scMIAC: Single-Cell Multi-modality Integration via cell type filtered Anchors using Contrastive learning

Overview

scMIAC is a comprehensive framework for single-cell multi-modality data integration, designed to tackle the most challenging problem in single-cell integration: diagonal integration (integrating unpaired cells from different feature spaces across modalities).

The methodological innovations of scMIAC include:

• scMIAC utilizes cell type information to select high-quality anchor cells for contrastive learning, improving integration of challenging cells such as imbalanced, rare, or isolated cell types.
• scMIAC innovatively introduces contrastive learning to diagonal integration task, where previous methods could only be applied to horizontal or vertical integration scenarios.
• As a diagonal integration approach, scMIAC preserves each modality's original biological characteristics through modality-specific VAEs, which serves as a regularizer preventing over-emphasis on modality alignment.

---

Installation

conda create -n scmiac python=3.11
conda activate scmiac
pip install scmiac

---

Usage

scMIAC provides two usage modes:

1. CLI Mode - Quick start for non-interactive, command-line workflow:

scmiac train \
  --rna-h5ad data/10x/input/adata_rna_10x.h5ad \
  --atac-h5ad data/10x/input/adata_atac_10x.h5ad \
  --output-dir data/10x/output/scmiac_results/ \
  --rna-latent-key X_pca \
  --atac-latent-key lsi49 \
  --rna-celltype-key cell_type \
  --atac-celltype-key pred

scmiac train -h  # For viewing all available parameters

MUST Required parameters:

• --rna-h5ad: Path to RNA AnnData file
• --atac-h5ad: Path to ATAC AnnData file
• --output-dir: Output directory

Output files:

• anchors.csv: Anchor pairs
• rna_vae.pth: RNA VAE model weights
• atac_vae.pth: ATAC VAE model weights
• rna_embeddings.csv: RNA cell embeddings
• atac_embeddings.csv: ATAC cell embeddings
• scmiac_latent_umap.png: UMAP visualization

2. API Mode - For flexible research and experimentation:

Refer to the Full API Tutorial & Examples for detailed usage and examples.

---

Reproducing Paper Results

The scripts/ directory contains all experimental code for reproducing the results in the scMIAC paper, including:

• Competing method comparisons
• Ablation studies
• Hyperparameter experiments
• Evaluation framework

For detailed information, see scripts/README.md.