Physics beats diffusion: Agentic AI-driven virtual screening benchmark on a GPCR target

Virtual screening (VS) campaigns require expert decisions at every stage, from active compound curation and decoy generation to receptor preparation, docking engine selection, and statistical evaluation. I show that an autonomous large language model (LLM) coding agent (Claude Code, Anthropic) can design and execute a complete VS benchmark pipeline without human coding intervention that requires only high-level scientific direction. The agent curated 1,000 FPR2 (a GPCR receptor) actives from ChEMBL (pChEMBL ≥ 5), generated ca. 10,000 property-matched decoys, prepared ligand libraries using two protocols (naive defaults and expert-guided), configured and ran docking with two fundamentally different engines; 1) Uni-Dock (GPU-accelerated physics-based) and 2) DiffDock (diffusion-based machine learning). Then it performed full statistical evaluation including ROC AUC, BEDROC, enrichment factors, DeLong tests, and paired bootstrap confidence intervals. Uni-Dock achieved ROC AUC = 0.70–0.73 with significant discrimination (permutation p < 0.0001), while DiffDock confidence scores yielded near-random performance (AUC = 0.54–0.56; Cliff’s d = negligible), consistent with its known underrepresentation of GPCR targets in training data. Expert-guided protocols improved Uni-Dock AUC by +0.020 (DeLong p = 0.003; paired bootstrap p = 0.002). Single-ligand redocking confirmed Vina reproduces the crystal pose (RMSD 0.22–0.39 Å) whereas both Uni-Dock batch mode (5.2–5.7 Å) and DiffDock (23–29 Å) failed. All code, data, and the agent’s skill file are openly available. Scientific contribution: This is the first demonstration of an LLM coding agent autonomously constructing a reproducible VS benchmark from scratch. The resulting benchmark provides the first head-to-head comparison of Uni-Dock and DiffDock on a GPCR target that reveals that physics-based docking (AUC = 0.70–0.73) substantially outperforms diffusion-based ML docking (AUC = 0.54–0.56, near-random) for this underrepresented target class.