Posted March 1, 2017

Due to the great work of researchers over the last decades, there are numerous drugs and surgical techniques to cure patients of cancer — if the cancer is caught early. And that’s notoriously hard to do, especially when symptoms often only arise in late stages of some of the most lethal cancers. For years, early detection has been the missing piece in the cure to many cancers. Existing tests are not sufficiently accurate, with many false positives and false negatives: 70% to 80% of men with an elevated PSA (prostate-specific antigen) who have a biopsy, for example, do not have cancer. Mammograms, too, were recently deprecated for younger women, due to unacceptably high false positive rates. Without higher accuracy, tests are useless in prevention.

Examining cell-free DNA has been a major breakthrough in recent years, with the most attention on examining tumor DNA. But in fact, only about 0.1% of cell-free DNA in cancer patients comes from tumors. That’s a needle in a haystack search problem. What’s the solution? For Freenome‘s cancer test, it’s looking at the whole haystack — for all kinds of changes. Our blood contains many additional analytes that indicate immunological and metabolic changes. So, for example, rather than only studying DNA fragments from tumor cells, Freenome has learned to decode complex signals coming from other cells in your immune system that change because of the tumor elsewhere.

How exactly can one identify which changes are associated with cancer? This is where machine learning shines, in particular deep learning. Much of the biological details here are complex, far too complex for typical approaches in biomarkers (looking for a specific gene or metabolite). Deep learning methods can quickly identify the relevant patterns (or “features”, to those performing machine learning) in a way that — because of the complexity — humans just couldn’t do.

This, however, leads to another challenge: how to train the machine learning methods. While these are powerful, they require data to learn these complex patterns. Freenome has developed a particularly valuable approach here, as they can work samples that others can’t, allowing them to build large datasets to achieve high accuracy.

This kind of approach marks a new era where machine learning allows us to embrace the complexity that permeates much of biology — without oversimplification, overfitting, or bias — resulting in considerably more accurate tests and earlier cancer detection.

I’m very pleased to announce that we are leading Freenome’s series A financing. Having worked with the company since leading their seed round, we have been delighted with their progress and with the quality of their team, including the original founders and the new hires since our last investment. We look forward to supporting Freenome as they continue and accelerate their novel research, expand to clinical trials, and bring their cancer screenings to market.