Pictured: Side-by-side comparison between a healthy bone (left) and a bone with osteoporosis (right).
A simple fall can be life-changing, especially if bones are not as strong as they used to be. For millions of people with osteoporosis, even minor accidents can lead to serious fractures and long recovery times.
Osteoporosis is as common as heart disease and diabetes, but many cases go undiagnosed. The condition weakens bones, making them fragile and easier to break. In the United States, osteoporosis affects about 1 in 10 people over the age of 50.
Treatments can help strengthen bones and ease pain, but some cause serious side effects. For example, one commonly used protein can lead to abnormal tissue growth.
Researchers are looking for safer options, and this past March, Biomedical Engineering Professor Esmaiel Jabbari and Department of Computer Science and Engineering Chair Homayoun Valafar launched a three-year research project funded by the National Science Foundation. Their goal is to develop new treatments that encourage bone growth without the harmful side effects by combining biology with machine learning, a subset of artificial intelligence.
For osteoporosis patients, Bone morphogenetic protein-2 (BMP-2) enhances the healing of damaged bones. But Jabbari’s team is focusing on peptides—tiny pieces of proteins that act like building blocks. These peptides may promote bone healing in the same way as current treatments, but more safely. The project will also utilize machine learning to connect a peptide’s amino acids with its ability to stimulate bone formation.
“BMP-2 encourages cells to produce more bone that will reduce the chance of fracture,” Jabbari says. “But those same proteins are involved in fetal development, which is why they can cause unwanted side effects.”
A key part of the project is understanding how a peptide’s shape influences its behavior. Even small changes in its structure can affect how well it encourages bone growth. Jabbari’s lab previously worked on developing shorter peptide sequences of BMP-2 to create bone formation without the side effects. When testing to determine if bone tissue could be formed, he observed different shapes of peptides compared to the same sequence in BMP-2.
“This led us to consider that if we started using machine learning by changing some of the sequences, could we see how the shape is changed? And could we get closer to the actual shape of the peptide on the protein in the same way?” Jabbari says.
Jabbari’s team developed a simulation method to predict the shape of different sequences. Using those shapes, machine learning was incorporated to search for the sequence that was closest to the actual shape of the peptide. Jabbari will now test 10 new peptides to determine whether they form more bone tissue.
"Machine learning will be integrated to build models and replace labor-intensive experimental screening, enabling rapid identification of candidate peptides with improved biological performance,” Valafar says. “Ultimately, these discoveries will support the development of targeted, controlled-delivery biomaterials that enhance tissue regeneration and improve patient treatment outcomes."
Jabbari plans to test the different peptide sequences with mesenchymal stem cells, which can differentiate into bone to aid in tissue repair. Bone cells come from mesenchymal stem cells, which migrate to the tissue and form bone. The tests will determine whether their peptides can transform mesenchymal stem cells into bone cells.
The testing will be performed using two-dimensional tissue culture plates, which are specialized vessels used for growing and maintaining cells in a controlled environment. It will then be switched to three-dimensional testing, which is more complex but mirrors the shapes of tissues in the body.
“We’ll start simple and build from there,” Jabbari says. “Ultimately, we want to find the best option and see how it performs in a living system.”