 | 16.353887 | 18.37000 | -1.37607 |
time: 41.4 ms (started: 2021-08-22 18:49:17 +00:00)
The feature attribute table from any assay can similarly be quickly inspected
[9]:
ds.RNA.feats.head()
[9]:
I | ids | names | I__hvgs | dropOuts | nCells | |
---|---|---|---|---|---|---|
0 | False | ENSG00000243485 | MIR1302-2HG | False | 7865 | 0 |
1 | False | ENSG00000237613 | FAM138A | False | 7865 | 0 |
2 | False | ENSG00000186092 | OR4F5 | False | 7865 | 0 |
3 | False | ENSG00000238009 | AL627309.1 | False | 7853 | 12 |
4 | False | ENSG00000239945 | AL627309.3 | False | 7865 | 0 |
time: 16.2 ms (started: 2021-08-22 18:49:17 +00:00)
Even though the above ‘head’ command may make you think that ds.cells
and ds.RNA.feats
are Pandas dataframe, they in fact are not. If you wish to obtain a full table as Pandas dataframe, then you can export the columns of your choice as shown below.
[10]:
ds.cells.to_pandas_dataframe(['ids', 'RNA_UMAP1', 'RNA_UMAP2', 'RNA_cluster']).set_index('ids')
[10]:
RNA_UMAP1 | RNA_UMAP2 | RNA_cluster | |
---|---|---|---|
ids | |||
AAACCCAAGATTGTGA-1 | 29.421942 | -7.417822 | 4 |
AAACCCACATCGGTTA-1 | 30.801842 | -12.649998 | 4 |
AAACCCAGTACCGCGT-1 | 29.397541 | 4.256729 | 5 |
AAACCCAGTATCGAAA-1 | 5.441914 | -32.227364 | 3 |
AAACCCAGTCGTCATA-1 | 7.166490 | -30.397615 | 3 |
... | ... | ... | ... |
TTTGTTGGTTCAAGTC-1 | -5.545609 | 21.742104 | 18 |
TTTGTTGGTTGCATGT-1 | -26.972895 | -0.762715 | 9 |
TTTGTTGGTTGCGGCT-1 | 27.239552 | -5.230516 | 4 |
TTTGTTGTCGAGTGAG-1 | -23.030291 | -20.647554 | 2 |
TTTGTTGTCGTTCAGA-1 | -10.845096 | -2.644977 | 2 |
7865 rows × 3 columns
time: 16.5 ms (started: 2021-08-22 18:49:17 +00:00)
If you wish to export all the columns as a dataframe then simply provide ds.cells.columns
or ds.RNA.feats.columns
as an argument rather than a list of columns. If you are interested in one of the columns then it can fetched using either fetch
or fetch_all
command.
fetch
will provide values for a subset of cells (by default, only for those that have True value in column I
, but any other boolean column can be given).
[11]:
clusters = ds.cells.fetch('RNA_cluster')
clusters.shape
[11]:
(7549,)
time: 4.85 ms (started: 2021-08-22 18:49:17 +00:00)
fetch_all
will return values for all the cells.
[12]:
clusters_all = ds.cells.fetch_all('RNA_cluster')
clusters_all.shape
[12]:
(7865,)
time: 3.2 ms (started: 2021-08-22 18:49:17 +00:00)
If you wish to add a new column then insert
command can be used for either cell or feature attributes. insert
method will take care of inserting the values in the right rows even when a subset of values are provided. The default value for key
parameter is I
in insert
method, so it will add values in same order as cells that have True value for I
[13]:
is_clust_1 = clusters == 1 # is_clust_1 has just 7549 elements (same as the number of cells with value True in `I`)
ds.cells.insert(column_name='is_clust1', values=is_clust_1)
time: 3.84 ms (started: 2021-08-22 18:49:17 +00:00)
If we try to add values for a column that already exists then Scarf will throw an error. For example, if we simply rerun the command above, we should get an error
[14]:
ds.cells.insert(column_name='is_clust1', values=is_clust_1)
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
/tmp/ipykernel_6260/1353646084.py in <module>
----> 1 ds.cells.insert(column_name='is_clust1', values=is_clust_1)
~/checkouts/readthedocs.org/user_builds/scarf/envs/0.8.5/lib/python3.8/site-packages/scarf/metadata.py in insert(self, column_name, values, fill_value, key, overwrite, location, force)
364 raise ValueError(f"ERROR: {col} is a protected column name in MetaData class.")
365 if col in self.columns and overwrite is False:
--> 366 raise ValueError(f"ERROR: {col} already exists. Please set `overwrite` to True to overwrite.")
367 if type(values) == list:
368 logger.warning("'values' parameter is of `list` type and not `np.ndarray` as expected. The correct dtype "
ValueError: ERROR: is_clust1 already exists. Please set `overwrite` to True to overwrite.
time: 185 ms (started: 2021-08-22 18:49:17 +00:00)
To override this behaviour, you can use ‘overwrite’ parameter and set it to True
[15]:
ds.cells.insert(column_name='is_clust1', values=is_clust_1, overwrite=True)
time: 4.15 ms (started: 2021-08-22 18:49:17 +00:00)
Please checkout the API of the Metadata
class to get information on how to perform other operations like delete and update columns.
3) Count matrices and data normalization¶
Scarf uses Zarr format so that data can be stored in rectangular chunks. The raw data is saved in the counts
level within each assay level in the Zarr hierarchy. It can easily be accessed as a Dask array using the rawData
attribute of the assay. Note that for a standard analysis one would not interact with the raw data directly. Scarf internally optimizes the use of this Dask array to minimize the memory requirement of all operations.
[16]:
ds.RNA.rawData
[16]:
|
time: 3.37 ms (started: 2021-08-22 18:49:17 +00:00)
The normalized data can be accessed through the normed
method of the assay. In Scarf a user doesn’t need to perform normalization step manually, Scarf only stores the raw data and generates normalized data whenever needed. This means that Scarf may need to perform normalization several times. However, in practise we have noted that time spent in normalizing the data is only a small fraction of routine workflows.
[17]:
ds.RNA.normed()
[17]:
|
time: 12.4 ms (started: 2021-08-22 18:49:17 +00:00)
Users can override how the normalization is performed in Scarf. Normalization is performed using the normMethod
attribute which references the function responsible for performing the normalization.
[18]:
ds.RNA.normMethod
[18]:
<function scarf.assay.norm_lib_size(assay, counts: <module 'dask.array' from '/home/docs/checkouts/readthedocs.org/user_builds/scarf/envs/0.8.5/lib/python3.8/site-packages/dask/array/__init__.py'>) -> <module 'dask.array' from '/home/docs/checkouts/readthedocs.org/user_builds/scarf/envs/0.8.5/lib/python3.8/site-packages/dask/array/__init__.py'>>
time: 1.65 ms (started: 2021-08-22 18:49:17 +00:00)
Let’s checkout the source of function that is referenced by ds.RNA.normMethod
[19]:
import inspect
print (inspect.getsource(ds.RNA.normMethod))
def norm_lib_size(assay, counts: daskarr) -> daskarr:
return assay.sf * counts / assay.scalar.reshape(-1, 1)
time: 2.28 ms (started: 2021-08-22 18:49:17 +00:00)
Following is an example of how one can override the method of normalization
[20]:
def my_cool_normalization_method(assay, counts):
import numpy as np
lib_size = counts.sum(axis=1).reshape(-1, 1) # Calculate total counts for each cell
return np.log2(counts/lib_size) # Library size normalization followed by log2 transformation
ds.RNA.normMethod = my_cool_normalization_method
time: 499 µs (started: 2021-08-22 18:49:17 +00:00)
Now whenever Scarf internally requires normalized values, this function will be used. Scarf provides a dummy normalization function (scarf.assay.norm_dummy
) that does not perform normalization. This function can be useful if you have pre-normalized data and need to disable default normalization.
[21]:
ds.RNA.normMethod = scarf.assay.norm_dummy
time: 231 µs (started: 2021-08-22 18:49:17 +00:00)
Please note: If you are using a custom function or disabling normalization, then everytime you load a DataStore object you will need to reassign normMethod
to the function of your choice.
4) Data caching during graph creation¶
All the results of make_graph
step are saved under a name on the form ‘normed__{cell key}__{feature key}’ (placeholders used in brackets here). In this case, since we did not provide a cell key it takes default value of I
, which means all the cells that were not filtered out. The feature key (feat_key
) was set to hvgs
. The Zarr directory is organized such that all the intermediate data is also saved. The intermediate data is organized in a hierarchy which triggers
recomputation when upstream changes are detected. The parameter values are also saved in hierarchy level names. For example, ‘reduction_pca_31_I’ means that PCA linear dimension reduction with 31 PC axes was used and the PCA was fit across all the cells that have True
value in column I.
[22]:
ds.show_zarr_tree(start='RNA/normed__I__hvgs')
normed__I__hvgs
├── data (7549, 1995) float64
├── reduction__pca__15__I
│ ├── ann__l2__50__50__48__4466
│ │ └── knn__11
│ │ ├── distances (7399, 11) float64
│ │ ├── graph__1.0__1.5
│ │ │ ├── edges (81389, 2) uint64
│ │ │ └── weights (81389,) float64
│ │ └── indices (7399, 11) uint64
│ ├── kmeans__100__4466
│ │ ├── cluster_centers (100, 15) float64
│ │ └── cluster_labels (7399,) float64
│ ├── mu (497,) float64
│ ├── reduction (497, 15) float64
│ └── sigma (497,) float64
└── reduction__pca__30__I
├── ann__l2__63__63__48__4466
│ └── knn__21
│ ├── distances (7549, 21) float64
│ ├── graph__1.0__1.5
│ │ ├── dendrogram (7548, 4) float64
│ │ ├── dendrogram_coalesced_20
│ │ │ ├── edgelist (38, 2) uint64
│ │ │ └── nodelist (39, 3) int64
│ │ ├── edges (158529, 2) uint64
│ │ └── weights (158529,) float64
│ └── indices (7549, 21) uint64
├── kmeans__100__4466
│ ├── cluster_centers (100, 30) float64
│ └── cluster_labels (7549,) float64
├── mu (1995,) float64
├── reduction (1995, 30) float64
└── sigma (1995,) float64
time: 8.95 ms (started: 2021-08-22 18:49:17 +00:00)
The graph calculated by make_graph
can be easily loaded using the load_graph
method, like below. The graph is loaded as a sparse matrix of the cells that were used for creating a graph.
Next, we show how the graph can be accessed if required. However, as stated above, normally Scarf handles the graph loading internally where required.
Because Scarf saves all the intermediate data, it might be the case that a lot of graphs are stored in the Zarr hierarchy. load_graph
will load only the latest graph that was computed (for the given assay, cell key and feat key).
[23]:
ds.load_graph(from_assay='RNA', cell_key='I', feat_key='hvgs', symmetric=False, upper_only=False)
[23]:
<7549x7549 sparse matrix of type '<class 'numpy.float64'>'
with 158529 stored elements in Compressed Sparse Row format>
time: 15.5 ms (started: 2021-08-22 18:49:17 +00:00)
The sparse matrix above will load the If you would like to load a graph generated using some other parameter, then simply run make_graph
again using those parameters. make_graph
will not recalculate the graph if it already exists and will simply set it as the latest graph. This graph can then be loaded using load_graph
5) Fetching marker features¶
Marker features (ex. marker genes) for a selection of cells can be calculated using run_marker_search
. The markers for each group are stored under ‘marker’ group under the assay’s group. Within the ‘{assay}/markers’ level there are sublevels (only 1 here) created based on which cell set and cell groups were used: {cell_key}__{group_key}. Stored below this Zarr level are the individual group labels that contain the maker feature ids and their corresponding scores.
[24]:
ds.show_zarr_tree(start='RNA/markers', depth=2)
markers
└── I__RNA_cluster
├── 1
├── 10
├── 11
├── 12
├── 13
├── 14
├── 15
├── 16
├── 17
├── 18
├── 19
├── 2
├── 20
├── 3
├── 4
├── 5
├── 6
├── 7
├── 8
└── 9
time: 5.31 ms (started: 2021-08-22 18:49:17 +00:00)
Marker list for any group (e.x. cell cluster) can be fetched like below
[25]:
ds.get_markers(group_key='RNA_cluster', group_id=7)
[25]:
score | names | |
---|---|---|
ids | ||
ENSG00000144290 | 0.696736 | SLC4A10 |
ENSG00000062524 | 0.52321 | LTK |
ENSG00000143365 | 0.473731 | RORC |
ENSG00000111796 | 0.381826 | KLRB1 |
ENSG00000113088 | 0.379597 | GZMK |
ENSG00000218357 | 0.362173 | LINC01644 |
ENSG00000099282 | 0.359214 | TSPAN15 |
ENSG00000197635 | 0.354979 | DPP4 |
ENSG00000080573 | 0.311975 | COL5A3 |
ENSG00000174946 | 0.297659 | GPR171 |
ENSG00000204475 | 0.287263 | NCR3 |
ENSG00000168824 | 0.282651 | NSG1 |
ENSG00000109906 | 0.281518 | ZBTB16 |
ENSG00000000971 | 0.273301 | CFH |
ENSG00000139187 | 0.269599 | KLRG1 |
ENSG00000206561 | 0.252777 | COLQ |
time: 45.5 ms (started: 2021-08-22 18:49:17 +00:00)
One can also export the names of markers for all the groups to a CSV file like below:
[26]:
ds.export_markers_to_csv(group_key='RNA_cluster', csv_filename='test.csv')
time: 1.01 s (started: 2021-08-22 18:49:17 +00:00)
That is all for this vignette.