gastrulation_mmusculus.py

import anndata as ad
import pyarrow as pa
import pandas as pd
import datasets

# parameters per dataset
CITATION = """
Blanca Pijuan-Sala & Jonathan Griffiths (2018). 
Timecourse single-cell RNAseq of whole mouse embryos harvested between days 6.5 and 8.5 of development. 
BioStudies, E-MTAB-6967. 
Retrieved from https://www.ebi.ac.uk/biostudies/arrayexpress/studies/E-MTAB-6967"""

DESCRIPTION = """
We have captured 100,000 single cells for single-cell RNAseq from whole mouse embryos during gastrulation and organogenesis, 
spanning days 6.5 to 8.5 of development, including embryonic and extraembryonic tissues. 
Cells were sampled every six hours, providing a continuous molecular characterisation of these processes. 
Cell libraries were prepared using the 10X Genomics Chromium platform."""

URL = "https://www.ebi.ac.uk/biostudies/arrayexpress/studies/E-MTAB-6967"
RAW_COUNTS = "X"
DATA_URL = "./data/mmusculus_gastrulation.h5ad"
FEATURES_TO_INCLUDE = ["cell_type"]

class RNAExp(datasets.ArrowBasedBuilder):
    """RNA Expression Baseclass."""

    def _info(self):

        self.batch = 10000

        # create a dictionary of features where raw_counts are ints and the rest are strings
        features = {"raw_counts": datasets.features.Sequence(datasets.features.Value("uint32")),"rows": datasets.features.Sequence(datasets.features.Value("uint32")),"size":datasets.Value("uint32")}
        for feature in FEATURES_TO_INCLUDE:
            if not features.get(feature):
                features[feature] = datasets.Value("string")

        return datasets.DatasetInfo(
            description = DESCRIPTION,
            features = datasets.Features(features), 
            homepage = URL,
            citation = CITATION
        )

    def _split_generators(self, dl_manager):
        self.anndata_file = dl_manager.download_and_extract(DATA_URL)
        adata = ad.read_h5ad(self.anndata_file, backed = "r")
        demarcation = int(len(adata)*80/100)
        
        return [
            datasets.SplitGenerator(
                name = datasets.Split.TRAIN,
                gen_kwargs = {"split": "train", "adata": adata[:demarcation], "batch_size":self.batch},
            ),
            datasets.SplitGenerator(
                name = datasets.Split.TEST,
                gen_kwargs = {"split": "test", "adata": adata[demarcation:], "batch_size":self.batch},
            )
        ]

    def _generate_tables(self, adata, batch_size, split):
        idx = 0
        
        # save the gene names as the id for the raw_counts feature
        self.info.features["raw_counts"].id = f"{','.join(adata.var.index.tolist())}"

        # iterate over the data in batches 
        for batch in range(0, adata.shape[0], batch_size):

            # raw counts
            if RAW_COUNTS == "X":
                chunk = adata.X[batch:batch+batch_size].tolil().astype('uint32')
            elif RAW_COUNTS == "raw.X":
                chunk = adata.raw.X[batch:batch+batch_size].tolil().astype('uint32')
            else:
                raise("Not valid raw_counts")
            df = pd.DataFrame([chunk.data,chunk.rows]).T
            df.columns = ['raw_counts','rows']  
            df['size'] = chunk.shape[1]
         
            # other features are all mapped to a list of strings
            for feature in FEATURES_TO_INCLUDE:
                df[feature] = list(map(str, adata.obs[feature][batch:batch+batch_size].tolist()))

            pa_table = pa.Table.from_pandas(df)
            yield idx, pa_table
            idx += 1