How Noetik uses images to give its AI model a deeper under­standing of cancer

Feed enough images of tumor slices into Noetik’s machine learning model, and it will build a powerful internal repre­sen­ta­tion of cancer itself. This image from the DCVC-backed company is a composite of thousands of slices or ​“cores” from human lung cancer tissue, with a few healthy controls mixed in. Many tumor subtypes from many patients are represented, each stained to highlight a specific protein and/or a different type of cell (purple and red cores are mostly tumor cells; yellow and green areas are immune cells.) For each sample, Noetik has also generated paired genomics, tran­scrip­tomics, and histology data. 

As the model processes thousands of these cores — and is forced to learn how to plausibly reconstruct 100 percent of an image from just the first 2 percent — it learns how all the data modalities relate. It’s essentially ​“building a nuanced under­standing of the rela­tion­ship between different parameters across these tumors to become a foundation model of cell and tissue biology,” says Noetik co-founder and CEO Ron Alfa. That allows the model to predict, for example, how raising or lowering the abundance of certain drug targets would attract tumor-killing immune cells. The company aims to use its model to guide the development of new forms of precision cancer therapies.