My Chemical Harness uses LLM agents to design synthesizable drug molecules

The My Chemical Harness framework reimagines computational molecular design by operating directly on synthetic pathways rather than static molecular graphs. Instead of having an LLM hallucinate candidate molecules with no guarantee of synthesizability, the system restricts the LLM to a high-level strategic role: it selects preferences such as route length, reaction family, or exploration pressure. A deterministic local engine then constructs, validates, deduplicates, scores, and evolves actual pathways using purchasable building blocks and verified reaction templates. This architectural separation eliminates the risk of LLM-generated unrealistic steps while still leveraging the model's ability to guide global search.

In benchmarking against the soluble epoxide hydrolase (sEH) proxy task—a common drug discovery target—the approach achieved state-of-the-art results across three key metrics: sEH score, synthetic accessibility score, and AiZynthFinder success rate. The LLM agent outperformed both single-pass LLM baselines and deterministic evolutionary controllers without any training, fine-tuning, or dedicated generative models. The paper (27 pages, 10 figures) demonstrates that constrained LLM agents can significantly accelerate molecular discovery by ensuring every candidate comes with a feasible synthesis route—a critical requirement for translating computational hits into real-world experiments. This work points toward a future where AI guides drug design not by dreaming up molecules, but by reasoning about how to build them.