scGO: interpretable deep neural network for cell status annotation and disease diagnosis

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals cellular heterogeneity but faces challenges with deep learning model interpretability.
• The
• black box
• nature of current models hinders accurate cell status annotation.

Purpose of the Study:

• Introduce scGO, a Gene Ontology (GO)-inspired deep learning framework for interpretable cell status annotation in scRNA-seq data.
• Enhance the understanding of cellular heterogeneity and biological processes through interpretable AI models.

Main Methods:

• Developed scGO, a framework utilizing sparse neural networks to integrate gene-GO term relationships.
• Leveraged intrinsic biological relationships among genes, transcription factors, and GO terms for enhanced interpretability.
• Incorporated in silico gene manipulations for therapeutic target discovery.

Main Results:

• scGO significantly improves interpretability and reduces computational cost in scRNA-seq analysis.
• Demonstrated superior performance over state-of-the-art methods in cell subtype characterization across diverse datasets.
• Showcased efficacy in disease diagnosis, developmental stage prediction, and assessment of senescence and disease severity.

Conclusions:

• scGO provides an interpretable deep learning model for accurate cell status annotation in scRNA-seq.
• The framework captures latent biological knowledge, offering valuable insights for clinical practice and therapeutic target identification.