SATURN: Uniting Single-cell Gene Expressions with Protein Sequences for Cross-Species Integration

PyTorch implementation of SATURN, a deep learning approach that couples gene expression with protein representations learnt using large protein language models for cross-species integration. The key idea in SATURN is to map cells from all datasets to a shared space of functionally related genes that we name macrogenes. Using macrogenes, SATURN is uniquely able to detect functionally related genes co-expressed across species.

SATURN takes as an input:

• multiple scRNA-seq count datasets from different species (AnnDatas), with cell type annotations.
• protein embeddings generated by large language models (TorchDicts)

SATURN is composed of three modules:

• Macrogene initialization with Kmeans (scipy)
• Pretraining conditional autoencoder (scVI ZINB loss)
• Fine tuning cell clusters with weakly supervised metric learning

Vignettes/frog_zebrafish_embryogenesis/Train SATURN.ipynb has an example of running SATURN, scoring the results, and running differential expression on Macrogenes.

protein_embeddings/Generate Protein Embeddings.ipynb has an example of creating and formatting protein embeddings.

Setting up SATURN

Protein Embeddings

To run SATURN you will need protein embeddings. To generate your own protein embeddings, please see instructions in the protein_embeddings directory.

Alternatively, you can use one of the publicly available protein embedding datasets we have provided.

Requirements

SATURN requires installation of a number of python modules. Please install them via:

pip install -r requirements.txt
pip install torch==1.10.2+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

This install should talke under 10 minutes.

Running SATURN

To run SATURN, use the train-saturn.py file.

Vignettes/frog_zebrafish_embryogenesis/Train SATURN.ipynb has an example of running SATURN, scoring the results, and running differential expression on Macrogenes.

train-saturn.py accepts a number of required arguments:

• in_data: a csv file with the following columns

path: path to an Scanpy h5 file with scRNA-seq count values in the X field, and at least one column of cell type labels. species: name of the Adata species embedding_path: optionally, you can specify a path to gene embedding torch files per adata. Otherwise, SATURN will map protein embeddings based on the default paths in data/gene_embeddings.py

• in_label_col: the name of a column present in all input AnnDatas that contains cell type labels that should be used for metric learning.
• ref_label_col: the name of a column present in all input AnnDatas that contains an additional cell type label that can be used to aid metric learning if the additional argumetn --use_ref_labels is also passed, otherwise this column will just be included in SATURN's output AnnData. If you don't have an additional column, just set this value to the same column as in_label_col and do not pass the --use_ref_labels argument.

Optional Arguments

train-saturn.py also contains a number of optional arguments, a number of which are detailed below:

• hv_genes: Number of highly variable genes each AnnData should be subset to. Defaulted to 8000
• num_macrogenes: The number of macrogenes.
• score_adatas: should the pretraining AnnData and metric learning AnnDatas be scored? See the section on scoring AnnDatas for more info.
• centroids_init_path: This is a path to a .pkl file that contains a copy of gene to macrogene scores after HV subset and centroids initialization, but before SATURN pretraining. When you pass this argument, SATURN will look for a file at this location and use that file to initialize macrogene weights. If the file does not exist, SATURN will instead create centroids and save to this location. This can be used to skip the centroids creation step if running SATURN multiple times on the same datasets with the same genes (after HV gene selection).
• device_num which GPU to use

SATURN Output

train-saturn.py will output a number of files during and following training.

All of these files will be in the same directory, and have the same prefix, run_name, which SATURN will display at the conclusion of training.

AnnDatas:

• Final integrated AnnData: {run_name}.h5ad

This AnnData will have SATURN embeddings in the .X slot. In .obs, there will be columns for the original unmodified labels from in_label_col in the slot labels2, and the slot labels will contain those labels but with their corresponding species value preprended. The ref_labels column will contain values from the ref_label_col, and the species column will contain species values.

In the AnnData's .obsm, there will be a slot called macrogenes that will contain the macrogene values for each cell.

• Pretraining AnnData: {run_name}_pretrain.h5ad

The pretraining AnnData has the same format the final AnnData.

Final Macrogene Weights:

• Gene to macrogene final weights file: {run_name}_genes_to_macrogenes.pkl

Log Files:

There are a number of additional log files outputted:

• {run_name}_triplets.csv A csv with information about which triplets were mined during metric learning
• {run_name}_epoch_scores.csv A csv with information about scoring during metric learning
• {run_name}_celltype_id.pkl A pkl of a dictionary containing cell type to categorical codings used for interpreting the other log files

Scoring SATURN outputs

To score SATURN outputs, either after training or during training, you will need a csv file that maps cell types between each species.

This csv should have the columns:

{species 1 name}_cell_type, {species 2 name}_cell_type... {species n name}_cell_type

and each row should contain the cell types from each species that should be mapped to eachother.

If a cell type is unique, you can just leave the other species' values blank.

To score while training SATURN, add the argument --score_adatas and pass the name of this csv file to --ct_map_path.

To score SATURN outputs after training is finished, use the score_adata.py file.

score_adata.py takes the following arguments:

• adata: path to the SATURN formatted AnnData to score
• species1: The species whose embeddings will be used to train a simple cell type classifier
• species2: The species whose embeddings will be used to test a simple cell type classifier
• ct_map_path: path to the csv file mapping cell types between species
• label: the label column for cell types. If scoring a SATURN output, this should be labels2
• scores: the number of scores that should be calculated.

Citing

If you find our paper and code useful, please consider citing the preprint:

@article{saturn2023,
  title={Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN},
  author={Rosen, Yanay and Brbi{\'c}, Maria and Roohani, Yusuf and Swanson, Kyle and Ziang, Li and Leskovec, Jure},
  journal={bioRxiv},
  doi = {10.1101/2023.02.03.526939},
  year={2023},
}

Data Availability

Link	Description
http://snap.stanford.edu/saturn/data/ah_atlas_export.tar.gz	5 Species Alignment of Aqueous Humor Outflow (AH) Atlas
http://snap.stanford.edu/saturn/data/frog_zebrafish_export.tar.gz	Frog and Zebrafish Embryogenesis Alignment
http://snap.stanford.edu/saturn/data/tabula_mammal_export.tar.gz	Tabula Sapiens, Muris and Microcebus Coarse Whole Atlas Alignments and Individual Tissue alignemnts
http://snap.stanford.edu/saturn/data/protein_embeddings.tar.gz	Protein Embeddings for analyzed species