scPhase: Exploring phenotype-related single-cells through attention-enhanced representation learning

The analysis of single-cell RNA-seq (scRNA-seq) data is pivotal for linking cellular heterogeneity to disease phenotypes, yet current methods often struggle with data sparsity, batch effects, and limited interpretability. Here, we present scPhase (phenotype prediction with attention mechanisms for single-cell exploring), a unified deep learning framework designed to predict clinical phenotypes directly from raw scRNA-seq profiles. To ensure robustness and scalability, scPhase integrates an efficient LinFormer-attention module to capture complex inter-cellular dependencies, a Mixture-of-Experts (MoE)-based Attention Multiple Instance Learning (AMIL) module to aggregate cellular information into patient-level representations, and an adversarial domain adaptation component to effectively mitigate batch effects across cohorts. Unlike existing approaches, scPhase does not rely on pre-defined cell type labels and offers multi-level interpretability at both the gene and cellular levels. We validated scPhase on five large-scale public datasets spanning COVID-19, aging, Alzheimer's disease, non-small cell lung cancer, and colorectal cancer. The results demonstrate that scPhase not only achieves superior predictive accuracy but also reveals critical biological insights into disease mechanisms.