Multimodal weakly supervised learning to identify disease-specific changes in single-cell atlases

Multimodal analysis of single-cell samples from healthy and diseased tissues at various stages provides a comprehensive view that identifies disease-specific cells, their molecular features and aids in patient stratification. Here, we present MultiMIL, a novel weakly-supervised multimodal model designed to construct multimodal single-cell references and prioritize phenotype-specific cells via patient classification. MultiMIL effectively integrates single-cell modalities, even when they only partially overlap, providing robust representations for downstream analyses such as phenotypic prediction and cell prioritization. Using a multiple-instance learning approach, MultiMIL aggregates cell-level measurements into sample-level representations and identifies disease-specific cell states through attention-based scoring. We demonstrate that MultiMIL accurately identifies disease-specific cell states in blood and lung samples, identifying novel disease-associated genes and achieving superior patient classification accuracy compared to existing methods. We anticipate MultiMIL will become an essential tool for querying single-cell multiomic atlases, enhancing our understanding of disease mechanisms and informing targeted treatments.