AutoSMILES: An Agentic Workflow for Property-Directed Molecule Generation Using Large Language Models

The discovery of new molecules with specific properties represents a major challenge in modern chemistry and drug development. Traditional methods involve testing thousands or even millions of different molecules, which is time-consuming and costly. This thesis presents AutoSMILES, an innovative framework based on a multi-agent architecture for AI-driven molecular design. The core innovation lies in employing two agents: a generative agent that creates molecules and a prompt agent that provides dynamic, contextual feedback based on historical generation attempts. This agentic approach o!ers paths of solutions to solves fundamental problems of iterative generation and inefficiency that limit traditional systems. The research reveals crucial discoveries in prompt engineering for computational chemistry. The prompt agent produces drastic performance improvements, demonstrating that prompt engineering strategies represent more than a critical factor. The study also establishes universal trends: higher temperatures systematically improve success rates while eliminating ine"cient loop behaviors. Most importantly, the work shows that sophisticated agent-based feedback mechanisms can transform molecular generation e!ectiveness, often providing greater improvements than scaling model size or complexity. These results open new perspectives for molecular generation systems based on agents.