Learning multi-cellular representations of single-cell transcriptomics data enables characterization of patient-level...

Area of Science:

• Computational Biology
• Genomics
• Translational Medicine

Background:

• Single-cell RNA sequencing (scRNA-seq) provides high-resolution insights into cellular heterogeneity in human diseases.
• Existing single-cell foundation models primarily focus on individual cell representations, limiting holistic understanding of disease states.
• Analyzing complex tissue ecosystems requires methods that integrate information from multiple interacting cells.

Purpose of the Study:

• To develop a machine learning model, PaSCient, for generating patient-level representations from multi-cellular expression contexts.
• To enable fine-grained analysis of disease characteristics by providing cell and gene importance scores.
• To improve disease classification and downstream analyses using comprehensive tissue-level data.

Main Methods:

• PaSCient employs a multi-level representation learning paradigm to process scRNA-seq data.
• The model integrates expression data from millions of cells across thousands of patients.
• Cellular and gene-level importance scores are derived for interpretability.

Main Results:

• PaSCient demonstrated superior performance in disease classification compared to existing methods.
• The model effectively identified patient subgroups and prioritized key genes and cell types associated with disease.
• Benchmarking confirmed PaSCient's utility in dimensionality reduction and feature prioritization.

Conclusions:

• PaSCient offers a powerful approach for understanding disease biology at the patient and tissue ecosystem levels.
• The model's ability to provide interpretable importance scores facilitates deeper insights into disease mechanisms.
• This work advances the application of machine learning to large-scale scRNA-seq data for clinical relevance.