Researchers have created an artificial intelligence model that forecasts which DNA molecules connect with others. Better insight into these intricate binding patterns supports uses in medical diagnostics and DNA-based computing systems.
Binding relationships in biology are rarely simple, as one molecule can attach to many others at different strengths. Modeling this complexity aids understanding of genetic systems and development of advanced tools.
Deep learning approaches can identify such patterns, yet they require large training datasets. Earlier efforts used smaller collections of DNA binding data combined with biophysical models, limiting accuracy.
The team generated a much larger dataset containing 144 million sequence pairs through new experiments. This resource enabled training of a deep learning system called BINND, which predicts binding between DNA sequences.
Tests showed BINND achieved 83.5 percent accuracy and outperformed prior models by at least 10 percent. Errors usually involved predicting no binding when connections actually occurred.
The model produced a matrix illustrating binding among 96 sequences of 20 characters each with 26 additional sequences. This output supports DNA computing by revealing sequence properties needed for data storage and retrieval.
BINND is available publicly on GitHub. Developers expect it to assist scaling of DNA technologies for practical applications in storage and other fields.
The study appears in Nature Communications.
