SmileyLlama: Modifying Large Language Models for Directed Chemical Space Exploration

The recent transformation of the Llama-3.1-8B-Instruct model into SmileyLlama marks a significant advancement in the application of large language models (LLMs) for chemical exploration. By employing supervised fine-tuning of engineered prompts, researchers have created a chemical language model that can generate novel drug-like molecules tailored to user specifications. This capability was benchmarked against traditional CLMs trained from extensive ChEMBL data, showcasing SmileyLlama's ability to produce valid and innovative compounds. Furthermore, the integration of direct preference optimization enhances the model's adherence to prompts, while the iMiner reinforcement learning framework aids in predicting new drug molecules with optimized 3D conformations and high binding affinity to targets, such as the SARS-Cov-2 Main Protease. Although the current dataset focuses on drug discovery, the methodologies developed could be extended to various chemical applications, including chemical …