FYI.

This story is over 5 years old.

Tech

How Artificial Intelligence and Physics Will Help Beat Cancer

Someday statistical mechanics might save your life.
Image: Robert Couse-Baker/Flickr

Cancer has a data problem. Better and faster gene sequencing techniques mean mountains of new genomic information. And within all of that noise is most certainly valuable data that might help researchers in the hunt for new and better approaches to beating cancer. Yet, as data accumulates, so does statistical noise. And so useful signals become obscured.

We could say that science itself has a data problem, generally, from astrophysics to climate science to indeed medical research. The solution, according to a group of researchers led by Harvard Medical School's Peter Sorger, lies in looking outside of medicine. In a paper out today in Nature Genetics, Sorger and his team describe a mathematical toolkit drawing on the techniques of statistical physics and artificial intelligence that's capable of identifying the "social networks" of cell proteins.

Advertisement

"[The Cancer Genome Atlas] and similar projects have generated extensive data on the mutational landscape of tumors," Sorger and his team write. "To understand the functional consequences of these mutations, it is necessary to ascertain how they alter the protein-protein interaction networks involved in regulating cellular phenotypes."

Interpreting all of this data has been problematic, according to the current paper, in the absence of a unified framework that might allow for the integration of diverse measurements and the creation of solid models of cancer mutations and their effects.

In particular, Sorger's work focuses on protein interaction domains (PID), social networks of sorts that act as venues for cell signalling. When something goes wrong with these signals, cancer and other diseases potentially get a boost.

The model described in Sorger's paper is based on statistical mechanics, a way of predicting the likelihood that some system, such as a PID network, will wind up in a particular state using the thermodynamic energy of a current state. It's great for physics, but less so for complex proteins in a solution.

The new model gets around this by applying machine learning techniques to the more usual principles of statistical mechanics. From here, it's possible to predict how an individual mutation might propagate through or influence a network, even if that network happens to be vast and deeply complex. Using their model, the Harvard team was able to generate the signaling network below, for kidney tumors.

Advertisement

Image: Al Quraishi

One revelation so-far gleaned from the model has to do with rare genetic mutations. Being rare, these mutations are usually set aside (relative to common mutations), yet they appear to have a similar influence on their surrounding network.

"Both kinds of mutations are equally strong," Mohammed Al Quraishi, the paper's first author, noted in a statement. "In both cases, about one percent of the common and one percent of the rare mutations alter the tumor networks we studied. But rare mutations are being largely ignored. We need to start paying attention to them."

If anything, take this as evidence that science bleeds through. A certain category of doofus likes to deride things like deep astrophysics as inconsequential and, inevitably, ground zeros for boondoggles. There are, however, general methods and ideas that are constantly being improved and refined across disparate scientific fields. This isn't the first.

A method used to catalog distant supernovae might one day save your life.