TRACE: LLM agent transforms drug lead optimization with trajectory-aware planning

Drug discovery's lead optimization phase—refining hit compounds into viable candidates by improving ADMET properties while preserving binding affinity—has long relied on one-step molecular optimization methods that ignore the long-term consequences of sequential decisions. To bridge this gap, researchers Lingxiao Li, Haobo Zhang, Ruohao Fan, Bin Chen, and Jiayu Zhou propose TRACE, a trajectory-aware LLM-reasoning agent that frames tool selection as a sequential decision-making problem over action trajectories. Given a lead molecule and an optimization objective, TRACE makes forward-looking decisions under structural constraints, enabling smarter, multi-step refinement.

Tested on multiple ADMET optimization tasks, TRACE outperformed baseline models with higher optimization success rates, larger property improvements, and superior validity—all while maintaining molecular similarity. The agent's ability to reason about future steps using an LLM allows it to avoid local optima common in one-shot approaches. By effectively combining tool planning with an understanding of molecular structure constraints, TRACE represents a practical leap for AI-driven drug development, potentially cutting the time and cost of bringing new drugs to clinical trials.