Machine learning-assisted substrate binding pocket engineering based on structural information.

Area of Science:

• Computational biology
• Biochemistry
• Machine learning

Background:

• Enzyme engineering relies on modifying enzyme-substrate binding pockets for altered catalytic activity.
• Distinct substrate binding sites pose a challenge for traditional engineering approaches.

Purpose of the Study:

• To develop a novel 3D convolutional neural network (DUnet) for accurate prediction of protein-ligand binding sites.
• To demonstrate the utility of DUnet in guiding enzyme engineering strategies.

Main Methods:

• A 3D convolutional neural network integrating DenseNet, UNet, and self-attention was developed.
• Data augmentation techniques were employed to enlarge the training dataset.
• The model was validated on SC6K, COACH420, and BU48 datasets, and applied to predict binding sites in Klebsiella variicola acid phosphatase (KvAP) and Bacillus anthracis proline 4-hydroxylase (BaP4H).

Main Results:

• DUnet achieved high prediction accuracy, with distances between predicted and real binding site centers ≤4 Å.
• Successful prediction of critical binding sites for KvAP (53.8%) and BaP4H (56%) was achieved, impacting catalysis.
• Virtual saturation mutagenesis based on predicted sites identified mutations enhancing enzyme-substrate binding.

Conclusions:

• Accurate prediction of key binding sites is essential for successful enzyme engineering.
• DUnet offers a powerful tool for identifying critical residues and improving enzyme function through targeted mutagenesis.