RADAR: Reference-guided Anomalous-cell Detection, Alignment, and Resolution

RADAR is a generative adversarial framework for marker-free fine-grained discovery of anomalous cells (ACs) in multi-sample and multimodal single-cell omics data.

RADAR is designed for Cross-sample Fine-grained Anomalous-cell Discovery (CFAD), where anomalous cells are defined as cell populations or cell states that are present in affected tissues but absent from healthy reference tissues. Given a healthy reference dataset and one or more target datasets, RADAR performs three coupled tasks in a unified workflow:

1. AC detection: identify anomalous cells in target datasets relative to a healthy reference;
2. Multi-sample alignment: reduce cross-sample or cross-modal technical variation while preserving AC-associated biological heterogeneity;
3. AC resolution: resolve detected ACs into biologically coherent subtypes.

Overview

RADAR contains three collaborative phases:

Phase I: Anomalous-cell detection

RADAR first trains a reconstruction-GAN on the healthy reference dataset only. Since the model learns to reconstruct normal cellular states, target cells with larger-than-expected reconstruction deviations are identified as putative anomalous cells.

Phase II: Multi-sample alignment

RADAR excludes Phase-I-detected ACs from alignment training to reduce the risk of treating AC-associated biological variation as batch effects. For predicted normal target cells, RADAR uses an Integrated-Gradients-guided pairing module to identify biologically relevant "kin" reference cells. These kin-cell pairs are then used to train a transferring-GAN that maps target datasets into a common reference space.

Phase III: Anomalous-cell resolution

Detected ACs are aligned into the reference space and then resolved into fine-grained AC subtypes. RADAR combines post-alignment cellular embeddings with reconstruction-deviation signals, enabling more accurate resolution of biologically distinct AC populations across samples and modalities.

Dependencies

RADAR is implemented in Python and has been tested with Python 3.9.

Main dependencies include:

anndata>=0.10.7
numpy>=1.22.4
pandas>=1.5.1
scanpy>=1.10.1
scikit-learn>=1.2.0
scipy>=1.11.4
torch>=2.0.0
tqdm>=4.64.1
captum
matplotlib

Installation

RADAR is developed as a Python package. You will need to install Python, and the recommended version is Python 3.9.

You can download the package from GitHub and install it locally:

git clone https://github.com/Catchxu/RADAR.git
cd RADAR/
pip install .

Quick start

After installation, RADAR can be run through three consecutive phases: anomalous-cell detection, multi-sample alignment, and anomalous-cell resolution.

Phase I: anomalous-cell detection

Train the reference-guided reconstruction model on the healthy reference dataset and predict anomalous cells in the target dataset:

bash scripts/01_phase1.sh

This step generates anomalous-cell prediction results, including cell-level anomaly labels and reconstruction-deviation scores.

Phase II: multi-sample alignment

Use predicted normal target cells to perform reference-guided multi-sample alignment:

bash scripts/02_phase2.sh

This step identifies kin reference cells for predicted normal target cells and maps target datasets into the reference-aligned space.

Phase III: anomalous-cell resolution

Resolve detected anomalous cells into fine-grained AC subtypes:

bash scripts/03_phase3.sh

This step generates AC subtype labels and downstream results for anomalous-cell resolution.

Data availability

All experimental datasets used in this project are available from their respective original sources. Due to file size and data-usage considerations, raw datasets are not included in this repository.

10x scRNA-seq datasets

• Healthy human intestinal epithelial tissue dataset (10xG-hInt-N) is available at GEO: GSE185224.
• Human colorectal cancer tissue datasets (10xG-hCRC-T-A and 10xG-hCRC-T-B) are available at GEO: GSE178341.
• Mouse embryo dataset (10xG-mEmb) is available at GEO: GSE186069.
• Healthy human peripheral blood mononuclear cell dataset (10xG-hHPBMC), used as a cross-modality reference for AC detection, is available from 10x Genomics.
• PBMC datasets for trajectory inference (10xG-hPBMC-A, 10xG-hPBMC-B, and 10xG-hPBMC-C) are available at GEO: GSE146974.
• Systemic lupus erythematosus PBMC datasets for AC alignment (10xG-hPBMCSLE-A, 10xG-hPBMCSLE-B, and 10xG-hPBMCSLE-C) are available at GEO: GSE96583.
• Human skin tissue datasets (10xG-hSCC-N1 to 10xG-hSCC-N5 and 10xG-hSCC-P1 to 10xG-hSCC-P5) are available at GEO: GSE144240.

10x scATAC-seq datasets

• Healthy and basal cell carcinoma human peripheral blood mononuclear cell scATAC-seq datasets (10xC-hHPBMC and 10xC-hPBMCBCC) are available from 10x Genomics.
• Mouse brain scATAC-seq datasets (10xC-mBrain-0, 10xC-mBrain-1, and 10xC-mBrain-2) are available from their original sources.

Spatial transcriptomics datasets

• Slide-seqV2 mouse embryo dataset (ssq-mEmb-33) is available at GEO: GSE197353.