Back to Models

scDiffusion

G Generative

Diffusion Model for Single-Cell

RNA

Conditional diffusion model for high-quality single-cell data generation with cell-type control

Publications

scDiffusion: conditional generation of high-quality single-cell data using diffusion model

Luo et al.2024
Complexity
★★★
complex
Interpretability
★☆☆
low
Architecture
Conditional Diffusion Model
Latent Dim
128
Used in LAIOR Framework

Conditional Denoising Diffusion

scDiffusion combines diffusion models with foundation models to generate realistic single-cell data conditioned on cell types via iterative denoising (reconstruction process)

Main Idea

Generate high-quality synthetic single-cell data by learning to reverse a noise diffusion process through denoising reconstruction

Key Components

Forward Diffusion

Gradually add Gaussian noise over T timesteps

Noise Predictor (Encoder)

Transformer-based network predicts noise at each step

Conditional Generation

Condition on cell type, tissue, or experimental factors

Reverse Sampling (Decoder)

Iteratively denoise to reconstruct/generate new samples

Mathematical Formulation

q(x_t|x_{t-1}) = N(√(1-β_t)x_{t-1}, β_t I); p_θ(x_{t-1}|x_t,c) learned via denoising

Loss Functions

Denoising Loss
||ε - ε_θ(x_t, t, c)||² (reconstruction of clean data from noise)

Data Flow

Real Data + Condition → Forward Noise Addition → Noisy Data → Noise Predictor → Reverse Denoising → Reconstructed/Generated Data

Architecture Details

Architecture Type

Conditional Denoising Diffusion Probabilistic Model (VAE-like Architecture)

Input/Output Types

single-cell → reconstruction

Key Layers

TransformerBlockTimeEmbeddingConditionEncoderDenoisingDecoder

Frameworks

PyTorch

Tags

diffusiongenerativeconditionalreconstructionrna