AutoClass

AutoClass applies a dual-component deep learning architecture to denoise single-cell RNA sequencing (scRNA-Seq) data and preserve biological signal for downstream analyses.

Key Features:

• Integration of deep neural networks: A dual-component architecture combining an autoencoder and a classifier to capture underlying data structure and separate biological variation from technical artifacts.
• Noise removal without distributional assumptions: Cleans a broad spectrum of noises and artifacts, including dropouts, without relying on explicit distributional models.
• Signal retention: Preserves essential biological signals while reducing technical noise.
• Hyperparameter robustness: Maintains consistent performance across bottleneck layer size, number of pre-clustering steps, and classifier weight adjustments.

Scientific Applications:

• Data recovery: Recovers higher-quality expression profiles from noisy scRNA-Seq inputs.
• Differential expression analysis: Improves identification of differentially expressed genes by reducing noise interference.
• Clustering analysis: Enhances reliability of clustering results for cell-type and cell-state identification by mitigating artifacts.
• Batch effect removal: Attenuates batch effects in multi-sample scRNA-Seq studies to facilitate cross-sample comparisons.

Methodology:

A dual-component deep learning architecture comprising an autoencoder and a classifier, with pre-clustering steps and adjustable hyperparameters (bottleneck layer size and classifier weight), performs denoising without explicit distributional assumptions.