Annotation-free phenotype prediction using knowledge-augmented clustering from single-cell RNA sequencing data

Area of Science:

• Computational biology
• Genomics
• Biotechnology

Background:

• Single-cell RNA sequencing (scRNA-seq) enables precise phenotype prediction and identification of disease-associated cell subpopulations.
• Existing computational methods often depend on predefined cell-type annotations, leading to sensitivity to annotation quality, inconsistencies, and dataset biases, limiting generalizability.

Purpose of the Study:

• To develop an annotation-free computational framework, scCap, for robust phenotype prediction from scRNA-seq data.
• To overcome the limitations of annotation-dependent models and improve the generalizability of predictive performance across diverse patient cohorts.

Main Methods:

• scCap employs knowledge-augmented clustering by refining initial clusters within the embedding space of a pretrained single-cell foundation model.
• It integrates these knowledge-augmented clusters into a hierarchical multiple instance learning framework with dual-level attention for interpretable cell- and cluster-level predictions.

Main Results:

• scCap consistently outperformed baseline models in predictive accuracy across three public scRNA-seq datasets.
• The framework successfully identified known disease-associated subpopulations without relying on prior cell-type annotations.

Conclusions:

• scCap offers a robust and interpretable solution for annotation-free phenotype prediction in scRNA-seq data.
• This approach enhances the reliability and applicability of scRNA-seq analysis in diverse biological and clinical contexts.