Proteins are the molecular workhorses of the human body. They perform a vast range of essential functions, from building tissues and transporting molecules to regulating cellular communication and defending against infection. Many medicines, including antibody therapies for cancer and insulin therapy for diabetes, among many others, work by interacting with specific proteins or by replacing proteins that are missing or malfunctioning. Because proteins carry out so many critical biological tasks, the ability to predict and engineer how they interact with one another could open new possibilities for treating disease.

For this reason, accurate structural prediction and precise engineering of protein–protein interactions have great potential to accelerate therapeutic development and address unmet medical needs. Such progress would complement rapid advances in protein delivery technologies, such as adeno-associated virus and mRNA lipid nanoparticle delivery.

Now, researchers from the Shanghai Institute of Organic Chemistry of the Chinese Academy of Sciences have made an important advance in the atomic-scale prediction of protein–protein interactions by developing Void-X, a generative artificial intelligence (AI) model that designs protein interfaces from the bottom up.

The study was published in PNAS on June 9.

Void-X contrasts with most existing AI-based protein design frameworks, which follow a top-down strategy. They typically begin by generating an overall protein scaffold that fits a target site and then design protein sequences to optimize binding.

Void-X takes a fundamentally different approach by using an atomic filling model. It is trained to capture atomic-scale interaction patterns and fill atomic voids within protein interfaces, on the assumption that optimal atomic packing in a stable macromolecular complex comes from local interactions among neighboring atoms and higher-order couplings with more distant atoms. Rather than designing entire protein shapes, the model directly generates atomic clusters optimized for tight packing within specified structural regions, establishing a physically grounded foundation for protein–protein interface design.

To train the model, Drs. YANG Jing, YUAN Junying, and James J. Chou assembled a dataset of more than eight million spherical atomic clusters derived from experimentally determined structures deposited in the Protein Data Bank. For each cluster, approximately 30% of the peripheral and spatially contiguous atoms were masked while the remaining atoms served as the "context"—or prompt—that the model used to predict the missing atoms.

The resulting model contains 172 million parameters and achieved predictive accuracies of 78.3% for intra-chain atomic clusters and 68.2% for inter-chain clusters.

According to the researchers, these capabilities enable the de novo generation of atomic-resolution protein interactions, offering a complementary and physically intuitive route for protein design. By integrating atomic-level detail with generative modeling, Void-X expands the toolkit for the rational design of biomolecular interfaces, offering broad applications in drug discovery, synthetic biology, and other fields.