A VAE-Driven One-Class Generative Fusion Approach for Cell-Type Annotation in scRNA-seq

Accurate cell-type annotation in single-cell RNA sequencing (scRNA-seq) is hindered by extreme sparsity, high dimensionality and strong overlap between related populations. We propose a hybrid framework that combines biologically motivated feature selection, supervised dimensionality reduction and class-specific generative modeling. First, we select highly variable genes (HVGs) and apply standard preprocessing to obtain a denoised expression matrix. On top of this matrix, we learn several complementary low-dimensional embeddings using principal component analysis (PCA), supervised PCA, a one-dimensional convolutional network and a multi-layer perceptron (MLP). In parallel, we train a bank of one-class variational autoencoders and compute SIMCA-style reconstruction and latent distance scores for every sample and class. These VAE-SIMCA scores are concatenated with each embedding and fed to lightweight classifiers such as multinomial logistic regression and linear SVM. Finally, Generative Topographic Mapping (GTM) is applied for interpretability, providing a probabilistic two-dimensional visualization of the fused representation. Experiments on four benchmark datasets (PBMC, Multiple Sclerosis, Zhengh۶۸K and Pancreatic) show that the fused feature sets achieve competitive or slightly improved accuracy and macro-F۱ compared with strong baselines, while GTM offers an interpretable view of the learned cell-type manifold.

کلیدواژه ها:

single-cell RNA-seq ، cell-type annotation ، highly variable genes ، supervised PCA ، VAE-SIMCA ، dimensionality reduction ، GTM ، manifold visualization