Large Language\nModels as Molecular Design Engines

The design of small molecules is crucial for technological\napplications\nranging from drug discovery to energy storage. Due to the vast design\nspace available to modern synthetic chemistry, the community has increasingly\nsought to use data-driven and machine learning approaches to navigate\nthis space. Although generative machine learning methods have recently\nshown potential for computational molecular design, their use is hindered\nby complex training procedures, and they often fail to generate valid\nand unique molecules. In this context, pretrained Large Language Models\n(LLMs) have emerged as potential tools for molecular design, as they\nappear to be capable of creating and modifying molecules based on\nsimple instructions provided through natural language prompts. In\nthis work, we show that the Claude 3 Opus LLM can read, write, and\nmodify molecules according to prompts, with impressive 97% valid and\nunique molecules. By quantifying these modifications in a low-dimensional\nlatent space, we systematically evaluate the model’s behavior\nunder different prompting conditions. Notably, the model is able to\nperform guided molecular generation when asked to manipulate the electronic\nstructure of molecules using simple, natural-language prompts. Our\nfindings highlight the potential of LLMs as powerful and versatile\nmolecular design engines.