SEDIM

SEDIM performs imputation of gene expression in single-cell RNA sequencing (scRNA-seq) data by automatically designing and optimizing deep neural network architectures using a surrogate-assisted evolutionary approach to address gene sparsity.

Key Features:

• Deep learning imputation: Employs deep learning models specifically tailored to impute gene expression levels in scRNA-seq datasets.
• Surrogate-assisted evolutionary algorithm: Uses a surrogate-assisted evolutionary algorithm to guide the search for optimal neural network architectures.
• Automated neural architecture design: Automatically designs and optimizes neural network architectures to adapt to dataset-specific characteristics without manual tuning.
• Offline surrogate model: Constructs an offline surrogate model to accelerate the architectural search and reduce computational cost.
• Gene sparsity handling: Targets inherent gene sparsity in single-cell datasets to improve robustness of imputation.
• Improved accuracy and clustering: Demonstrates improved imputation accuracy and enhanced clustering performance relative to benchmark methods.

Scientific Applications:

• scRNA-seq data imputation: Recovers missing or zero-inflated gene expression values in single-cell RNA-seq experiments.
• Clustering and cell-type identification: Improves downstream clustering to support more accurate cell-type identification.
• Mass cytometry: Applies to high-throughput mass cytometry datasets for data denoising and imputation.
• Metabolic profiling: Extends to metabolic profiling datasets for signal recovery and analysis.
• Marker gene detection: Facilitates detection of marker genes from imputed expression profiles.
• Gene ontology enrichment analysis: Supports downstream gene ontology enrichment analyses using imputed data.
• Pathological assessments: Assists pathological assessments to investigate underlying biological mechanisms.

Methodology:

SEDIM applies deep learning–based imputation combined with a surrogate-assisted evolutionary algorithm for automated neural architecture search and employs an offline surrogate model to accelerate the architectural search process.