Foldseek-Interface reveals a protein interface universe far from complete

Life inside a cell depends on a constant, intricate conversation between proteins. These molecules are the workers of biology, and to get anything done, they must find one another and lock together. Sometimes they form rigid, permanent structures, like the beams of a building. Other times, they touch briefly to send a signal and then drift apart, like a quick handshake. The specific shape where two proteins touch is called an interface. It is the unique geometry of that meeting point that determines whether the interaction happens, how strong it is, and what the result will be. For decades, scientists have mapped the shapes of individual proteins, cataloging the vast library of molecular forms that exist. But while we know what the individual pieces look like, we have lacked a systematic way to map the billions of different ways those pieces can fit together. Without a map of these meeting places, it is difficult to understand the full scope of how cells function, how they evolve, or how diseases might hijack these connections.

A team of researchers has now created that missing map. They developed a new method called Foldseek-Interface, which treats the three-dimensional shapes of protein interfaces as if they were simple strings of text. This clever conversion allows computers to search and compare millions of these meeting points in a fraction of the time it would take with older tools. By applying this method to every protein complex currently known to science, the team discovered that the universe of protein interfaces is far larger and more diverse than previously thought. They found that while we have been busy cataloging the shapes of individual proteins, the ways those proteins connect are still largely unexplored territory.

The researchers began by taking the entire library of experimentally determined protein structures, known as the Protein Data Bank, which contains over three million pairs of interacting proteins. Using their new tool, they extracted the specific atoms where these proteins touch and converted those shapes into a searchable format. This allowed them to group similar interfaces together, regardless of whether the proteins themselves looked different. The result was a massive reduction in redundancy: the three million pairs were organized into 77,167 distinct clusters. Each cluster represents a unique type of meeting place, a recurring topological family that nature has used over and over again. This resource is now freely available to scientists, offering a way to search for similar interfaces across the entire known universe of protein structures.

One of the most striking findings was that our current view of protein interactions is heavily skewed. The database is dominated by stable, rigid interfaces where both proteins are fully folded and ordered. However, the researchers suspect this is an artifact of how science has been done, not a reflection of reality. In living cells, many interactions involve proteins that are floppy or disordered, touching briefly to regulate signals. These flexible interactions are much harder to capture in a lab, so they are underrepresented in the data. The study suggests that the true landscape of protein interfaces is likely much richer in these disordered, transient connections than our current maps show.

The team also used this new map to look at how these interfaces have evolved. They found that nature often arrives at the same solution from different starting points. In some cases, proteins with completely different overall shapes evolved to use the exact same interface geometry to bind to a partner. This phenomenon, known as convergent evolution, suggests that certain shapes are simply the most efficient way to solve a binding problem. The researchers also identified cases where pathogens, such as bacteria and viruses, have evolved to mimic the interfaces of human proteins. By copying these shapes, the invaders can trick human cells into interacting with them, effectively hijacking the cell's machinery. The new tool allowed the team to spot these mimicry events with high precision, revealing how viruses often use short, flexible segments to copy human interfaces, while bacteria tend to use more rigid, folded structures.

Finally, the researchers tested whether this method could help make sense of the flood of predicted protein structures generated by artificial intelligence. They searched through a database of predicted human protein complexes to see if the AI had invented new types of interfaces that do not yet exist in the experimental record. They found thousands of predicted interfaces that had no match in the known database. While many of these were likely errors or artifacts of the prediction process, a careful review revealed several promising candidates. These included potential new interactions involved in ribosome function, cell signaling, and the transport of materials within the cell. These findings suggest that artificial intelligence is not just confirming what we already know, but is actively pointing toward a hidden layer of protein interactions that experimental science has yet to discover.

The work provides a new lens through which to view the molecular machinery of life. By shifting the focus from the whole protein to the specific point of contact, the researchers have opened a window into the diversity of biological interaction. The map they have built is not complete; the authors note that the experimental data is still biased toward stable structures, and the true extent of the interface universe remains to be fully charted. However, the tool they created offers a powerful way to navigate this space, allowing scientists to find patterns in the chaos of molecular binding. As more predicted structures become available, this method will likely reveal even more of the hidden connections that drive the complexity of living systems.