New AI model maps the ion binding sites that control how proteins work

Researchers at Constructor University and Constructor Labs have created BiteNetI, an advanced AI model that swiftly identifies the binding sites of 14 crucial ions within protein structures. This innovative model, described in a study published in Communications Biology, is twice to three times more accurate than current predictors, including Google DeepMind's AlphaFold 3.

BiteNetI's efficiency stems from its use of deep learning, enabling it to process entire protein structures in mere seconds. The tool's open-access nature could significantly accelerate drug development and deepen our understanding of protein function. Proteins, the cellular workhorses, often rely on ions to stabilize their structure, drive reactions, and transmit signals.

Malfunctions in these ion interactions can lead to severe neurological, cardiovascular, or metabolic disorders. Experimental determination of ion-binding sites typically requires time-consuming and expensive high-resolution X-ray crystallography or spectroscopy. Current computational alternatives are limited in their scope and scalability.

BiteNetI overcomes these limitations by employing physics-based AI, converting proteins into a three-dimensional grid and scanning their complex 3D shape across 11 different atom types, including the 14 ion types it predicts. This multitask approach allows BiteNetI to simultaneously determine both the exact coordinates and coordinating residues for these ions in less than a minute per structure.

Comparatively, while AlphaFold 3 is a powerful generalist model that can also predict ion positions, BiteNetI specifically targets ion binding, delivering superior accuracy, especially for key ions like calcium, potassium, magnesium, phosphate, and sulfate. The model's sensitivity allows it to detect subtle genetic changes that could affect ion bonds.

When water molecules are included in the analysis, BiteNetI's precision further improves, even when the protein structure was derived from different imaging techniques, such as cryo-electron microscopy. Looking forward, the researchers aim to extend this concept to a single all-encompassing model that can identify and annotate all types of binding sites, from ions to peptides and small molecules, streamlining the process for researchers working on varied molecular interactions.