Machine learning around the clock

From letting us know when to sleep to influencing the effectiveness of medications, our circadian clock exerts a powerful influence on our lives. However, you can’t simply check the time of your circadian clock with a watch. A new paper published today Genome Medicine showcases a new machine learning tool to monitor the clock in human blood named ZeitZeiger. Here, author of the study, Dr. Jacob Hughey, tells us more.

Quick, what’s your time? Not the time of day outside, but the time inside your own body. You might not know it, but almost every organ in your body has its own clock, which scientists call a circadian clock. Your circadian clocks work together to tell you when to go to sleep and can affect your performance on a test or in a sport. In addition, because the clocks coordinate rhythms throughout your body, some medications (and possibly the flu shot) may be more or less effective, depending on the time of day at which you take them.

Unfortunately, our modern environment interferes with our circadian clocks in multiple ways. For example, we’re exposed to too little sunlight during the day and too much electrical lighting at night. Better methods to monitor the clocks throughout the body could help us better diagnose and treat sleep and circadian clock-related disorders.

Importantly, everyone’s clock is different. For example, some people naturally wake up early, whereas others (such as myself) rely on an alarm clock and/or a significant other to get up in the mornings. Although you can check the time of day outside by simply looking at your watch or phone, monitoring the time of day inside your body is not as easy.

In this paper, I developed a computational method to monitor the clock in human blood. In three previous studies, researchers collected blood samples every 3-4 h for a day from a total of 60 individuals, then measured how strongly almost every gene in the genome (~17,000) was turned on in each sample.

I applied the data from these studies, which the researchers made publicly available, to a machine learning program I previously developed called ZeitZeiger (German for “time revealer”). ZeitZeiger took those data, almost 9 million data points, and figured out how to accurately predict time of day.

Given a single blood sample, ZeitZeiger predicted time of day in the blood to within 2.1 hours on average. Surprisingly, the predictor contains only 15 genes. This compactness is important for making the predictor interpretable and practical. For the most part, these 15 genes are not the gears that control the clock, but the hands that tell the time (the gears have been discovered through a lot of hard work by researchers over the last 20 years or so).

To predict time of day in human blood even more accurately, I developed two extensions to ZeitZeiger (for details, please check out the paper). For example, instead of using a single blood sample, using two samples taken 12 hours apart improves accuracy by 28%. Finally, I used the 15 genes to quantify how the clock in the blood ticks differently in each of the 60 individuals.

These results are exciting, but there’s still much more work to do. For one, these data were collected from humans under controlled conditions in the lab, so it will be important to validate the results with humans in the real world. Ultimately, however, these results and the machine learning approach in this paper could help us harness the circadian clock to improve human health. It’s about time.

View the latest posts on the On Medicine homepage