scDeepCluster

Back to Models

Autoencoder with Zero-Inflated Negative Binomial reconstruction and DEC-style clustering

Publications

scDeepCluster: Clustering single-cell RNA-seq data with deep learning and ZINB

Tian et al.2019

Complexity

★★☆

moderate

Interpretability

★★☆

medium

Architecture

DEC-based Autoencoder

Latent Dim

Used in LAIOR Framework

Joint Clustering and Reconstruction

scDeepCluster uses autoencoders with ZINB loss and clustering refinement via DEC (Deep Embedded Clustering) approach

Main Idea

Learn clusters by jointly optimizing reconstruction loss and clustering consistency with DEC-style centroid-based refinement

Key Components

Encoder

Maps expression to latent embeddings

ZINB Decoder

Zero-Inflated NB reconstruction loss for count data

KMeans Initialization

Pre-clustering for warm start

DEC Refinement

Iterative cluster assignment refinement

Mathematical Formulation

p(x|z) = ZINB(π, μ, θ); clustering via centroid distances

Loss Functions

ZINB Reconstruction

Zero-Inflated NB loss

Clustering

KL divergence to soft cluster assignments

Data Flow

Expression → AE Encoder → Pretrain → KMeans Init → Joint Optimization → ZINB Decoder → Clusters + Latent

Architecture Details

Architecture Type

Autoencoder with ZINB + DEC (VAE Architecture)

Input/Output Types

single-cell → reconstruction

Key Layers

EncoderZINBDecoderClusteringLayer

Frameworks

PyTorch

Deep Clustering with ZINB