AI Model Maps Ion Binding Sites That Control Protein Function

Researchers at Constructor University and Constructor Labs in Bremen, Germany, have introduced BiteNetI, a specialized deep-learning architecture designed to rapidly and accurately map ion-binding sites within three-dimensional protein structures. Authored by Igor Kozlovskii and Petr Popov, the study was published in Communications Biology in 2026. The model addresses a critical bottleneck in structural biology by identifying precisely where physiologically vital ions interact with proteins, a process essential for cellular signaling, enzymatic activity, and structural stability. Unlike conventional computational predictors that often rely on amino acid sequences or focus on a single ion type, BiteNetI operates directly on 3D atomic coordinates. The architecture transforms each protein structure into a spatial grid, scanning 11 distinct atom types to evaluate electrostatic and steric environments. Trained on a curated dataset comprising over 12,000 protein complexes and more than 35,000 experimentally mapped ions, the multitask framework simultaneously predicts binding coordinates and coordinating residues for 14 biologically relevant ion species. The system processes a single structure in seconds, delivering speed improvements of approximately tenfold compared to running individual specialized models. Benchmark evaluations against established platforms, including MIonSite and IonCom, demonstrate that BiteNetI achieves two- to threefold higher accuracy than current general-purpose predictors for key physiological ions such as calcium, potassium, magnesium, phosphate, and sulfate. While AlphaFold 3 maintains a marginal advantage for carbonate and sodium prediction, BiteNetI surpasses Metal3D, a tool engineered exclusively for zinc binding. The researchers note that incorporating explicit water molecules significantly enhances prediction fidelity, as solvent networks frequently stabilize ion coordination. Performance remains robust across diverse experimental datasets, including cryo-electron microscopy reconstructions. The rapid availability of precise ion-binding annotations positions BiteNetI as a valuable asset for computational drug discovery and protein engineering. Disrupted ion-protein interactions are frequently implicated in neurological, cardiovascular, and metabolic diseases, making atomic-level mapping crucial for targeted therapeutic intervention. Popov emphasized that task-specific architectures like BiteNetI excel at detailed annotation, complementing broader structural prediction tools that focus on global folding. The team has deployed the model as an open-access web application, enabling immediate integration into academic and industrial workflows. Looking ahead, the research group aims to unify binding site prediction by developing a single framework capable of identifying all molecular ligands, small molecules, and ions within a protein environment. This progression mirrors broader shifts in artificial intelligence toward consolidated prediction engines. By democratizing access to high-fidelity ion mapping, BiteNetI establishes a new baseline for accelerating functional genomics and precision drug design.