scAURA

scAURA: Alignment- and Uniformity-based Graph Debiased Contrastive Representation Architecture for Self-Supervised Clustering of Single-Cell Transcriptomics.

Overview

scAURA is a unified framework for single-cell RNA sequencing (scRNA-seq) clustering that integrates graph debiased contrastive learning with self-supervised clustering to robustly identify cellular heterogeneity under high dimensionality, sparsity, and technical noise.

Key Contributions of scAURA

• Adaptive k-nearest neighbor (kNN) graph construction that dynamically adjusts neighborhood sizes, enabling improved detection of rare and small cell populations
• Noise-robust cell–cell relationship modeling by combining shared nearest neighbor (SNN)–based edge weighting with a debiased graph contrastive learning objective
• Modified contrastive loss with alignment and uniformity, ensuring biologically similar cells are embedded closer together while maintaining a well-dispersed latent space
• Self-supervised clustering module that iteratively refines cluster assignments during representation learning
• Consistent state-of-the-art performance across diverse scRNA-seq datasets, including robustness to high dropout rates and extreme sparsity, with demonstrated utility for biological discovery such as identifying novel marker genes and regulatory signals

Overall, scAURA provides a noise-aware, and biologically meaningful framework for accurate single-cell clustering and downstream analysis.

Model Architecture

Install Dependencies

pip install -r requirements.txt

Running scAURA

Two implementations are provided based on dataset size.

Option 1: CPU version (small datasets)

Use this version when the dataset contains fewer than 2,500 cells.

python scAURA.py

Option 2: GPU version (large datasets)

Use this version when the dataset contains 2,500 cells or more.

python scAURA_gpu.py