jo-mengr/cellxgene_pseudo_bulk_3_5k_pairs_natural_language_annotation

Description

This dataset contains RNA sequencing data (or actually references to anndata files stored on a cloud) and text descriptions. Dataset type: pairs (suitable for relevant training or inference tasks).

The RNA sequencing data used for training was originally gathered and annotated in the CellWhisperer project. It is derived from CellxGene and GEO. Detailed information on the gathering and annotation of the data can be read in the CellWhisperer Manuscript.

The Test data is partly taken from the cellwhisperer project (bowel disease dataset) and from Luecken et. al. It was processed and converted into a Hugging Face dataset using the adata_hf_datasets Python package. The dataset can be used to train a multimodal model, aligning transcriptome and text modalities with the sentence-transformers framework. See mmcontext for examples on how to train such a model.

The anndata objects are stored on nextcloud and a sharelink is provided as part of the dataset to download them. These anndata objects contain intial embeddings generated like this: Included embeddings: ['X_hvg', 'X_pca', 'X_scvi', 'X_geneformer'], stored in AnnData .obsm fields. These initial embeddings are used as inputs for downstream model training / inference. $caption_info

Source

• Original Data: CZ CELLxGENE Discover: A single-cell data platform for scalable exploration, analysis and modeling of aggregated data CZI Single-Cell Biology, et al. bioRxiv 2023.10.30 Publication

  Bowel Disease: Parikh, Kaushal, Agne Antanaviciute, David Fawkner-Corbett, Marta Jagielowicz, Anna Aulicino, Christoffer Lagerholm, Simon Davis, et al. 2019. “Colonic Epithelial Cell Diversity in Health and Inflammatory Bowel Disease.” Nature 567 (7746): 49–55 GEO accession

  Other Test Data: Luecken, Malte D., M. Büttner, K. Chaichoompu, A. Danese, M. Interlandi, M. F. Mueller, D. C. Strobl, et al. “Benchmarking Atlas-Level Data Integration in Single-Cell Genomics.” Nature Methods 19, no. 1 (January 2022): 41–50. Publication.

• Annotated Data: Cell Whisperer: Multimodal learning of transcriptomes and text enables interactive single-cell RNA-seq data exploration with natural-language chats Moritz Schaefer, Peter Peneder, Daniel Malzl, Mihaela Peycheva, Jake Burton, Anna Hakobyan, Varun Sharma, Thomas Krausgruber, Jörg Menche, Eleni M. Tomazou, Christoph Bock Publication Annotated Data: CellWhisperer website

• Embedding Methods: scVI: Lopez, R., Regier, J., Cole, M.B. et al. Deep generative modeling for single-cell transcriptomics. Nat Methods 15, 1053–1058 (2018). https://doi.org/10.1038/s41592-018-0229-2 geneformer: Theodoris, C.V., Xiao, L., Chopra, A. et al. Transfer learning enables predictions in network biology. Nature 618, 616–624 (2023). Publication

• Further important packages anndata: Isaac Virshup, Sergei Rybakov, Fabian J. Theis, Philipp Angerer, F. Alexander Wolf. anndata: Annotated data. bioRxiv 2021.12.16.473007 Publication scnapy: Wolf, F., Angerer, P. & Theis, F. SCANPY: large-scale single-cell gene expression data analysis. Genome Biol 19, 15 (2018). Publication

Usage

To use this dataset in Python:

from datasets import load_dataset

dataset = load_dataset("jo-mengr/cellxgene_pseudo_bulk_3_5k_pairs_natural_language_annotation")

The anndata reference is a json string which contains a share_link to the remotly stored anndata object. It can be obtained like this:

import json
import anndata
import requests
import numpy as np

adata_ref = json.loads(dataset["train"]["anndata_ref"][0])
caption = dataset["train"]["caption"] #For dataset_type "pairs"
#caption = dataset["train"]["positive"] #For dataset_type "multiplet"
adata_share_link = adata_ref["file_record"]["dataset_path"]
embedding_matrix_share_link = adata_ref["file_record"]["embeddings"]["X_scvi"] #directly access a seperatly stored embedding matrix
sample_id = adata_ref["sample_id"]
save_path = "../data/adata.h5ad"
#read the whole adata
response = requests.get(adata_share_link)
if response.status_code == 200:
  # Write the content of the response to a local file
  with open(save_path, "wb") as file:
    file.write(response.content)
else:
  print("Failed to read data from share link.")
adata = anndata.read_h5ad(save_path)

save_path = "../data/embedding.npy"
# The dataset contains several pre-computed embeddings. Lets for example get the embeddings computed with "scvi":
sample_idx = adata.obs.index == sample_id
sample_embedding = adata.obsm["X_scvi"][sample_idx]
# This sample embedding is described the the caption (loaded above)
# The same embedding should be obtainable from the embedding matrix directly
response = requests.get(embedding_matrix_share_link)
if response.status_code == 200:
  # Write the content of the response to a local file
  with open(save_path, "wb") as file:
    file.write(response.content)
else:
  print("Failed to read data from share link.")
# Load the .npz file
npzfile = np.load(save_path, allow_pickle=True)
# Extract arrays from the keys
emb_matrix = npzfile["data"]       # Assuming "data" contains your embeddings
sample_ids = npzfile["sample_ids"] # Assuming "sample_ids" contains the corresponding sample IDs
sample_embedding_2 = emb_matrix[sample_ids == sample_id]
assert np.allclose(sample_embedding, sample_embedding_2)