Back to Models

CLEAR

🎯 Predictive Models

Contrastive Learning for Enhanced scRNA-seq

RNA

MoCo-based contrastive learning framework for learning robust cell representations (encoder-only)

Publications

CLEAR: Contrastive Learning for Single-cell RNA-seq

Zhang et al.2022
Complexity
★★☆
moderate
Interpretability
★★☆
medium
Architecture
MoCo Encoder
Latent Dim
128
Used in LAIOR Framework

Momentum Contrast Learning

CLEAR applies Momentum Contrast to single-cell RNA-seq data with carefully designed augmentations (masking, noise, jittering) to learn invariant cell embeddings

Main Idea

Learn robust, batch-invariant representations through contrastive learning with momentum contrast queue

Key Components

Query Encoder

Main encoder network for cell embeddings

Momentum Encoder

Slow-moving encoder maintaining feature consistency

Augmentation Module

Feature masking, noise injection, and scaling jitter

Contrastive Queue

Large memory bank of negative samples

Mathematical Formulation

L = -log(exp(sim(z_q, z_k^+)/τ) / Σ_i exp(sim(z_q, k_i)/τ))

Loss Functions

InfoNCE Loss
Contrastive loss with queue-based negatives

Data Flow

Data → Two Augmented Views → Query/Momentum Encoders → Feature Similarity → Contrastive Loss (No Decoder)

Architecture Details

Architecture Type

Momentum Contrast Framework (Encoder-Only)

Input/Output Types

single-cell → latent

Key Layers

MLPEncoderMoCoHeadAugmentation

Frameworks

PyTorch

Tags

contrastiveembeddingbatch-robustencoder-onlyrna