AI predicts super-complex DNA bonds, discovered by North Carolina state

Researchers at North Carolina State University have developed an artificial intelligence model that can predict which DNA molecules will bind to each other with 83.5% accuracy. This will greatly improve our understanding of the complex relationships within genetic systems. The new model, called BINND: Binding and Interaction Neural Network for DNA, was trained on a dataset of 144 million sequence pairs and was able to capture a level of complexity previously unattainable with biophysical modeling. “We often think of binding as a simple relationship where molecule A binds to molecule B,” says Albert Kuhn, co-corresponding author of the study and a Goodnight Fellow in Innovation in Biotechnology and Biomolecular Engineering. “But in biological systems, it’s not that simple.” According to the researchers, this advance has the potential for applications ranging from more sensitive biomedical diagnostic tools to new DNA computing systems.

Deep learning model BINND predicts DNA-DNA binding accuracy

BINND achieved 83.5% accuracy in predicting DNA pair binding, representing a significant improvement over previous prediction tools. These tools often relied on small datasets and biophysical modeling that failed to capture the full range of binding relationships and struggled with the inherent complexity of DNA interactions. “We took a different experimental approach and were able to generate substantially more data about which DNA sequences bind to each other,” explains Dr. Karishma Matanje. A graduate of North Carolina State University, he is co-lead author of a paper published in Nature Communications. This expanded dataset allows the team to go beyond extrapolation and leverage the pattern recognition capabilities of deep learning. This model tends to err on the side of predicting that a join is absent, rather than incorrectly predicting a join when it is not present, a notable feature of its performance. “BINND is at least 10% more accurate than the best existing models,” said co-lead author Gunabalan Brihadiswaran, Ph.D.

Beyond improving accuracy, the researchers used BINND to create a database of superconnectivity of DNA-DNA bonds that mapped interactions between 96 different 20-letter DNA sequences and 26 other DNA sequences. Molecule A can combine to varying degrees with dozens of other molecules, Keung added, noting that this database is particularly promising for advances in DNA computing, providing important information for using DNA to capture and retrieve data.

But in biological systems, it’s never that simple. Molecule A can bind to dozens of other molecules to varying degrees.

AI training with 144 million sequence pair datasets

The quest to understand DNA interactions has entered a new phase, moving beyond limited data sets and increasingly relying on artificial intelligence to decipher the complex relationships between sequences. Previous predictive models often suffer from the inherent complexity of biological systems, estimating from small datasets or biophysical modeling. However, researchers are now leveraging vastly expanded resources to train more accurate AI. The North Carolina State University team created a database of 144 million sequence pairs. This has significantly increased the amount of data available for analysis. The move toward AI-driven prediction is driven by the recognition that DNA binding is not a simple one-to-one correspondence. We achieved 83.5% accuracy, and the model tended to predict no joins rather than false positives, especially when errors occurred.

This enhanced predictive capability has implications beyond basic biological understanding and could impact fields such as biomedical diagnostics and DNA computing. “Capturing that complexity is also important if we want to develop tools that take full advantage of biomolecules, such as diagnostic tools that are sensitive to genetic differences or DNA computing systems that rely on DNA for data storage and retrieval,” Keung said, highlighting the research’s wide applicability. The team has published BINND on GitHub in hopes of spurring further innovation within the scientific community.

We often think of bonding as a very simple relationship where molecule A binds to molecule B.

Albert Kuhn, co-corresponding author of this study and associate professor of chemical and biomolecular engineering at North Carolina State University