This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: Researchers at Constructor University and Constructor Labs have developed BiteNetI, a deep-learning model that locates the binding sites of 14 biologically important ion types directly in three-dimensional protein structures. The model needs only several seconds per structure and reaches two- to threefold higher accuracy than most existing predictors, including Google DeepMind's AlphaFold 3.
The study, authored by Igor Kozlovskii and Petr Popov, has been published in the journal Communications Biology. The tool provides an open-access platform that could help accelerate drug discovery and the understanding of protein function. Proteins are the molecular workhorses of human cells, but they rarely act alone.
They rely on ions—electrically charged atoms or small charged groups like calcium, sodium and potassium—to stabilize their structure, trigger chemical reactions and transmit signals. When these ion interactions malfunction, they can lead to severe neurological, cardiovascular or metabolic disorders. Knowing exactly where an ion-binding site is located at atomic resolution is valuable for several reasons: Determining such sites experimentally, however, requires high-resolution X-ray crystallography or spectroscopy, which is laborious and expensive.
Existing computational alternatives are typically limited to a single ion type, operate on the sequence rather than the structure, or do not scale to large numbers of proteins. The physics-based AI model BiteNetI addresses this bottleneck by bringing the image-recognition technology found in modern smartphone cameras into molecular biology. The system translates a protein into a three-dimensional grid, scanning its complex 3D shape like an animated video across 11 different atom types.
The model learned by studying a curated digital library of more than 12,000 protein complexes containing more than 35,000 precisely mapped bound ions. Through this multitask design, BiteNetI simultaneously predicts both the exact coordinates and coordinating residues for 14 different ion types in several seconds—making it roughly 10 times faster than running individual specialized models per ion type without sacrificing accuracy. Generalized AI models like Google DeepMind's AlphaFold 3 have recently made major advances in predicting overall protein shapes, and they can also place ions during their predictions.
Benchmark results showed that the specialized BiteNetI model achieved a two- to threefold accuracy improvement over existing methods for vital physiological ions, delivering superior scores for calcium, potassium, magnesium, phosphate and sulfate, while AlphaFold 3 held a slight edge for carbonate and sodium. BiteNetI's sensitivity allows it to recognize intricate atomic patterns and evaluate how tiny genetic changes, like point mutations, might weaken or restore necessary ion bonds. "AlphaFold 3 and BiteNetI answer different questions, and the results should be read in that light," said Popov.
Extract — continue reading at the source.
