Scientists at North Carolina State University have developed a new artificial intelligence model that can predict which DNA molecules bind to each other with significantly greater accuracy than previous methods. The breakthrough could accelerate progress in DNA-based computing systems and genetic diagnostic tools by solving a problem that has long challenged researchers: understanding the hypercomplex web of interactions between DNA sequences.

What Makes DNA Binding Prediction So Difficult?

Most people think of molecular binding as straightforward: Molecule A binds to Molecule B. But biological reality is far messier. A single DNA molecule can bind to dozens of other sequences, each with varying degrees of strength and specificity. Capturing this complexity has been nearly impossible using traditional computational approaches.

"We often think about binding as a very simple relationship, but in biological systems, it's far from simple. Molecule A may bind to dozens of other molecules, to varying degrees," explained Albert Keung, an associate professor of chemical and biomolecular engineering at North Carolina State University and co-corresponding author of the study.

Albert Keung, Associate Professor of Chemical and Biomolecular Engineering, North Carolina State University

Previous attempts to predict DNA-DNA binding relied on small datasets and biophysical modeling, which struggled to capture the full complexity of real-world interactions. The resulting tools were often inaccurate and slow.

How Did Researchers Train a Better AI Model?

The NC State team took a different approach. Instead of relying on limited data and traditional biophysical principles, they developed a novel experimental method that generated vastly more information about DNA binding behaviors. Their dataset ultimately contained 144 million sequence pairs, providing the deep learning model with far richer training material than previous efforts.

Using this larger dataset, the researchers trained a deep learning model they named BINND, which stands for Binding and Interaction Neural Network for DNA. In testing, BINND predicted which DNA pairs would bind with 83.5% accuracy. When the model did make errors, it tended to be conservative, predicting that sequences would not bind when they actually would.

"BINND is at least 10% more accurate than the state-of-the-art model," noted Gunavaran Brihadiswaran, a co-lead author and Ph.D. student at NC State.

Gunavaran Brihadiswaran, Ph.D. Student, North Carolina State University

The model also runs dramatically faster than existing tools, completing predictions 50 times quicker than current approaches. This speed improvement matters enormously for researchers trying to design and test new DNA systems at scale.

What Are the Real-World Applications?

The researchers demonstrated BINND's utility by creating a searchable database showing how 96 different 20-character DNA sequences bind with 26 other sequences. This type of information is critical for several emerging fields:

• DNA Computing: Researchers are exploring DNA as a medium for storing and retrieving data, similar to how computers use silicon. Accurate binding predictions are essential for designing DNA systems that reliably encode and retrieve information without unwanted cross-reactions.
• Genetic Diagnostics: More sensitive diagnostic tools could be built using DNA sequences that bind specifically to disease markers or genetic variations, enabling earlier detection of health conditions.
• Synthetic Biology: Scientists designing artificial biological systems need to understand which DNA sequences will interact with each other, allowing them to build more complex and reliable engineered organisms.

"This particular demonstration has real utility from a DNA computing standpoint, as it provides us with key information about the characteristics of these sequences, which is critical for efforts to capture and retrieve information using DNA," said James Tuck, a professor of electrical and computer engineering at NC State and co-corresponding author.

James Tuck, Professor of Electrical and Computer Engineering, North Carolina State University

How to Access and Use BINND for Your Research

• Open Access: The researchers have made BINND publicly available on GitHub at https://github.com/dna-storage/BINND, allowing any researcher worldwide to use the model without licensing fees or restrictions.
• Documentation and Training: The team published their work in the journal Nature Communications with full methodological details, enabling other scientists to understand how the model works and adapt it for their own applications.
• Community Collaboration: By releasing BINND openly, the researchers hope to accelerate progress across multiple fields, from academic labs with small teams to large pharmaceutical companies developing new therapies.

Why Does This Matter for the Future of DNA Technology?

One of the biggest questions surrounding DNA computing and storage has been whether these technologies can scale up for practical, real-world use. Current DNA systems are still largely experimental, limited by our ability to design sequences that behave predictably. BINND addresses a critical bottleneck by making it faster and more accurate to predict how DNA sequences will interact.

"One of the challenges for DNA data storage and computing has been whether it can be scaled up for practical use. We're optimistic that BINND will be a valuable tool for facilitating efforts to scale up those technologies, among other potential applications," said Keung.

Albert Keung, Associate Professor of Chemical and Biomolecular Engineering, North Carolina State University

The work was supported by the National Science Foundation, the National Institutes of Health, the Department of Education, and the Simons Foundation, reflecting broad recognition that this type of foundational research has significant potential impact.

Beyond DNA computing, the success of BINND illustrates a broader trend in AI-driven biology: deep learning models trained on massive datasets are increasingly outperforming traditional computational methods that rely on hand-coded rules and biophysical principles. As researchers generate more biological data and train more sophisticated AI systems, we can expect similar breakthroughs across genomics, drug discovery, and synthetic biology in the coming years.