Phenotype prediction from single-cell RNA-seq data using attention-based neural networks.

Area of Science:

• Computational Biology
• Genomics
• Immunology

Background:

• Disease phenotypes are driven by specific cell groups, often with unknown markers detectable only late.
• Conventional bulk assays like RNA-Seq lack cell-level resolution for early detection.
• Single-cell RNA sequencing (scRNA-seq) offers cell-level gene expression profiling but faces challenges with limited annotated samples for deep learning.

Purpose of the Study:

• To develop a novel method, ScRAT, for accurate phenotype prediction using scRNA-seq data.
• To address the challenge of limited annotated samples in training predictive models.
• To identify informative cells driving disease phenotypes without prior cell type knowledge.

Main Methods:

• ScRAT employs a mixup module to augment limited training samples.
• A multi-head attention mechanism identifies phenotype-informative cells without relying on cell type annotation.
• The method was validated using three public coronavirus disease (COVID) datasets.

Main Results:

• ScRAT significantly outperforms existing phenotype prediction methods on COVID datasets.
• The method's performance advantage increases with fewer training samples, demonstrating the effectiveness of sample mixup.
• High-attention cells identified by ScRAT align with novel findings in relevant literature, suggesting potential for new discoveries.

Conclusions:

• ScRAT effectively predicts disease phenotypes from scRNA-seq data, overcoming limitations of missing marker genes and small sample sizes.
• The approach holds significant potential for revealing novel molecular mechanisms and therapeutic strategies.
• The ScRAT code is publicly available for further research and application.