scdataloader

This single cell pytorch dataloader / lighting datamodule is designed to be used with:

lamindb

and:

It allows you to:

load thousands of datasets containing millions of cells in a few seconds.
preprocess the data per dataset and download it locally (normalization, filtering, etc.)
create a more complex single cell dataset
extend it to your need

built on top of lamindb and the .mapped() function by Sergei: https://github.com/Koncopd

Portions of the mapped.py file are derived from Lamin Labs
Copyright 2024 Lamin Labs
Licensed under the Apache License, Version 2.0 (http://www.apache.org/licenses/LICENSE-2.0)
The rest of the package is licensed under MIT License, see LICENSE for details
Please see https://github.com/laminlabs/lamindb/blob/main/lamindb/core/_mapped_collection.py
for the original implementation

The package has been designed together with the scPRINT paper and model.

I needed to create this Data Loader for my PhD project. I am using it to load & preprocess thousands of datasets containing millions of cells in a few seconds. I believed that individuals employing AI for single-cell RNA sequencing and other sequencing datasets would eagerly utilize and desire such a tool, which presently does not exist.

scdataloader.drawio.png

Install it from PyPI

pip install scdataloader
# or
pip install scDataLoader[dev] # for dev dependencies

lamin init --storage ./testdb --name test --schema bionty

if you start with lamin and had to do a lamin init, you will also need to populate your ontologies. This is because scPRINT is using ontologies to define its cell types, diseases, sexes, ethnicities, etc.

you can do it manually or with our function:

from scdataloader.utils import populate_my_ontology, _adding_scbasecamp_genes

populate_my_ontology() #to populate everything (recommended) (can take 2-10mns)

populate_my_ontology( #the minimum to the tool
organisms: List[str] = ["NCBITaxon:10090", "NCBITaxon:9606"],
    sex: List[str] = ["PATO:0000384", "PATO:0000383"],
    celltypes = None,
    ethnicities = None,
    assays = None,
    tissues = None,
    diseases = None,
    dev_stages = None,
)
# if you want to load the gene names and species for the arc scbasecount species, also add this:
_adding_scbasecamp_genes()

Dev install

If you want to use the latest version of scDataLoader and work on the code yourself use git clone and pip -e instead of pip install.

git clone https://github.com/jkobject/scDataLoader.git
pip install -e scDataLoader[dev]

Usage

DataModule usage

# initialize a local lamin database
#! lamin init --storage ./cellxgene --name cellxgene --schema bionty
from scdataloader import utils, Preprocessor, DataModule


# preprocess datasets
preprocessor = Preprocessor(
    do_postp=False,
    force_preprocess=True,
)
adata = preprocessor(adata)

art = ln.Artifact(adata, description="test")
art.save()
ln.Collection(art, key="test", description="test").save()

datamodule = DataModule(
    collection_name="test",
    organisms=["NCBITaxon:9606"], #organism that we will work on
    how="most expr", # for the collator (most expr genes only will be selected)
    max_len=1000, # only the 1000 most expressed
    batch_size=64,
    num_workers=1,
    validation_split=0.1,
)

see the notebooks in docs to learn more

lightning-free usage (Dataset+Collator+DataLoader)

# initialize a local lamin database
#! lamin init --storage ./cellxgene --name cellxgene --schema bionty

from scdataloader import utils, Preprocessor, SimpleAnnDataset, Collator, DataLoader

# preprocess dataset
preprocessor = Preprocessor(
    do_postp=False,
    force_preprocess=True,
)
adata = preprocessor(adata)

# create dataset
adataset = SimpleAnnDataset(
    adata, obs_to_output=["organism_ontology_term_id"]
)
# create collator
col = Collator(
    organisms="NCBITaxon:9606",
    valid_genes=adata.var_names,
    max_len=2000, #maximum number of genes to use
    how="some" |"most expr"|"random_expr",
    # genelist = [geneA, geneB] if how=='some'
)
# create dataloader
dataloader = DataLoader(
    adataset,
    collate_fn=col,
    batch_size=64,
    num_workers=4,
    shuffle=False,
)

# predict
for batch in tqdm(dataloader):
    gene_pos, expression, depth = (
        batch["genes"],
        batch["x"],
        batch["depth"],
    )
    model.predict(
        gene_pos,
        expression,
        depth,
    )

Gathering a pre-training database

Here I will explain how to gather and preprocess all of cellxgene (scPRINT-1 pretraining database) with scDataLoader, and the scPRINT-2 corpus (scPRINT-2 pretraining database).

Getting all of cellxgene

Here is an example of how to download and preprocess all of cellxgene with scDataLoader as a script (a notebook version is also available in ./notebooks/update_lamin_or_cellxgene.ipynb).

# initialize a local lamin database
#! lamin init --storage ./cellxgene --name cellxgene --schema bionty

from scdataloader import utils
from scdataloader.preprocess import LaminPreprocessor, additional_postprocess, additional_preprocess

# preprocess datasets
DESCRIPTION='preprocessed by scDataLoader'

cx_dataset = ln.Collection.connect(instance="laminlabs/cellxgene").filter(name="cellxgene-census", version='2023-12-15').one()
cx_dataset, len(cx_dataset.artifacts.all())

# (OPTIONAL) if you want to do you preprocessing on a slurm cluster without internet connections,
# you can first do this:
load_dataset_local(
    cx_dataset,
    download_folder="/my_download_folder",
    name="cached-cellxgene-census",
    description="all of it topreprocess",
)

# preprocessing
do_preprocess = LaminPreprocessor(additional_postprocess=additional_postprocess, additional_preprocess=additional_preprocess, skip_validate=True, subset_hvg=0)

preprocessed_dataset = do_preprocess(cx_dataset, name=DESCRIPTION, description=DESCRIPTION, start_at=6, version="2")

After this you can use the preprocessed dataset with the DataModule below.

# create dataloaders
from scdataloader import DataModule
import tqdm

datamodule = DataModule(
    collection_name="preprocessed dataset",
    organisms=["NCBITaxon:9606"], #organism that we will work on
    how="most expr", # for the collator (most expr genes only will be selected)
    max_len=1000, # only the 1000 most expressed
    batch_size=64,
    num_workers=1,
    validation_split=0.1,
    test_split=0)

for i in tqdm.tqdm(datamodule.train_dataloader()):
    # pass #or do pass
    print(i)
    break

# with lightning:
# Trainer(model, datamodule)

You can use the command line to preprocess a large database of datasets like here for cellxgene. this allows parallelizing and easier usage.

scdataloader --instance "laminlabs/cellxgene" --name "cellxgene-census" --version "2023-12-15" --description "preprocessed for scprint" --new_name "scprint main" --start_at 10 >> scdataloader.out

Getting the rest of the scPRINT-2 corpus

by now, using the command / scripts above you should be able to get all of cellxgene (and preprocess it). laminlabs now also hosts the rest of the scPRINT-2 corpus in laminlabs/arc-virtual-cell-atlas and they can be downloaded and preprocessed the same way as cellxgene above. Be careful however that there is no metadata for these datasets.

You can have a look at my notebooks: ./notebooks/adding_tahoe.ipynb and ./notebooks/adding_scbasecount.ipynb where I create some remmaping to retrive metadata that can be used by scdataloader and lamindb from these datasets.

If you do not have access for some reason to these datasets, please contact laminlabs. But another solution, is to download them from the original sources and add them one by one in your instance and then do the same preprocessing but this time use your_account/your_instance instead of laminlabs/arc-virtual-cell-atlas.

This is actually what I did in my own instance to create the full scPRINT-2 corpus and you can see some of it in the notebooks above.

Getting even more

They also host a pertubation a

ScDataLoader

Install / Use

README