Back to Models

scGCC

🎯 Predictive Models

Graph Contrastive Clustering

RNA

Graph neural network with MoCo-based contrastive learning for single-cell clustering and embedding (encoder-only)

Publications

scGCC: Graph Contrastive Learning for Single-cell RNA-seq

Tian et al.2023
Complexity
★★★
complex
Interpretability
★★☆
medium
Architecture
Graph Attention Encoder
Latent Dim
32
Used in LAIOR Framework

Graph Contrastive Learning

scGCC uses Graph Attention Networks combined with Momentum Contrast to learn robust cell embeddings while preserving cell-cell relationships

Main Idea

Learn cell embeddings by contrasting augmented graph views while maintaining cell adjacency information

Key Components

Graph Construction

Build cell graphs from expression similarity

GAT Encoder

Graph Attention layers for neighbor aggregation

MoCo Head

Momentum Contrast for contrastive learning

Augmentation

Feature masking and graph dropout for robustness

Mathematical Formulation

L = InfoNCE(z_q, queue); z learned via GAT + augmentation

Loss Functions

Contrastive Loss
InfoNCE(query, momentum_updated_queue)

Data Flow

Expression → kNN Graph → GAT Encoder → MoCo Head → Contrastive Embeddings (No Decoder)

Architecture Details

Architecture Type

Graph Attention Network with Contrastive Head (Encoder-Only)

Input/Output Types

single-cell → latent

Key Layers

GATMoCoHeadGraphAugmentation

Frameworks

PyTorchPyTorch Geometric

Tags

contrastivegraphclusteringencoder-onlyrna