Interpretable Molecular Modeling via LLM Agents with Chain-of-Thought Reasoning

In the past decade, the application of AI and machine learning has significantly accelerated progress in materials discovery. Recently, the emergence of large language models (LLMs) has opened new opportunities in complex and fine modeling in materials science. Predicting properties (e.g., toxicity, melting point) of unknown molecules has been a central task. While existing LLM-based property predictors often achieve high prediction accuracy, their reasoning processes remain opaque, offering limited or no interpretability to researchers. To address this gap, we propose an agent-based framework that incorporates chain-of-thought (CoT) reasoning to enhance the interpretability of molecular property predictions. Our approach utilizes LLM agents to infer intermediate fundamental properties (e.g. HOMO energy, LUMO energy, dipole moment, etc.) prior to estimating the final target property. When a user submits a query to predict a property for a given molecule (e.g., melting point), an orchestrator agent dynamically identifies relevant intermediate properties, delegates their prediction to specialized agents. Also, the orchestrator drives a structural analysis agent to interpret key molecular features such as symmetry, rigidity, and polarity etc. in natural language. These outputs are then integrated by a reasoning agent to generate a final prediction with interpretable justification. We conducted preliminary analyses in domains such as organic dyes, demonstrating that our CoT-guided, multi-agent workflow improves both predictive performance and interpretability. This approach highlights the potential of LLM agents to support human-interpretable reasoning in AI-assisted materials discovery process.