Single cell RNA sequencing has become one of the most powerful tools for studying complex tissues because it allows researchers to examine the gene activity of individual cells. One of its most common uses is identifying different cell types within a sample. However, distinguishing closely related cell subtypes, especially in cancer, remains a major challenge because tumors contain highly diverse and constantly changing cell populations.

Researchers at the University of Michigan developed a computational framework called ScPanKD to improve the identification of fine grained T cell subtypes in cancer samples. Rather than relying on conventional annotation methods that often struggle when reference datasets differ from new samples, ScPanKD uses artificial intelligence to learn features that remain consistent across many different cancer types.

The framework is based on a machine learning technique known as knowledge distillation. In this approach, information learned from large, highly detailed reference datasets is transferred into a streamlined model that can accurately classify new cells. ScPanKD also uses a two step strategy that first identifies reliable anchor cells before applying knowledge distillation, helping reduce differences between reference and query datasets.

The researchers evaluated ScPanKD using large pan cancer collections of both CD4 positive and CD8 positive T cells. Across multiple experiments, the framework consistently outperformed existing cell annotation methods as well as recently developed single cell foundation models. It produced more accurate and more reliable identification of T cell subtypes despite the biological variability found in cancer tissues.

Accurate T cell subtype identification is becoming increasingly important because different T cell populations play distinct roles in tumor growth and immune responses. Some T cells actively attack cancer cells, while others suppress immune activity or become exhausted during prolonged disease. Correctly identifying these populations helps researchers better understand the tumor microenvironment and may improve the development of immunotherapies.

The authors have made ScPanKD and its source code freely available, allowing other researchers to incorporate the framework into their single cell RNA sequencing analysis pipelines. As cancer atlases continue to expand, tools that improve cell subtype annotation will become increasingly valuable for discovering new biology and identifying potential therapeutic targets.

Availability – ScPanKD and all reproducible scripts are available at https://github.com/marvinquiet/ScPanKD