An AI Program Maps How Proteins Pair Up Inside the Cell's Power Plants, Turning Up More Than a Thousand Possible New Partnerships
Researchers used an AI program to sift about 630,000 possible pairings among the proteins inside mitochondria, the parts of our cells that make energy. They ended up with 2,895 likely partnerships, more than half of them never reported before, which gives scientists new places to look for the causes of disease. These are mostly predictions: only some have been tested in the lab, and nothing here is a treatment.
Researchers at Texas A&M University, working with colleagues at Washington University School of Medicine in St. Louis, published a paper today in the journal Nature Communications that uses an AI program to draw a new map of the inside of human mitochondria.
Mitochondria are the tiny structures inside our cells that turn food into energy the body can use, which is why they are often called the cell's power plants. They contain about 1,100 kinds of protein. Proteins rarely work alone, so finding out which ones work together is one of the first steps to learning what each one does. This matters for health: nearly 30 percent of these proteins have been linked to human disease when the gene for them carries a harmful change. But checking every possible pairing by hand in a lab is out of reach.
So the team turned to AlphaFold Multimer, a program from Google DeepMind that predicts the shape proteins take when they come together. It was not designed as a matchmaker. The researchers used its predictions to judge which pairs were likely to truly work together, and ran it on about 630,000 possible pairs.
They ended up with 2,895 likely partnerships, more than half of which had never been reported. The list also gives first likely partners to 85 proteins whose jobs were unknown, which is a first clue to what those proteins do.
The team then tested some of the predictions in yeast and human cells. Those experiments backed up a group of proteins that work together around a molecule called coenzyme Q, and helped clear up the route copper takes to reach one of the enzymes the cell uses to make energy.
There are clear limits. The great majority of the list is still prediction, not proof, and only some of it has been checked in a lab. The paper does not offer a new treatment or test for any disease, and none of this has been tried in patients. What it offers is a starting point: a scientist studying a poorly understood protein now has a short list of partners to check first. The authors report no competing interests.