Using algorithms against cancer

To diagnose cancer earlier and fight and treat the disease more effectively, machine learning (ML) methods are increasingly being used in cancer research. Rebekka Burkholz, senior researcher at CISPA since September, is working, among other things, on the further development of ML models so that they can predict the development and course of the disease ever more accurately in the future. She presents an innovative approach in the paper "Scaling up Continuous-Time Markov Chains Helps Resolve Underspecification," which will be presented at the prestigious NeurIPS 2021 conference (Conference on Neural Information Processing Systems).

The cause of cancer is out-of-control cells that proliferate uncontrollably, spread to surrounding tissue, and destroy it, explains the German Cancer Research Center in the Helmholtz Association (dkfz). A normal cell becomes a tumor cell when some harmful changes in and on the genetic information take place beforehand. According to the dkfz, such mutations can be promoted by hereditary predisposition but also by environmental influences such as UV radiation or harmful behaviors such as smoking. Not every mutation necessarily leads to cancer. Often, several gene mutations have to come together before the disease breaks out.

"An important aspect of the development of cancer is how different gene mutations interact and in what order they occur," says Rebekka Burkholz. She is an expert in machine learning methods and is involved in their application in the medical field, among other things. "Determining when which mutation is added has been very difficult until now because we can almost always only collect data when cancer has already progressed," explains the 34-year-old. Many different mutations can then be detected in the tumors. Where the chain started and what dependencies exist has so far only been traceable for a few mutations. "The common model we looked at so far only included about 20 out of more than a thousand mutations in its prediction," Burkholz says. Together with colleagues at MIT and Harvard University, the researcher has developed an approach that improves existing ML models so that they can predict cancer development much more accurately in the future. "Our approach allows us to extend the process to hundreds of mutations and is much faster."

Machine learning uses algorithms that infer patterns and regularities from observations of data to build statistic models that enable automated predictions. "There are many different approaches to machine learning. Many people immediately think of neural networks, which need a lot of data to be trained. But in the medical field, we often work with very manageable amounts of data," Burkholz says. ML methods help researchers in these areas to refine prediction models and thus gather more useful information. In this way, the available data can be better evaluated and ultimately new treatment methods can be developed by understanding correlations.

The 34-year-old moved to CISPA from Harvard University in the summer and focuses primarily on artificial intelligence and machine learning. The Frankfurt native earned her doctorate after studying math and physics at ETH Zurich, where she received the Zurich Dissertation Prize for her work. "I am very excited about what awaits me at CISPA and am already looking forward to the exchange with my new colleagues. The perspective of the cybersecurity experts here is very different from my own. We can certainly enrich each other. "