As reported in Nature Communications on July 14, 2026, researchers at North Carolina State University have developed a deep learning model that predicts which DNA sequences will bind to each other with 83.5% accuracy. Trained on 144 million sequence pairs, the model captures the web of molecular interactions that earlier tools missed, offering a path to more reliable DNA data storage and biomedical diagnostics.
"We often think about binding as a very simple relationship-Molecule A binds to Molecule B," said Albert Keung, co-corresponding author and associate professor of chemical and biomolecular engineering at NC State. "But in biological systems, it's far from simple. Molecule A may bind to dozens of other molecules, to varying degrees." Existing predictive methods relied on small datasets and biophysical rules, which struggled to represent this hypercomplexity.
Dataset scale unlocks deep learning
Previous attempts to model DNA-DNA binding used biophysical modeling on limited data. The NC State team took a different experimental approach, generating a database of 144 million sequence pairs. "We knew that deep learning models-artificial intelligence models capable of capturing complex patterns-had the potential to help us explore this type of hypercomplex system," said Gunavaran Brihadiswaran, co-lead author and Ph.D. student. "However, we also knew that we would need a large dataset in order to train the model. A model is only as good as the data you train it on."
"Altogether, our database consists of 144 million sequence pairs," said Karishma Matange, co-lead author and Ph.D. graduate. "This broader dataset allowed us to make use of AI models rather than extrapolating based on biophysical or biochemical principles."
BINND outperforms prior models
The researchers trained a deep learning model they named BINND (Binding and Interaction Neural Network for DNA). In proof-of-concept tests, BINND predicted which DNA pairs would bind with 83.5% accuracy. When it erred, it tended to predict no binding for sequences that actually did bind. "BINND is at least 10% more accurate than the state-of-the-art model," said Brihadiswaran.
Applying the model to DNA computing
To demonstrate utility, the team used BINND to produce a matrix showing how 96 DNA sequences of 20 characters each interact with 26 other such sequences. This hyperconnected network is valuable for DNA data storage and retrieval. "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, co-corresponding author and professor of electrical and computer engineering at NC State. Keung added, "We're optimistic that BINND will be a valuable tool for facilitating efforts to scale up those technologies."
Public access to the model
The researchers posted the BINND code to GitHub, making it freely available. The repository is at https://github.com/dna-storage/BINND.
Researchers interested in applying AI to biochemical problems can explore the AI Learning Path for Biochemists. For broader coverage of machine learning in the laboratory, see our AI for Science & Research resources.
Why this matters for Science and Research
For biochemists and molecular biologists, BINND offers a practical tool to design DNA interactions with higher precision, which is essential for building DNA-based data storage systems that can store and retrieve information reliably. The work also demonstrates that pairing large-scale experimental data generation with deep learning can overcome long-standing modeling challenges in biomolecular binding. Researchers in adjacent fields-such as RNA-protein interactions or drug-target binding-may adopt similar data-driven strategies to improve their own predictive models.
Your membership also unlocks:
