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

The data used to train and validate the CSD-NN and PDB-NN models. These neural networks built in Scikit-learn and are trained to recognize the geometry and coordination number of a metal site from the geometric features including distances and angles of the six atoms closest to the metal. The sites are stored as .pdb files composed of all atoms falling within 10 angstroms of the central metal atom and they include both metal complexes deposited in the Cambridge Structural Database (CSD, Version 5.42-5.43) and bioinorganic sites deposited in the Protein Data Bank (PDB), retrieved through the MetalPDB interface (up to the end of 2018). The sites are divided in seven geometry classes: linear (LIN), trigonal planar (TRI), tetrahedral (TET), square planar (SPL), square pyramidal (SQP), trigonal bipyramidal (TBP) and octahedral (OCT).

The number of metal sites in each file is the following:

CSD_dataset_pdbs.zip - 110K 

CSD_validation_dataset_pdbs.zip - 1369

PDB_dataset_pdbs.zip file - 2960

PDB_validation_dataset_pdbs.zip - 106

The file images_and_data_analysis.zip contains all data and code required to replicate the figures shown in the paper.

The benchmark_fp.zip file contains code for the benchmark of Metalhawk and its comparison to Findgeo, another tool for coordination geometry classification.