| Jun 16, 2026 |
Piecing the puzzle of how proteins fit together
A team of researchers say that the current computational methods for predicting how proteins will bind are inadequate. They offer a simpler, and more accurate, way.
|
|
(Nanowerk News) How the proteins in our bodies bind together to form protein complexes plays a critical role in numerous cell functions - staving off diseases, for instance, or transporting ions across cell membranes. A better understanding of how they bind could lead to new medicines and possibly the design of new protein complexes.
|
|
But there are around 10,000 distinct types of proteins in the human body, and we know only a small fraction of the millions of possible ways for them to bind.
|
|
When two distinct proteins (also known as “monomers”) bind together to form what’s known as a “heterodimer,” they typically fit together like puzzle pieces. But not all monomers fit together. As for predicting which monomers will fit each other, that’s the challenge.
|
|
“We only have high-resolution structures for a couple thousand proteins, so we have little to no structural information on a large majority of proteins,” said Prof. Corey O’Hern, who led the study (Physical Review E, "Assessment of scoring functions for computational models of protein-protein interfaces"). “So we have tons of proteins in our bodies, but we don't know what complexes they form, what cell functions they carry out, or if they are implicated in disease.”
|
|
Because it is too time-intensive and costly to test the potential pairs one-by-one experimentally in a lab, researchers use computer models to predict which ones will bind.
|
|
“We need to build computational methodologies to accurately assess whether two proteins bind and where so we can try to understand their function and develop therapeutics,” said O’Hern, professor of mechanical engineering.
|
|
The problem, though, is that the computational methods currently in use to assess the accuracy of the models are significantly flawed.
|
|
“When the previous literature assesses current methods for scoring computational models, they claim that they're accurate, but when we do rigorous tests, we find that they’re not,” he said. These scoring functions factor in many physical features and can perform well on classification tasks, which categorize the models, but poorly on regression tests that assess the actual quality of the model more accurately.
|
|
The researchers developed their own computational scoring function, a simple support-vector regression (SVR) model with only two physical features—the size of the interface between the two monomers and how well-intertwined the proteins are at the interface. They compared its accuracy to seven state-of-the-art scoring functions, which are commonly used in the field. They found that their scoring function did at least as well, and often outperformed the others.
|
|
“In the current study, we focused on the simple problem of identifying the binding interface between two rigid protein pairs,” he said. “In the future, we will develop methods for identifying binding pairs when we do not know the bound form of the monomers.”
|
