A new AI model can accurately predict how various proteins interact with DNA.
Researchers at USC have developed an innovative artificial intelligence model, featured in Nature Methods, that can reliably forecast how diverse proteins bind to DNA. This breakthrough technology is expected to shorten the time needed for the development of new drugs and medical treatments.
The tool, named Deep Predictor of Binding Specificity (DeepPBS), utilizes a geometric deep learning framework to determine protein-DNA binding specificity based on the structures of protein-DNA complexes. DeepPBS enables scientists to upload the structural data of these complexes into a web-based computational tool.
“Protein-DNA complex structures typically involve proteins attached to a single DNA sequence. To better understand gene regulation, it’s crucial to know how a protein can bind to various DNA sequences or genomic regions,” explained Remo Rohs, professor and founding chair of the department of Quantitative and Computational Biology at USC Dornsife College of Letters, Arts and Sciences. “DeepPBS serves as an AI solution that eliminates the requirement for high-throughput sequencing or structural biology experiments to determine protein-DNA binding specificity.”
AI assesses and forecasts protein–DNA interactions
DeepPBS uses a geometric deep learning model, a machine-learning technique that interprets data through geometric shapes. This AI tool is crafted to grasp the chemical features and geometric arrangements of protein-DNA interactions, allowing for accurate predictions of binding specificity.
By processing this information, DeepPBS creates spatial graphs that represent protein structures and the connections between protein and DNA models. Unlike many current methods that focus on specific protein families, DeepPBS can predict binding specificity across multiple protein families.
“Having a universal method available for researchers that isn’t confined to a particular well-studied protein family is essential. This capability also enables us to design new proteins,” Rohs stated.
Significant progress in predicting protein structures
The area of protein-structure prediction has made remarkable strides since the introduction of DeepMind’s AlphaFold, which can identify protein structures from sequences. These advancements have expanded the amount of structural data for scientists and researchers to analyze. DeepPBS complements structure prediction techniques to forecast specificity for proteins that lack experimental structures.
Rohs highlighted the vast potential applications of DeepPBS. This pioneering research approach could accelerate the creation of new drugs and treatments targeting specific mutations in cancer cells, as well as facilitate breakthroughs in synthetic biology and RNA research.
About the study: Along with Rohs, the research team includes Raktim Mitra, Jinsen Li, Yibei Jiang, Ari Cohen, and Tsu-Pei Chiu from USC; Jared Sagendorf from the University of California, San Francisco; and Cameron Glasscock from the University of Washington.
This study was primarily funded by NIH grant R35GM130376.