SEISMO: Increasing Sample Efficiency in Molecular Optimization with a Trajectory-Aware LLM Agent

TL;DR

SEISMO tackles the challenge of sample-efficient molecular optimization under expensive evaluations by deploying a trajectory-aware LLM agent that performs strictly online, inference-time optimization. By conditioning each proposal on the full optimization trajectory, task descriptions, scalar oracle scores, and explanatory feedback, SEISMO achieves substantially higher sample efficiency than established baselines on the Practical Molecular Optimization benchmark, often reaching near-optimal solutions within a few dozen oracle calls. The findings show that providing task descriptions and richer explanations significantly improves search efficiency, highlighting the value of integrating domain knowledge and post-hoc explainability as direct optimization signals. The approach offers a training-free, scalable framework that leverages advances in LLM reasoning to accelerate early-stage drug design without requiring population-based learning or large evaluation budgets.

Abstract

Optimizing the structure of molecules to achieve desired properties is a central bottleneck across the chemical sciences, particularly in the pharmaceutical industry where it underlies the discovery of new drugs. Since molecular property evaluation often relies on costly and rate-limited oracles, such as experimental assays, molecular optimization must be highly sample-efficient. To address this, we introduce SEISMO, an LLM agent that performs strictly online, inference-time molecular optimization, updating after every oracle call without the need for population-based or batched learning. SEISMO conditions each proposal on the full optimization trajectory, combining natural-language task descriptions with scalar scores and, when available, structured explanatory feedback. Across the Practical Molecular Optimization benchmark of 23 tasks, SEISMO achieves a 2-3 times higher area under the optimisation curve than prior methods, often reaching near-maximal task scores within 50 oracle calls. Our additional medicinal-chemistry tasks show that providing explanatory feedback further improves efficiency, demonstrating that leveraging domain knowledge and structured information is key to sample-efficient molecular optimization.

Figures (7)

• Figure 1: Overview of the iterative optimisation cycle of SEISMO. The initial message and prediction feedback are modular and vary between tasks. Colors indicate different levels of information available in different agent modes, corresponding to the evaluation settings used to analyse performance. Italicized text shows an example. The full system prompt and an example conversation are provided in Appendix \ref{['sec:appendix_examples']}.
• Figure 2: Panels (a--c) show best-so-far optimization curves under different oracle feedback variants. In (a), SEISMO optimizes structural similarity to randomly selected molecules together with drug-likeness (QED), averaged over 20 target molecules. Please note that the y-axis does not start at 0 to focus on the practically relevant score range. In (b), SEISMO performs inhibitory concentration (IC50) optimization against SARS-CoV-2 Mpro under novelty and drug-likeness constraints. Please note the logarithmic y-axis. In (c), SEISMO optimizes binding probability for a protein with no known binders using an expensive co-folding-based oracle. Shaded regions indicate 95% confidence intervals over repeated runs.
• Figure 11: Visualisation of the 20 randomly sampled molecules from PubChem that were used as reference molecules in the lead optimisation proxy.
• Figure 21: Comparison of multiple large language models on the 20 random-target similarity+QED task. The line and band denote mean and standard deviation respectively.
• Figure 31: Optimization of a compound for binding affinity to SARS-CoV-2 Mpro. The line and band denote mean and standard deviation respectively, computed over five repeats for each model.