Machine learning for cancer diagnostics

We are merging our knowledge of embryonic development with our knowledge of cancer to build the next generation of cancer diagnostics. Our hypothesis is that many cancers are ultimately  developmental diseases, with activation of normal embryonic programs in the wrong time and place.  We are merging datasets cataloging mammalian development with datasets cataloging cancer to build a comprehensive developmental map of human tumors.  We can use these maps to build predictive machine learning models that can aid diagnosis of cancers in the clinic.  This will enable better precision medicine.

Shown above is the developmental deconvolution profile of a single hepatocellular carcinoma tumor. Clockwise around the circle are 214 main mammalian developmental programs; distance from the center represents how strongly each program operates in this tumor. Developmental programs are also grouped into 10 main programs (colors), with particularly high signal for hepatic programs (peach, 7 o'clock) in this sample. Reproduced with permission from Moiso et al,  Cancer Discovery 2022.