Back to Models

scDAC

✨ Generative Models

Deep Adaptive Clustering

RNA

Deep autoencoder with Dirichlet Process Mixture Model for adaptive cell clustering

Publications

scDAC: Deep Adaptive Clustering for Single-cell RNA-seq

Tian et al.2021
Complexity
★★☆
moderate
Interpretability
★★★
high
Architecture
Adaptive Autoencoder
Latent Dim
32
Used in LAIOR Framework

Adaptive Clustering via DPMM

scDAC combines deep autoencoders with Dirichlet Process Mixture Models to perform nonparametric Bayesian clustering on learned embeddings

Main Idea

Learn both embeddings and cluster assignments adaptively using infinite mixture models without pre-specifying cluster count

Key Components

Deep Autoencoder

Learns compressed cell representations

DPMM Clustering

Nonparametric Bayesian clustering of embeddings

Adaptive K

Automatically determines optimal number of clusters

Flexible Activation

Supports ReLU, Mish, Sigmoid activations

Mathematical Formulation

z ~ AE(x); z | π,μ,Σ ~ ΣΠ_k π_k N(μ_k, Σ_k)

Loss Functions

Reconstruction
MSE(x, AE_decode(z))
DPMM
Bayesian clustering objective

Data Flow

Expression → Autoencoder Encoder → Embeddings → DPMM Inference → Autoencoder Decoder → Clusters + Posterior

Architecture Details

Architecture Type

Deep Autoencoder with DPMM (VAE Architecture)

Input/Output Types

single-cell → reconstruction

Key Layers

MLPEncoderDPMMLayerMLPDecoder

Frameworks

PyTorchscikit-learn

Tags

clusteringadaptiveautoencodervaerna