Multivalent ion binding site identification with structure-based deep learning

Area of Science:

• Biochemistry
• Structural Biology
• Computational Biology

Background:

• Protein-ion interactions are crucial for cellular functions like catalysis and signaling.
• Experimental methods for mapping ion-binding sites are costly and time-consuming.

Purpose of the Study:

• To develop BiteNet_I, a structure-based deep learning model for predicting ion-binding sites on proteins.
• To identify ion-binding centers and predict binding residues for 14 biologically relevant ions.

Main Methods:

• Utilized 3D convolutional neural networks for a structure-based deep learning approach.
• Trained the model on a dataset of over 10,000 high-resolution protein-ion complexes.
• Ensured consistent annotation of binding sites by transferring ions between homologous structures.

Main Results:

• BiteNet_I demonstrates strong generalization across diverse ions within a multitask architecture.
• Achieved state-of-the-art performance on two test benchmarks, outperforming existing predictors and AlphaFold3 for protein-ion structures.
• Showcased a two- to three-fold accuracy improvement for physiologically relevant ions like Ca²⁺, Na⁺, and K⁺.

Conclusions:

• BiteNet_I offers a powerful and accurate computational tool for predicting protein-ion interactions.
• The model significantly advances the study of ion-binding sites, facilitating research in various cellular processes.
• Provides a more efficient and cost-effective alternative to experimental methods for mapping ion-binding sites.