The chemical space containing all molecules that can be made with the elements of the periodic table and the laws of physics that determine their possible combinations is extremely large. For example, the space of all drug-like molecules is estimated to contain ~1060 molecules. Upon removing constraints on structure and composition, this number quickly increases up to ~10200 molecules.
Similar to the universe, one can think of the chemical space as a constellation of galaxies, each concentrating a massive amount of molecules sharing a common feature. One of these "galaxy subspaces" is that of the transition metal complexes (TMCs), which are compounds made of organic molecules rich in carbon and hydrogen bound to the transition metals of the periodic table.
In the chemical and materials sciences, TMCs are at the core of highly relevant processes; for example, the industrial production of medicines and, in the green shift context, the recycling of CO2 and the use of renewable energies. There is thus a pressing need for discovering the TMCs that enable these technologies. However, finding these compounds within the vast TMC space is a hard needle-in-a-haystack problem. This was the challenge addressed by Hylleraas members Hannes Kneiding, Ainara Nova, and David Balcells.[1]
The exploration of the TMC space for finding systems with optimal properties requires two key elements: A) An extensive dataset of organic ligands, and B) An efficient exploration algorithm. For the dataset, Kneiding and co-workers developed tmQMg-L, a large and diverse library containing 30K ligands, which was derived from previous work of the group[2,3]. The main difficulty was to define the charge of the ligands and how they bind to the transition metals, a problem that was solved using an approach based on quantum mechanics known as natural bond orbital analysis.
For the algorithm, the authors developed a generative AI approach based on genetic algorithms (GAs). In general, generative AI predicts the X features that define a system from the Y target properties that enable a given a application of interest. This task is the reverse of that commonly done by machine learning (ML) models, which instead predict Y from X. In the context of generative AI for chemicals and materials, GAs are increasingly regarded as one of the most powerful methods. The implementation of GAs require to express the chemical space of interest in a modular manner (chromosome), using components (genes) that can be manipulated with the same genetic transformations observed in nature (mutation and crossover).
The Hylleraas team implemented a novel GA for TMCs called PL-MOGA. By using the ligands of the tmQMg-L library, this algorithm was used to explore vast chemical spaces; for example, billions of TMCs based on a subset of 252 organic ligands. Within these spaces, the PL-MOGA allowed for finding TMCs within selected regions of the Pareto front, which defines the limits for the joint optimization of two or more target properties. In a known space of 1.37 million TMCs used as a benchmark, the algorithm found the Pareto solutions after exploring only 1% of the total space. In much larger spaces containing billions of TMCs, the PL-MOGA also found Pareto solutions relative to the experimental CSD database[4]. The generation of the TMCs could be easily explained with the concepts commonly used by chemists and, relative to other state-of-the-art GAs, the PL-MOGA did not require previous knowledge on the ranges spanned by the target properties.
In drug discovery applications, this method can be used for pre-screening massive amounts of candidates based on properties related to their stability and activity like, for example, the HOMO-LUMO gap, polarizability, and lipophilicity. Other interesting perspectives include the further expansion of the tmQMg-L library and the augmentation of the PL-MOGA algorithm with other ML methods[5].
The main contributor to this work was Hannes Kneiding, a PhD candidate in the Oslo node of the Hylleraas Centre at the UiO Department of Chemistry, supported by the COMPSI MSCA program affiliated to the dScience Centre. Ainara Nova, Hylleraas PI of RT5, contributed to the development of the tmQMg-L dataset, and David Balcells, Hylleraas PI of RT7, designed and supervised the research in the context of the FRIPRO catLEGOS project.
References:
[1] Kneiding, H.; Nova, A.; Balcells, D. "Directional multiobjective optimization of metal complexes at the billion-system scale", Nat. Comput. Sci., 2024, in press.
[2] Balcells, D.; Skjelstad, B. B. "tmQM Dataset — Quantum Geometries and Properties of 86k Transition Metal Complexes", J. Chem. Inf. Model., 2020, 60, 6135.
[3] Kneiding, H.; Lukin, R.; Lang, L.; Reine, S.; Pedersen, T. B.; de Bin, R.; Balcells, D. "Deep Learning Metal Complex Properties with Natural Quantum Graphs", Digital Discovery, 2023, 2, 618.
[4] The Cambridge Structural Database.
[5] Kneiding, H.; Balcells, D. "Augmenting Genetic Algorithms with Machine Learning for Inverse Molecular Design", ChemRxiv, 2024, preprint.
Abbreviations:
tmQMg-L = transition metal Quantum Mechanics graph Ligand dataset
PL-MOGA = Pareto Lighthouse Multi-Objective Genetic Algorithm