A Comprehensive Machine Learning Model for Metal-Ligand Binding Prediction: Applications in Chemistry and Biology.

A machine-learning (ML) model that predicts metal-ligand binding constants was developed using the open-source Chemprop software. The model was trained on over 30,000 experimental log K1 values, which include both protonation and metal-ligand stability constants, comprising over 3500 ligands and 102 metal ions from 73 total elements, thus generalizing beyond existing limited approaches, which focus only on specific metals or ligand families. The best-performing model included a combination of SMILES-based molecular representations along with descriptors for the metal ion and experimental conditions. It had an external test R2 value of 0.942, and MAE value of 0.834. A "SMILES-only" simpler version also produced accurate predictions and preserved the binding trends, serving as a quick and easily accessible alternative for users without computational expertise. The SMILES-only model performed comparably to density functional theory (DFT) calculations but utilized a fraction of the computational resources. The model was successfully applied across diverse domains, including bioinorganic chemistry, heavy metal remediation, and sensor development and demonstrated its effectiveness as a rapid and reliable screening tool for both academic and industrial uses.