Deep interpretable learning of sample representations for characterizing disease states in single-cell...

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Single-cell transcriptomics reveals cellular heterogeneity but struggles with sample-level disease representation.
• Disease annotations are often coarse, limiting the analysis of systemic states.
• Existing methods face challenges in integrating diverse single-cell data for disease phenotyping.

Purpose of the Study:

• To introduce Phenoverse, an interpretable deep learning framework for learning sample-level disease representations from single-cell transcriptomic data.
• To demonstrate Phenoverse's ability to predict disease states and severity using cell type-aware analysis.
• To enhance the interpretability of complex single-cell data for patient-level biological insights.

Main Methods:

• Phenoverse employs cell type-aware residual encoding, prototype learning, and Perceiver-based aggregation.
• The framework was applied to large-scale single-cell transcriptomic cohorts of COVID-19, Alzheimer's disease, and systemic lupus erythematosus.
• Trajectory-derived genes and prototype learning were used for cross-cohort analysis and interpretability.

Main Results:

• Learned sample representations accurately predicted disease states and severity spectrum on unseen data.
• Results showed correlations between learned representations and clinical/pathological measures.
• Trajectory-derived genes revealed reproducible cross-cohort molecular programs, outperforming traditional comparisons.

Conclusions:

• Phenoverse offers an interpretable approach to disease phenotyping from single-cell transcriptomic data.
• The framework successfully translates complex cellular data into patient-level biological insights.
• This method enhances understanding of disease heterogeneity and severity across different conditions.