Comprehensive single-cell RNA-seq analysis using deep interpretable generative modeling guided by biological...

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Single-cell technologies enable exploration of cellular heterogeneity.
• Deep learning, particularly generative models, has advanced transcriptomic data analysis.
• Existing generative models often lack interpretability and depth, limiting biological insights.

Purpose of the Study:

• To develop a deep interpretable generative model (d-scIGM) for enhanced single-cell data analysis.
• To improve the biological interpretability and analytical capabilities beyond existing shallow models.
• To apply d-scIGM to diverse tasks like clustering, visualization, and drug response analysis.

Main Methods:

• Developed d-scIGM, a deep generative framework using sawtooth connectivity and residual networks.
• Incorporated hierarchical biological domain prior knowledge for enhanced interpretability.
• Evaluated performance on clustering, visualization, pseudo-temporal inference, and drug response data.

Main Results:

• d-scIGM demonstrated superior performance in clustering, visualization, and pseudo-temporal inference.
• Learned topics from d-scIGM were significantly enriched for biologically meaningful pathways.
• Successfully captured drug response patterns and identified key genes/pathways in a melanoma dataset.

Conclusions:

• d-scIGM offers a powerful and interpretable approach for single-cell data analysis.
• The model facilitates deeper understanding of cellular heterogeneity and biological mechanisms.
• d-scIGM shows promise for drug development and disease mechanism elucidation.