ProDCoNN

ProDCoNN predicts amino-acid sequences that fold into specified three-dimensional (3D) protein structures using deep convolutional neural networks.

Key Features:

• Deep Learning Framework: A nine-layer 3D deep convolutional neural network (CNN) that predicts residue type from the local 3D structural environment surrounding the Cα atom.
• Multi-Scale Structural Analysis: CNN layers capture structural information at multiple scales, including bond lengths, bond angles, torsion angles, and secondary structure.
• Input Representation: Processes gridded boxes of atomic coordinates and atom types around each residue to represent the local structural context for prediction.

Scientific Applications:

• Protein sequence design: Predicts amino-acid sequences compatible with target 3D structures.
• Protein engineering: Enables design of proteins with specified structural constraints to achieve desired functions or properties.
• Drug design: Facilitates generation of protein sequences for targets or scaffolds relevant to therapeutic development.
• Enzyme optimization and synthetic biology: Supports design of enzyme sequences and engineered proteins for synthetic biology applications.

Methodology:

Training a nine-layer 3D CNN on a large dataset of protein structures using gridded-box inputs of atomic coordinates and atom types; the model predicts residue type from the local 3D environment around the Cα with CNN layers configured to capture bond lengths, bond angles, torsion angles, and secondary structure, and performance validated on large test proteins and benchmark datasets.