DeepAE

DeepAE applies a deep auto-encoder to compress and reconstruct high-dimensional, sparse gene expression profiles from single-cell RNA sequencing (scRNA-seq) to identify key transcriptomic dimensions.

Key Features:

• Auto-encoder architecture: A neural network with an input layer, seven hidden layers, and an output layer implementing encoding and decoding phases.
• Encoding and decoding phases: Sequential compression (encoding) and decompression (decoding) of transcriptomic data to reveal lower-dimensional representations.
• Dimensionality reduction: Reduces dimensionality of scRNA-seq data while preserving essential biological information.
• High-dimensional and sparse data handling: Targets high-dimensional, sparse gene expression profiles generated by scRNA-seq technologies.
• Benchmark evaluation: Comparative experiments across nine diverse transcriptomic profiling datasets showing performance versus four benchmark methods.
• Cross-platform evaluation: Performance assessment on mass cytometry and metabolic profiling datasets.
• Biological interpretation analyses: Application of gene ontology enrichment and pathology analyses to investigate mechanisms underlying identified dimensions.
• Robustness and efficacy: Demonstrated ability to identify key transcriptomic dimensions across multiple datasets and platforms.

Scientific Applications:

• Single-cell transcriptome analysis: Elucidates complex high-dimensional scRNA-seq datasets by extracting salient transcriptomic dimensions.
• Pattern and relationship discovery: Uncovers significant patterns and relationships obscured in raw high-dimensional data.
• Cross-omics evaluation: Applies to mass cytometry and metabolic profiling for cross-platform dimensionality analysis.
• Functional interpretation: Supports gene ontology enrichment and pathology analyses to interpret discovered dimensions.
• Benchmarking and method comparison: Enables comparative evaluation across nine transcriptomic datasets against four benchmark methods.

Methodology:

Uses a deep auto-encoder with an input layer, seven hidden layers, and an output layer implementing encoding and decoding to compress and reconstruct data; includes comparative experiments on nine datasets against four benchmark methods and downstream gene ontology enrichment and pathology analyses.