MetalHawk: Enhanced Classification of Metal Coordination Geometries by Artificial Neural Networks

Area of Science:

• Inorganic Chemistry
• Computational Chemistry
• Structural Biology

Background:

• Metal complex properties depend on ligand coordination number and geometry.
• Current methods for determining these properties involve a trade-off between accuracy and computational cost.
• This limitation hinders the analysis of large structural datasets.

Purpose of the Study:

• To develop a machine learning-based approach for simultaneous classification of metal site coordination number and geometry.
• To overcome the limitations of existing methods in terms of accuracy and computational cost.
• To provide a tool for efficient analysis of large metal complex structural data.

Main Methods:

• Development of MetalHawk, a machine learning tool utilizing artificial neural networks (ANNs).
• Training ANNs using data from the Cambridge Structural Database (CSD) and Metal Protein Data Bank (MetalPDB).
• Validation of the model on CSD-deposited metal sites and bioinorganic metal sites from MetalPDB.

Main Results:

• The CSD-trained model achieved 96.51% balanced accuracy for classifying CSD metal sites.
• The model demonstrated 84.29% balanced accuracy on the PDB dataset and 91.66% on a manually reviewed PDB validation set.
• Output vectors from the CSD-trained model serve as a proxy for metal-site distortions, representing subtle geometrical features.

Conclusions:

• MetalHawk provides an accurate and computationally efficient method for classifying metal site coordination number and geometry.
• The tool is applicable to both synthetic metal complexes and bioinorganic sites.
• MetalHawk's output offers insights into metal-site distortions and geometrical nuances.