AI Model Predicts How DNA Sequences Bind to Each Other

Researchers from North Carolina State University have demonstrated a new AI model that predicts which DNA molecules bind with which other DNA molecules, offering a more thorough understanding of these hypercomplex binding relationships. The work has applications ranging from biomedical diagnostic tools to DNA computing.

“We often think about binding as a very simple relationship—Molecule A binds to Molecule B,” says Albert Keung, co-corresponding author of the study published in Nature Communications. “But in biological systems, it’s far from simple. Molecule A may bind to dozens of other molecules, to varying degrees. Capturing that hypercomplexity is a significant challenge, but it is critical if we want to better understand natural genetic systems,” says Keung. “And capturing that hypercomplexity is also critical if we want to develop tools that make full use of biomolecules, such as diagnostic tools that are sensitive to genetic differences or DNA computing systems that rely on DNA to store and retrieve data.”

Gunavaran Brihadiswaran, co-lead author of the paper, said the team knew deep learning models had potential to explore this hypercomplex system, but needed a robust dataset to train the model on. Previous attempts to predict DNA-DNA binding behaviors relied on relatively small datasets and biophysical modeling tools, producing predictive tools that struggled to capture the complexity of binding relationships.

Karishma Matange, co-lead author of the paper, added that the team took a different experimental approach that generated substantially more data on which DNA sequences bind to each other, resulting in a database of 144 million sequence pairs. This larger dataset allowed the team to train a deep learning model, named BINND (Binding and Interaction Neural Network for DNA), rather than extrapolating from biophysical or biochemical principles.

In proof-of-concept testing, BINND predicted which DNA pairs would bind with 83.5% accuracy, and Brihadiswaran said it is at least 10% more accurate than the state-of-the-art model. When it erred, it tended to predict that two sequences would not bind when they actually would.

To demonstrate BINND’s utility, the researchers used it to produce a database showing how 96 20-character DNA sequences bind, or don’t, with 26 other 20-character DNA sequences. James Tuck, co-corresponding author, said this demonstration provides key information about sequence characteristics important for capturing and retrieving information using DNA, and that BINND is being made publicly available on GitHub for the research community to use.