Back to Models

scSMD

Generative Models

ResNet-based Clustering

RNA

ResNet autoencoder with mutual information clustering for single-cell analysis

Publications

scSMD: Single-cell clustering via mutual information with ResNet

Song et al.2020
Complexity
moderate
Interpretability
low
Architecture
ResNet Autoencoder
Latent Dim
32
Used in LAIOR Framework

CNN-based Clustering via Mutual Information

scSMD reshapes gene expression as 2D images and uses ResNet autoencoders with mutual information-based clustering

Main Idea

Leverage CNN architectures by treating gene expression as 2D data, clustering via maximizing mutual information with reconstruction

Key Components

ResNet Encoder

2D CNN-based feature extraction from reshaped genes

Mutual Information

Clustering by maximizing MI between data and clusters

ResNet Decoder

Reconstructs 2D gene expression via transposed convolutions

NB Loss

Negative Binomial reconstruction loss

Mathematical Formulation

L = NB_loss(X, X̂) + β*MI(X; C)

Loss Functions

NB Reconstruction
Negative Binomial loss
MI Clustering
Mutual information between data and clusters

Data Flow

Expression → Reshape 2D → ResNet Encoder → MI Clustering → ResNet Decoder → Reconstructed Expression

Architecture Details

Architecture Type

2D CNN Autoencoder with ResNet (VAE Architecture)

Input/Output Types

single-cellreconstruction

Key Layers

Conv2DBottleneckBlockTransposedConvDecoder

Frameworks

PyTorch

Tags

clusteringcnnmutual-informationvaerna