Accelerating QSAR Modeling with Human-AI Collaboration: A Case Study from a Computational Blind Challenge

Large language models (LLMs) enable agents that can accelerate research in formally verifiable domains, but their utility in empirical or noisy settings such as drug discovery remains less well characterized. We examine this question through our participation in the activity track of the OpenADMET Pregnane X Receptor (PXR) Blind Challenge, which required predicting PXR induction potency $\(pEC_{50}\)$ from molecular structure. Over an approximately 12-day campaign, our approach coupled an LLM-powered agent with human strategic direction; the organizer's final-results post ranked the resulting entry 13th among 100 evaluated final activity entries, assigned it to Tier~1, and reported an MAE of $\(0.4221 \pm 0.0285\$) on the 260-compound AS2 set, whose structures were visible but outcomes remained hidden through submission. In this work, we study the discovery process by illustrating our modeling strategy, highlighting our lessons learned during the challenge and providing post-competition analysis of human-agent interactions. The human authors framed the problem, seeded and prioritized modeling hypotheses, revised evaluation policies, and authorized the final lock, while the agent implemented, orchestrated, and debugged experiments. This case study demonstrates the potential of mixed-initiative human--agent workflows to produce competitive solutions in scientific machine learning while retaining human oversight.