MoCoO

Momentum Contrast ODE-Regularized VAE for Single-Cell RNA Velocity

A unified deep learning framework combining Variational Autoencoders (VAE), Neural Ordinary Differential Equations (ODE), and Momentum Contrast (MoCo) for robust single-cell trajectory inference and representation learning.

---

Features

• VAE-based dimensionality reduction with multiple count-based likelihoods (MSE, NB, ZINB, Poisson, ZIP)
• Neural ODE for continuous trajectory modeling and pseudotime inference
• Momentum Contrast (MoCo) for robust contrastive representation learning
• Information bottleneck for hierarchical feature extraction
• Disentanglement losses (DIP-VAE, β-TC-VAE, InfoVAE) for interpretable latents
• Vector field analysis for RNA velocity visualization

---

Installation

From PyPI (recommended)

pip install mocoo

From source

git clone https://github.com/PeterPonyu/MoCoO.git
cd MoCoO
pip install -e .

Development installation

git clone https://github.com/PeterPonyu/MoCoO.git
cd MoCoO
pip install -e ".[dev]"

Publishing

The package is automatically published to PyPI when a GitHub release is created.

To create a new release:

1. Bump version:

   python release.py patch  # For bug fixes (0.0.1 → 0.0.2)
   python release.py minor  # For new features (0.0.1 → 0.1.0)
   python release.py major  # For breaking changes (0.0.1 → 1.0.0)
   

2. Commit and push:

   git add -A
   git commit -m "Bump version to X.Y.Z"
   git push
   

3. Create GitHub release:

   • Go to Releases
   • Click "Create a new release"
   • Tag: vX.Y.Z (e.g., v0.1.0)
   • Title: Release X.Y.Z
   • Description: List changes
   • Click "Publish release"

4. Automated publishing:

   • GitHub Actions will automatically build and publish to PyPI
   • Check the Actions tab for build status

---

Quick Start

Basic VAE

import scanpy as sc
from mocoo import MoCoO

adata = sc.read_h5ad('data.h5ad')

model = MoCoO(
    adata,
    layer='counts',
    loss_mode='nb',
    batch_size=128
)
model.fit(epochs=100)

latent = model.get_latent()
adata.obsm['X_mocoo'] = latent

With ODE + MoCo

model = MoCoO(
    adata,
    use_ode=True,
    use_moco=True,
    latent_dim=10,
    i_dim=2,
    moco_K=4096,
    aug_prob=0.5,
    batch_size=256
)
model.fit(epochs=400, patience=25)

latent = model.get_latent()
velocity = model.get_velocity()
pseudotime = model.get_time()
transition = model.get_transition(top_k=30)

---

Parameters

Parameter	Type	Default	Description
adata	AnnData	required	Annotated data matrix
layer	str	'counts'	Layer containing raw counts
loss_mode	str	'nb'	Likelihood: 'mse', 'nb', 'zinb', 'poisson', 'zip'
latent_dim	int	10	Latent space dimension
i_dim	int	2	Bottleneck dimension (< latent_dim)
use_ode	bool	False	Enable Neural ODE
use_moco	bool	False	Enable MoCo
moco_K	int	4096	MoCo queue size
batch_size	int	128	Mini-batch size
lr	float	1e-4	Learning rate

See docstrings for complete parameter list.

---

API

Training

model.fit(epochs=400, patience=25, val_every=5)

Inference

latent = model.get_latent()           # Latent embeddings
bottleneck = model.get_bottleneck()   # Bottleneck features
time = model.get_time()               # Pseudotime (ODE only)
velocity = model.get_velocity()       # RNA velocity (ODE only)
transition = model.get_transition()   # Transition matrix (ODE only)

Metrics

loss_hist = model.get_loss_history()
metrics_hist = model.get_metrics_history()
resources = model.get_resource_metrics()

---

Architecture

Input (n_genes)
    ↓
Encoder (log1p → MLP → latent_dim)
    ↓
[Optional ODE] Neural ODE dynamics
    ↓
Bottleneck (latent_dim → i_dim → latent_dim)
    ↓
Decoder (MLP → n_genes)
    ↓
Reconstruction (NB/ZINB/MSE/Poisson/ZIP)

[Optional MoCo] Contrastive learning on augmented views

---

Loss Functions

• Reconstruction: MSE, NB, ZINB, Poisson, ZIP
• KL Divergence: β-weighted regularization
• Disentanglement: DIP-VAE, β-TC-VAE, InfoVAE (MMD)
• ODE Regularization: MSE between VAE and ODE latents
• MoCo Contrastive: InfoNCE loss

---

Validation Metrics

• ARI: Adjusted Rand Index
• NMI: Normalized Mutual Information
• ASW: Silhouette Score
• CH: Calinski-Harabasz Index
• DB: Davies-Bouldin Index
• Corr: Latent correlation

---

Citation

@article{mocoo2025,
  title={MoCoO: Momentum Contrast ODE-Regularized VAE for Single-Cell Trajectory Inference},
  author={Ponyu, Peter},
  year={2025}
}

---

License

MIT License

---

Contact

GitHub: @PeterPonyu
Repository: MoCoO