scSorterDL: a deep neural network-enhanced ensemble LDAs for single cell classifications.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals cellular heterogeneity but poses challenges for accurate cell type annotation.
• High dimensionality and sparsity in scRNA-seq data complicate classification using traditional methods.

Purpose of the Study:

• To develop and validate scSorterDL, an advanced computational tool for robust and accurate cell type annotation in scRNA-seq data.
• To leverage a hybrid approach combining pLDA, swarm learning, and DNNs for improved classification performance.

Main Methods:

• scSorterDL employs penalized Linear Discriminant Analysis (pLDA) on data subsets to capture diverse cellular features.
• A deep neural network (DNN) consolidates pLDA outputs, identifying complex interactions for enhanced classification.
• The approach utilizes GPU computing for efficient processing of large-scale, high-dimensional gene expression data.

Main Results:

• scSorterDL demonstrated superior accuracy and robustness compared to nine existing cell annotation tools across 13 diverse scRNA-seq datasets.
• The method showed exceptional performance in both cross-validation and cross-platform validation scenarios.
• Validation on 20 cross-platform dataset pairs confirmed the tool's adaptability and reliability.

Conclusions:

• scSorterDL offers a powerful and adaptable solution for automated cell type annotation in scRNA-seq research.
• The integration of pLDA, swarm learning, and DNNs effectively addresses the challenges of high dimensionality and sparsity.
• The tool's performance highlights its potential to advance the analysis of cellular diversity.