Explainable deep neural networks for predicting sample phenotypes from single-cell transcriptomics.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-Seq) offers high-resolution molecular insights into cellular phenotypes.
• Analyzing scRNA-Seq data presents challenges due to unique data characteristics, necessitating advanced statistical methods.
• Existing sample classification methods are often unsuitable for scRNA-Seq data, highlighting the need for new algorithms.

Purpose of the Study:

• To develop an end-to-end pipeline, singleDeep, for streamlined analysis and deep neural network training on scRNA-Seq data.
• To enable robust prediction and characterization of sample phenotypes using scRNA-Seq data.
• To overcome limitations in usability and analytical power of current scRNA-Seq analysis methods.

Main Methods:

• Developed singleDeep, an integrated pipeline for scRNA-Seq data analysis.
• Utilized deep neural networks for phenotype prediction and characterization.
• Applied singleDeep to scRNA-Seq datasets from systemic lupus erythematosus, Alzheimer's disease, and COVID-19.

Main Results:

• singleDeep demonstrated strong diagnostic performance across multiple disease datasets, validated internally and externally.
• The pipeline outperformed traditional machine learning and alternative single-cell analysis methods.
• singleDeep provided valuable insights into cell type-specific gene importance for phenotypic characterization.

Conclusions:

• singleDeep offers a robust and user-friendly solution for scRNA-Seq data analysis and phenotype prediction.
• The pipeline effectively identifies disease-relevant genes and cell types, advancing precision medicine.
• singleDeep's ability to uncover cell-type-specific gene roles enhances understanding of complex diseases.