Abstract:The design of functional transition metal complexes (TMCs) is hindered by the combinatorial explosion of the search space spanned by various metals and ligands, necessitating efficient multi- objective optimization strategies. Traditional genetic algorithms (GAs) are frequently employed in this domain, utilizing random mutations and crossovers steered by explicit mathematical objective formulations to navigate the search space. The transfer and sharing of knowledge across different GA optimization tasks, however, remain challenging. Here, we introduce the integration of large language models (LLMs) into the evolutionary optimization framework (LLM-EO) for TMCs. LLM- EO significantly outperforms traditional GAs due to the intrinsic chemical knowledge embedded within LLMs, acquired during their extensive pretraining. Notably, without the need for supervised fine-tuning, LLMs can leverage the entirety of historical data amassed during the optimization processes, demonstrating superior performance compared to LLMs that are limited to the best TMCs identified in the evolutionary cycle. Specifically, LLM-EO identifies eight out of the top 20 TMCs with the largest HOMO-LUMO gaps by interrogating merely 200 candidates within a vast search space of 1.37 million TMCs. Through prompt engineering using natural language, LLM-EO introduces unparalleled flexibility in multi-objective optimizations, especially when guided by seasoned researchers, thereby circumventing the necessity for intricate mathematical formulations. As generative models, LLMs possess the capability to propose novel ligands and TMCs with unique chemical properties by amalgamating both internal knowledge and external chemistry data, thus combining the benefits of efficient optimization and molecular generation. With the increasing potential of LLMs, both in their capacity as pretrained foundational models and new strategies in post-training inference, we anticipate broad applications of LLM-based evolutionary optimization in the fields of chemistry and materials design.

Inverse design of ligands using a deep generative model semi‐supervised by a data‐driven ligand field strength metric

Deep Generative Model for the Dual-Objective Inverse Design of Metal Complexes

Generative Design of Functional Metal Complexes Utilizing the Internal Knowledge of Large Language Models

Ligand additivity relationships enable efficient exploration of transition metal chemical space

Learning to design drug-like molecules in three-dimensional space using deep generative models

A fully differentiable ligand pose optimization framework guided by deep learning and a traditional scoring function

Partial to Total Generation of 3D Transition Metal Complexes

LigandDiff: 3D Transition Metal Complex Generation with Diffusion Models

A fully differentiable ligand pose optimization framework guided by deep learning and traditional scoring functions

Design of Ligand-Binding Proteins with Atomic Flow Matching

Constructing and evaluating machine-learned interatomic potentials for Li-based disordered rocksalts

LigandDiff: de Novo Ligand Design for 3D Transition Metal Complexes with Diffusion Models

Exploiting Ligand Additivity for Transferable Machine Learning of Multireference Character Across Known Transition Metal Complex Ligands

Modelling ligand exchange in metal complexes with machine learning potentials

Data-driven prediction of complex crystal structures of dense lithium

ReaLigands: A Ligand Library Cultivated from Experiment and Intended for Molecular Computational Catalyst Design

Prediction of stability constants of metal–ligand complexes by machine learning for the design of ligands with optimal metal ion selectivity

Graph neural networks for predicting metal–ligand coordination of transition metal complexes

Ligand Many-Body Expansion as a General Approach for Accelerating Transition Metal Complex Discovery

Virtual Ligand-Assisted Optimization: A Rational Strategy for Ligand Engineering