summary: Researchers developed and trained a new deep learning model named . bind (Neural Networks of DNA Binding and Interactions). Leveraging an unprecedented dataset of 144 million sequence pairs, BINND predicts complex DNA-DNA binding affinities with superior accuracy, outperforming previous state-of-the-art models and establishing an important tool for scaling up DNA-based computing and data retrieval.
important facts
- The barrier to hyperconnectivity has broken: Traditional models treat DNA binding as a simple isolated yes/no interaction. BINND is specifically designed to handle “hyperconnect” networks, predicting how multiple different DNA strands interact simultaneously, mimicking the crowded environments of living cells and complex DNA computers.
- Avoid overestimation of empirical data: Rather than extrapolating predictions from basic biophysical or thermodynamic formulas (which have a hard time explaining nonlinear molecular behavior), the team 144 million sequence pairs We directly train BINND on empirical real-world binding events.
- High prediction accuracy: In proof-of-concept testing, the BINND deep learning model achieved the following: 83.5% accuracy It minimally outperforms current state-of-the-art models in predicting bond behavior. 10%.
- Asymmetric safety failure: When the AI model made a mistake, it demonstrated a predictable safety bias. The AI model tended to incorrectly predict that two DNA strands were present. do not have Rather than falsely claiming that a non-existent binding occurs, it binds when it actually does. This helps researchers avoid devastating background interference (crosstalk) in molecular diagnostics.
- Matrix demonstration: To demonstrate the utility of BINND, the team built an interactive database that maps the cross-linking relationships of 96 20-character DNA sequences and 26 other 20-character sequences, establishing a reliable “address book” for storing and retrieving molecular data.
- Unlocking the potential of scalable DNA computing: Storing vast archives of human data within microscopic DNA molecules requires rapid and error-free retrieval of physical data. BINND solves fundamental scaling challenges by providing a reliable roadmap of exactly which DNA strands will bind, paving the way for molecular hard drives that can store petabytes of data in a single droplet.
sauce: north carolina state university
Researchers have demonstrated a new AI model that can predict which DNA molecules will bind to which other DNA molecules. A more complete understanding of these ultra-complex binding relationships will aid applications ranging from biomedical diagnostic tools to DNA computing.
“We often think of binding as a very simple relationship, where molecule A binds to molecule B,” says Albert Kuhn, co-corresponding author of the study and associate professor of chemical and biomolecular engineering at North Carolina State University. “But in biological systems, it is never that simple: Molecule A can bind to dozens of other molecules to varying degrees.
“Capturing that hypercomplexity is a significant challenge, but it’s critical if we want to better understand natural genetic systems,” says Kuhn, a Goodnight Scholar in Innovation in Biotechnology and Biomolecular Engineering and director of the North Carolina Integrative Science Initiative’s biotechnology program.
“And capturing that hypercomplexity 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.”
“We knew that deep learning models (artificial intelligence models that can capture complex patterns) had the potential to help explore these types of hypercomplex systems,” said study co-lead author Gunabalan Brihadithwaran, a Ph.D. student at North Carolina State University. “However, we also knew that we needed a robust dataset to train our model. A model is only as good as the data you use it to train it on.”
Previous attempts to develop tools to predict DNA-DNA binding behavior have relied on relatively small datasets of DNA-DNA data and used biophysical modeling tools to predict which DNA sequences bind to which other DNA sequences. The resulting predictive tools struggle to capture the complexity of bond relationships.
“We took a different experimental approach and were able to generate substantially more data about which DNA sequences bind to each other,” says co-lead author Karishma Matanje, a Ph.D. Graduated from North Carolina State University. “Our database consists of a total of 144 million sequence pairs. This extensive dataset allows us to leverage AI models rather than extrapolating based on biophysical or biochemical principles.”
Specifically, the researchers used a large dataset to train a deep learning model to predict which DNA sequences bind to which other DNA sequences. They named this model BINND (Binding and Interaction Neural Network for DNA).
In a proof-of-concept test, the researchers found that the BINND model was able to predict which DNA pairs would join with 83.5% accuracy. And when it got it wrong, it tended to predict that the two DNA sequences wouldn’t bind, when in fact they didn’t. will do bind.
“BINND is at least 10% more accurate than state-of-the-art models,” Brihadiswaran says.
To demonstrate the utility of BINND, the researchers used the model to create a database that captures the superconnectivity of DNA-DNA binding behavior. This database is essentially a matrix, showing how 96 20-character DNA sequences will or will not combine with 26 other 20-character DNA sequences.
“This particular demonstration is actually useful from a DNA computing perspective because it provides us with important information about the characteristics of these sequences, which is critical for efforts to use DNA to capture and retrieve information,” said James Tuck, co-corresponding author of the paper and professor of electrical and computer engineering at North Carolina State University. “We hope others in the research community will take advantage of BINND, which is why we’re making it publicly available on GitHub.” The BINND repository can be found at https://github.com/dna-storage/BINND.
“One of the challenges in DNA data storage and computing is whether it can be scaled up for practical use,” Keung said. “We are optimistic that BINND will be a valuable tool to accelerate efforts to scale up these technologies, among other potential applications.”
The paper “Deep learning predicts differential DNA-DNA binding and designs hyperconnected networks” is published open access in the journal. nature communications. This paper was co-authored by Kyle Tomek and Dr. Kevin Forkel. Graduate of North Carolina State University. And Doug Townsend is currently a Ph.D. student at North Carolina State University.
Funding: This research was supported by the National Science Foundation under grants 2027655, 1901324, and 2403352. National Institutes of Health is supported by grant R41HG013877. Department of Education Graduate Support Fellowship in Areas of Need, P200A160061. and the Simons Foundation under grant 990252.
Answers to key questions:
answer: In high school biology, we learn that DNA strands are like puzzle pieces that only fit together if they are complete opposites. However, in real biological systems, DNA molecules are highly dynamic and sticky. One strand of DNA can be partially attached to dozens of other imperfect matches, dangling with varying degrees of strength. This “hypercomplexity” creates a chaotic web of interactions. If you are trying to build a biological computer, this unexpected stickiness can cause different data paths to intersect and fail, making it important to predict these partial matches in advance.
answer: DNA data storage works by converting digital ones and zeros into synthetic DNA bases A, C, T, and G. To retrieve a specific file from a molecular “hard drive,” small fluorescent DNA “probes” that act like search queries can be injected to find, bind to, and read the target data strand. If a search probe accidentally sticks to the wrong file, data will be corrupted. BINND acts like an incredibly smart safety coordinator, mapping out exactly how these sequences interact so that your search queries retrieve only the correct data without molecular crosstalk.
answer: There is a golden rule in machine learning. That said, the quality of a model is determined by the data used to train it. Previous attempts to predict DNA binding have relied on small data sets, requiring computers to make educated guesses based on common physics formulas. The North Carolina State team took a completely different approach. By generating a large-scale real-world database, 144 million sequence pairsthey gave their deep learning models a vast library of real molecular interactions to study. This large scale allows BINND to recognize complex hidden bonding patterns that cannot be simply calculated using physical equations.
Editorial note:
- This article was edited by the editors of Neuroscience News.
- Journal articles were reviewed in full text.
- Additional context added by staff.
About this AI/genetic research news
author: matt shipman
sauce: north carolina state university
contact: Matt Shipman – North Carolina State University
image: Image credited to Neuroscience News
Original research: Open access.
“Deep Learning Predicts Differential DNA-DNA Bonds and Designs Hyperconnected Networks” by Karishma Matunge, Gunavaran Brihadithwaran, Kyle J. Tomek, Kevin Forkel, Doug Townsend, James M. Tuck, and Albert J. Kuhn. nature communications
DOI:10.1038/s41467-026-75395-w
abstract
Deep learning predicts bonds between different DNA and designs hyperconnected networks
The general framework of molecular bioengineering and synthetic biology focuses on orthogonality and considers weak or nonspecific interactions as problems to be avoided. This limits the available sequence space, limits scalability, and ignores scenarios where synthetic systems must operate within the natural context of high sequence diversity. Lack of models that can accurately and quickly predict non-orthogonal interactions and validate them against ground truth data makes it difficult to exploit the entire space.
Here, we use DNA-DNA interactions as a testbed to develop BINND (DNA Binding and Interaction Neural Network). BINND combines an ultra-high throughput platform that measures millions of interactions with a deep learning model that achieves over 80% accuracy, generalizes across diverse sequences, and runs 50x faster than current models.
We demonstrate its value with a searchable DNA network of fictional story characters. BINND enables accurate predictions in diagnostics, biotechnology, and DNA origami, supporting the move toward exploiting the entire sequence space.
