The directed message passing neural network (D-MPNN) approach of machine learning (ML) was applied to predict the stability constants of metal-ligand complexes. The ML algorithms were trained on reliable experimental data extracted from a legacy NIST database based on the multi-volume work of R. M. Smith and A. E. Martell "Critical Stability Constants". Molecular properties used in the ML process include multipole moments, Mulliken population analysis, and chemical shield tensors computed by ab-initio methods as well as the number of freely rotating bonds in the ligand that are frozen on metal chelation, MM3 strain energy, and negative Gibb's free energy of hydration obtained by the de Novo molecular design software HostDesigner software. The LOGKPREDICT program was created to interface HostDesigner with the ML program Chemprop. LOGKPREDICT creates a mutually benefical synergy between HostDesginer and Chemprop. On the one hand, HostDesginer improves the machine-learning process by providing molecular properties to Chemprop. On the other hand, Chemprop improves the ligand design by providing logK values to HostDesginer. Examples of utilizing the novel LOGPREDICT interface for the design of new ligands are presented.

Prediction of stability constants of metal-ligand complexes by machine learning and design of optimal ligands for metal ion selectivity

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.