AI maps two distinct defects inside breast tumors

Under the microscope, a breast tumor is not one uniform mass. It is a patchwork of cells, blood vessels, dividing cells and dying cells, with minute structures hidden inside each cell. A team at the University of Southampton has used its open-source AI platform, CenSegNet, to map those structures across 911 breast tumor specimens from 127 patients.

The structures are centrosomes: cellular organizing hubs that help cells divide correctly and maintain their shape. For more than a century, abnormal centrosomes have been recognized as a hallmark of cancer, but examining them across patient tissue has been difficult. CenSegNet can analyze hundreds of thousands of cells and, in this study, more than 330,000 centrosomes at single-cell resolution.

The result was a split where researchers had often seen one process. Some cells had acquired too many centrosomes; others had centrosomes that were abnormally enlarged. The defects behaved independently and could occupy different parts of the same tumor. Enlarged centrosomes were associated with more aggressive features, including higher tumor grade, lymph node involvement and certain genetic alterations. Patients with fewer enlarged centrosomes in the tumor core tended to have better overall survival.

Practically, the near-term change is for cancer researchers and pathologists, not yet for people receiving treatment. Spatial maps from CenSegNet could help identify tumors with particularly aggressive characteristics, test new biomarkers and clarify why one region of a tumor behaves differently from another. Drugs targeting proteins that control centrosome function are already in development, raising the possibility of matching a specific defect with a therapy—but that application remains future work.

The platform is not ready for routine clinical use. The Southampton team, led by Dr. Salah Elias, plans to combine its maps with genomic, transcriptomic and proteomic data to test whether centrosome-based biomarkers can guide treatment decisions. The software is freely available and has also been applied to kidney, colon and appendix tissues, allowing other researchers to test it beyond breast cancer.