Interactive Benchmarking Portal: https://peterponyu.github.io/MCCVAE/

Publication status: Published in Biomedical Signal Processing and Control.

Fu, Z., Chen, C., Zhang, K. (2026). Islands and bridges: Momentum contrastive coupling unifies discrete and continuous structure in single-cell omics. Biomedical Signal Processing and Control, 122, 110376. DOI · ScienceDirect

MCC is a deep learning framework for single-cell genomics analysis. It combines a Momentum Contrastive (MoCo) module with an information-bottleneck coupling within a Variational Autoencoder (VAE) framework.

• Momentum Contrastive Learning: Uses MoCo with InfoNCE loss, momentum-updated encoders, and memory queues for contrastive representation learning.
• Information Bottleneck Coupling: Compresses and refines latent codes through a low-dimensional bottleneck layer.
• Dual Reconstruction Pathways: Two reconstruction paths - direct from primary latent representation and through the information bottleneck.
• Multiple Modalities: Supports scRNA-seq and scATAC-seq data.

1. Clone the repository:

   git clone https://github.com/PeterPonyu/MCCVAE.git
   cd MCCVAE

2. Install the required dependencies:

   pip install -r requirements.txt

Visualization training

python start_servers.py

Use the Agent class to train models and extract latent representations.

import scanpy as sc
from mccvae import Agent

# Load data
adata = sc.read_h5ad("your_data.h5ad")

# Initialize agent
agent = Agent(
    adata=adata,
    layer="counts",
    latent_dim=10,
    use_moco=True,
    irecon=1.
)

# Train
agent.fit(epochs=1000)

# Extract representations
latent_representations = agent.get_latent()       # Shape: (n_cells, latent_dim)
bottleneck_embeddings = agent.get_bottleneck()    # Shape: (n_cells, i_dim)
refined_embeddings = agent.get_refined()          # Shape: (n_cells, latent_dim)

Embeddings:

• get_latent(): Primary latent representations (z) with dimension latent_dim
• get_bottleneck() or get_iembed(): Compressed bottleneck (l_e) with dimension i_dim
• get_refined(): Refined representations (l_d) with dimension latent_dim

The framework consists of three components:

• Query Encoder: Maps input to latent representation z
• Decoder: Reconstructs data from latent representations
• Loss: Negative binomial reconstruction + KL divergence

• Momentum Encoder: Parameters updated via exponential moving average
• Memory Queue: FIFO mechanism for negative samples
• InfoNCE Loss: Contrastive loss with L2-normalized embeddings

• Information Bottleneck: z → l_e → l_d pathway
  • l_e: Compressed representation via bottleneck_encoder (dimension: i_dim)
  • l_d: Refined representation via bottleneck_decoder (dimension: latent_dim)
• Secondary Reconstruction: Reconstruction through l_d

The MCC objective combines four loss components:

L_MCC = L_NB + β·L_KLD + α·L_MoCo + γ·L_Cou

• L_NB: Negative binomial reconstruction loss (z → reconstruction)
• L_KLD: KL divergence regularization
• L_MoCo: InfoNCE contrastive loss
• L_Cou: Coupling reconstruction loss (z → l_e → l_d → reconstruction)

If you use MCCVAE or the MCC benchmark portal, please cite:

@article{fu2026islands,
  title = {Islands and bridges: Momentum contrastive coupling unifies discrete and continuous structure in single-cell omics},
  author = {Fu, Zeyu and Chen, Chunlin and Zhang, Keyang},
  journal = {Biomedical Signal Processing and Control},
  volume = {122},
  pages = {110376},
  year = {2026},
  doi = {10.1016/j.bspc.2026.110376}
}

This project is licensed under the MIT License - see the LICENSE file for details.