scCaT: An explainable capsulating architecture for sepsis diagnosis transferring from single-cell RNA sequencing.

Area of Science:

• Computational Biology
• Genomics
• Artificial Intelligence in Medicine

Background:

• Sepsis is a critical condition with high mortality, necessitating accurate and timely diagnosis.
• Current deep learning models for medical tasks require extensive data and often lack transparency, limiting their use in sepsis diagnosis.
• Single-cell RNA sequencing (scRNA-seq) offers high-resolution gene expression data but presents analytical challenges.

Purpose of the Study:

• To introduce scCaT, a novel deep learning framework for sepsis diagnosis.
• To leverage explainable AI by integrating a capsulating architecture with Transformer models for gene expression analysis.
• To demonstrate the model's effectiveness on both single-cell and bulk RNA sequencing data.

Main Methods:

• Developed scCaT, a deep learning framework combining a capsulating architecture for gene grouping with a Transformer decoder for classification.
• Trained and validated the model on single-cell RNA sequencing data for sepsis diagnosis.
• Transferred the trained model to analyze seven independent bulk RNA sequencing cohorts.

Main Results:

• Achieved a high diagnostic accuracy with an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.93 on the single-cell test set.
• Demonstrated excellent performance on bulk RNA data, with an average AUROC of 0.98 across seven cohorts.
• The model's capsules identified cell types and biological pathways distinguishing sepsis from control samples, providing explainability.

Conclusions:

• scCaT offers an accurate and explainable deep learning approach for sepsis diagnosis using RNA sequencing data.
• The framework effectively learns gene modules and demonstrates successful transferability across different data types (scRNA-seq to bulk RNA).
• This approach holds promise for diagnosing rare diseases with limited patient data.