Researchers from the Shanghai Institute of Organic Chemistry at the Chinese Academy of Sciences have developed a generative artificial intelligence model that predicts protein-protein interactions at atomic resolution, according to a study published in the Proceedings of the National Academy of Sciences (PNAS) on June 9, 2026. The model, named Void-X, takes a fundamentally different approach to protein design by building interfaces from individual atoms upward rather than sculpting an overall protein shape first.
The work, led by Drs. Yang Jing, Yuan Junying, and James J. Chou, was highlighted by the Chinese Academy of Sciences and reported by Phys.org. It adds a new tool to a fast-moving field in which AI is increasingly used to design the molecular machines that underpin medicines, materials, and synthetic biology. For more context on this story, see our ongoing latest AI developments.
Why protein-protein interactions matter
Proteins are the molecular workhorses of the human body. They build tissues, transport molecules, regulate cellular communication, and defend against infection. Many of the most important medicines — including antibody therapies for cancer and insulin for diabetes — work by interacting with specific proteins or by replacing ones that are missing or malfunctioning.
Because so many biological processes depend on how proteins bind to one another, the ability to accurately predict and engineer protein-protein interactions could accelerate drug discovery and address unmet medical needs. Progress here also complements rapid advances in delivery technologies such as adeno-associated virus vectors and mRNA lipid nanoparticles, which ferry therapeutic molecules into cells.
A bottom-up approach instead of top-down
Most existing AI frameworks for protein design follow a top-down strategy: they first generate an overall protein scaffold that fits a target site and then design a sequence to optimize binding. Void-X inverts that logic.
The model is an atomic filling system. It is trained to recognize atomic-scale interaction patterns and to fill the "voids" — the gaps — within protein interfaces. Its guiding assumption is that stable macromolecular complexes are held together by optimal atomic packing, which emerges from local interactions among neighboring atoms as well as higher-order couplings with more distant ones. Rather than designing an entire protein shape, Void-X directly generates compact atomic clusters optimized for tight packing within a specified structural region, giving the design a physically grounded foundation.
Trained on eight million atomic clusters
To train Void-X, the researchers assembled a dataset of more than eight million spherical atomic clusters drawn from experimentally determined structures deposited in the Protein Data Bank. For each cluster, roughly 30% of the peripheral, spatially contiguous atoms were masked, while the remaining atoms served as the "context" — effectively a prompt — that the model used to predict the missing atoms.
The resulting network contains 172 million parameters and achieved strong predictive accuracy: 78.3% for intra-chain atomic clusters and 68.2% for inter-chain clusters, the researchers reported.
What it means for drug discovery and beyond
According to the authors, these capabilities enable the de novo generation of atomic-resolution protein interactions, offering what they describe as a complementary and physically intuitive route for protein design. By combining atomic-level detail with generative modeling, Void-X expands the toolkit for rationally engineering biomolecular interfaces.
That matters because interface design sits at the heart of modern drug development. Better computational tools for predicting how a candidate molecule will bind to its protein target can shorten the path from idea to therapy, reduce the reliance on costly trial-and-error experiments, and make it easier to design proteins for entirely new functions in synthetic biology.
A growing field
Void-X arrives amid a surge of AI-driven protein work. From structure prediction to all-atom generative design, researchers are steadily pushing toward the ability to model biomolecular interactions at full atomic resolution. Earlier all-atom generative systems have already demonstrated the feasibility of designing and predicting interactions at that scale, and Void-X contributes a distinct, packing-first strategy to the same effort.
The masking trick that makes it work
Void-X's training method borrows a page from the language-model playbook. By masking roughly 30% of the atoms in each training cluster and asking the model to infer them from the surrounding context, the researchers turned structural biology into a fill-in-the-blanks task — the same principle that lets large language models predict missing words in a sentence. The difference is that Void-X's "vocabulary" consists of atoms arranged in three-dimensional space, and the grammar it learns reflects the physical packing rules that govern how real proteins fold and bind.
That structure-aware training is what separates Void-X from approaches that treat proteins as flat sequences of amino acids. Because the model reasons about local and long-range atomic couplings simultaneously, it can propose interface geometries that respect the physical constraints of a stable complex, rather than producing plausible-looking but biologically unstable designs.
Limits and open questions
The results are promising, but the researchers are candid that challenges remain. The inter-chain accuracy of 68.2% — relevant to the multi-protein complexes that define most real-world drug targets — is meaningfully lower than the intra-chain figure, reflecting the greater difficulty of modeling how two separate molecules come together. Designing an interface that binds tightly in a computational model is also only the first step; candidates must still be synthesized and validated experimentally, and predicted interactions can falter in the messier environment of a living cell.
Scaling up the training data, refining the model's handling of flexible or disordered protein regions, and integrating Void-X with complementary top-down systems are likely next steps. The authors frame their contribution as complementary rather than replacement technology — a physically grounded option to be combined with existing scaffolds as the field races toward reliable, end-to-end molecular design.
A broader trend
The study, published under DOI 10.1073/pnas.2607035123, signals that the competition to build ever more precise molecular design tools is now well underway — and that generative AI is becoming an indispensable part of the biochemist's workbench. As models like Void-X narrow the gap between computation and the laboratory, the hope is that researchers will be able to design new therapeutics, enzymes, and materials faster and more rationally than ever before.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
