Directional Multiobjective Optimization of Metal Complexes at the Billion-Scale with the tmQMg-L Dataset and PL-MOGA Algorithm

25 September 2023, Version 2
This content is a preprint and has not undergone peer review at the time of posting.


Transition metal complexes (TMCs) play a key role in several areas of high interest, including medicinal chemistry, renewable energies, and nanoporous materials. The development of TMCs enabling these technologies remains challenged by the need to optimize multiple properties within very large chemical spaces, in which the thirty transition metals can be combined with a virtually infinite number of ligands. In this work, we provide the open tmQMg-L dataset including 30K TMC ligands, which combines large chemical diversity with synthesizability. The charge and metal-coordination mode of the ligands were robustly defined with a novel algorithm based on graph and natural bond orbital theories. The tmQMg-L dataset was leveraged in the automated generation of 1.37M TMCs resulting from all possible combinations between a square planar palladium(II) scaffold and a pool of 50 different ligands. This TMC space was used to benchmark a multiobjective genetic algorithm (MOGA) that optimized two properties over a Pareto front; namely the polarizability (alpha) and the HOMO-LUMO gap (epsilon). The MOGA evolved 130 TMC hits with maximal (alpha, epsilon) values in a way that could be easily rationalized by analyzing the nature of the ligands selected. Instead of the traditional mutation and crossover of fragments within a single ligand, this MOGA implemented full-ligand genetic operations acting on all coordination sites, maximizing chemical diversity. Further, we extended this MOGA algorithm with the Pareto-Lighthouse functionality (PL-MOGA), which allows for controlling both the aim and scope of the multiobjective optimization over the Pareto front. In explicit spaces containing billions of TMCs, the PL-MOGA enabled the explainable generation of thousands of novel and highly diverse TMC hits. We believe that the combined use of the tmQMg-L dataset and PL-MOGA algorithm will facilitate the discovery of TMCs with optimal properties within untapped chemical spaces.


evolutionary learning
multiobjective optimization
Pareto front
ligand dataset
ligand charge
genetic algorithms
chemical space
transition metal complex
chemical diversity
directional optimization
explainable AI
metal coordination
machine learning

Supplementary materials

Supporting Information
The Supporting Information provides further details about the tmQMg-L dataset, the 1.37M chemical space, the PL-MOGA algorithm, the estimation of the chemical diversity with average Tanimoto coefficients, the DFT benchmark, repetitions from different random initial populations, additional information on the exploration of the implicit billion spaces, and general computational and chemoinformatics details.

Supplementary weblinks


Comments are not moderated before they are posted, but they can be removed by the site moderators if they are found to be in contravention of our Commenting Policy [opens in a new tab] - please read this policy before you post. Comments should be used for scholarly discussion of the content in question. You can find more information about how to use the commenting feature here [opens in a new tab] .
This site is protected by reCAPTCHA and the Google Privacy Policy [opens in a new tab] and Terms of Service [opens in a new tab] apply.