SCimilarity

US researchers have developed a deep metric learning framework called SCimilarity to address the challenges of analyzing and querying single-cell RNA sequencing (scRNA-seq) data across diverse studies and conditions. A report published in Nature describes how the model enables efficient searches of over 23.4 million human cell profiles, spanning 412 studies and representing a wide array of tissues, diseases, and experimental contexts.

SCimilarity organizes data in a way that groups similar cells based on their gene activity patterns, making it easier to identify connections between them. The system uses advanced learning techniques to balance sensitivity and consistency across different datasets. It was trained on data from 7.9 million cells, enabling it to recognize patterns in new, unseen datasets while working reliably across various research platforms.

To test SCimilarity, the researchers used macrophage and fibroblast profiles from interstitial lung disease (ILD) as queries. The model identified similar cells across datasets in a matter of seconds, revealing shared states in fibrotic diseases, COVID-19, and various cancers. Notably, SCimilarity demonstrated its ability to differentiate closely related cell types, outperforming other computational tools in precision and speed. For instance, querying macrophages associated with fibrosis showed that these cells are present in fibrotic lung diseases and certain cancers like pancreatic ductal adenocarcinoma.

In addition to in vivo searches, SCimilarity successfully identified an ex vivo model that mimics fibrosis-associated macrophages. Using public datasets, it pinpointed a 3D hydrogel culture system that replicated these cell states in vitro; the finding was later validated through experimental replication, highlighting the model’s ability to bridge the gap between observational studies and experimental biology.

By enabling scalable and interpretable cell queries, SCimilarity could become a foundational tool in single-cell research, with applications ranging from discovering new cell state to understanding disease mechanisms. 

The open-source framework provides researchers with a powerful resource to accelerate insights from the growing Human Cell Atlas. Future enhancements may expand its capabilities, supporting the integration of even more diverse data types and biological contexts.