SAFAARI

Single-cell Annotation and Fusion with Adversarial Open-Set Domain Adaptation Reliable for single-cell multi-omics Data Integration

SAFAARI: Single-Cell Annotation & Fusion using Adversarial Open-Set Adaptation

SAFAARI is a Single-cell Annotation and Fusion with Adversarial Open-Set Domain Adaptation Reliable for single-cell multi-omics Data Integration. It effectively removes batch effects, adapts to new cell types, and improves cross-modality single-cell analysis. It supports both open-set and closed-set annotation.

Features

• Open-Set & Closed-Set Adaptation: Handles novel cell types in the target dataset.
• Batch Effect Removal: Uses adversarial learning to mitigate batch effects.
• Class Imbalance Handling: Uses SMOTE oversampling to balance training data.
• Novel Cell Type Detection: Identifies unknown cell types in new datasets.
• Cross-Modality/Cross-species Integration

Installation

To install SAFAARI, run the following commands:

git clone https://github.com/VafaeeLab/SAFAARI.git
cd SAFAARI
pip install -e .

Running SAFAARI

You can run SAFAARI using the command-line interface (CLI):

Supervised Integration

safaari-supervised_integration

Unsupervised Annotation & Integration in open set OR closed set Setting

closed set mode:

safaari-unsupervised --open_set False

open set mode:

safaari-unsupervised --open_set True

Downloading Datasets for Supervised Integration

If you want to obtain the Supervised Integration Results for the following datasets:

• Ovary (RNA-ATAC)
• PBMC (ADT-SCT)
• SEURAT_PBMC (RNA-ATAC)

You can download the required datasets from the following link: Dataset Download

After downloading, you need to modify main_integration.py and uncomment the relevant dataset section to run the integration process.

Custom Configuration

You can specify additional parameters:

safaari-unsupervised --open_set True --epochs 500 --batch_size 512 --cuda 0

Data Structure

SAFAARI expects input data in CSV format, where:

• Each row represents a single cell
• Each column represents a gene
• The first column must be labeled 'cell types' and contain cell-type annotations

Dataset Organization

SAFAARI utilizes a subset of the Tabula Muris cell atlas, containing seven tissues:

• Bladder, Kidney, Heart, Mammary Gland, Muscle, Bone Marrow, Spleen

Data from FACS (Fluorescence-Activated Cell Sorting) serves as the source domain, while 10x Genomics is the target domain. The structured datasets are stored in: data/FACS10X/{Tissue_name} Each dataset follows the naming convention:

• dataset_source_CmnGenes.csv (e.g., Bladder_FACS_CmnGenes.csv)
• dataset_target_CmnGenes.csv (e.g., Bladder_10X_CmnGenes.csv)

These files contain normalized gene expression values across common highlu variable genes between domains.

Output Files

Unsupervised results will be stored in data/results/{Tissue_name}/:

• source_embeddings.csv (Processed source domain features)
• target_embeddings.csv (Processed target domain features)
• FACS_to_10X_labels.csv or FACS_to_10X_labels_op.csv (Open vs. Closed set labels)

Supervised integration results will be stored in SAFAARI_supervised_Integration_Resultsdata/{Tissue_name}/:

• dataset_FACS_to_10X_embeddings_op_source.csv (supervised source embeddings)
• dataset_FACS_to_10X_embeddings_op_target.csv (supervised target embeddings)

Example Run

safaari-unsupervised --open_set True

safaari-supervised_integration

Citation

If you use SAFAARI in your research, please cite:

@article{Aminzadeh2024SAFAARI,
  author = {Fatemeh Aminzadeh, Jun Wu, Jingrui He, Morteza Saberi, Fatemeh Vafaee},
  title = {Single-Cell Data Integration and Cell Type Annotation through Contrastive Adversarial Open-set Domain Adaptation},
  journal = {bioRxiv},
  year = {2024},
  doi = {10.1101/2024.10.04.616599},
  publisher = {Cold Spring Harbor Laboratory}
}

You can also find the preprint at: https://doi.org/10.1101/2024.10.04.616599.

---

SAFAARI GitHub Repository