SpheronizaTor: Spherical Voxelization for Interpretable Protein Microenvironment Modeling.

Area of Science:

• Structural biology
• Artificial intelligence
• Computational chemistry

Background:

• Current AI models for structural biology depend on effective protein environment encoding.
• Existing volumetric representations using Cartesian voxel grids face challenges with rotational invariance, residue specificity, and biochemical detail.

Purpose of the Study:

• Introduce SpheronizaTor, a novel residue-centered voxelization framework for protein structures.
• Address limitations in current volumetric representations for AI-driven structural biology.

Main Methods:

• Developed a residue-centered voxelization framework, SpheronizaTor.
• Built local spherical voxel maps for each residue, encoding atom types, covalent bonding, and residue relationships.
• Combined spherical alignment with chemically explicit feature channels.

Main Results:

• SpheronizaTor produces residue-specific voxel representations, capturing protein granularity.
• The framework enhances interpretability and compatibility with 3D neural network architectures.
• Demonstrated a chemically realistic and computationally efficient voxelization strategy.

Conclusions:

• SpheronizaTor offers a versatile foundation for downstream tasks like mutation effect prediction and binding site analysis.
• The residue-centric approach bridges global volumetric and graph-based models in structural biology.
• This framework advances AI applications in understanding protein structure and function.