Skill-File-Guided Virtual Screening: A Single-Target Formyl Peptide Receptor 2 Case Study of...

Background: Virtual screening outcomes depend on protocol choices normally made implicitly by experts. I tested whether encoding that expertise as a "skill file" (a Markdown reference read by a large-language-model coding agent) changes virtual screening outcomes versus naive agent defaults, in a single-target case study on formyl peptide receptor 2 (FPR2; Protein Data Bank 7T6S, 3.0 Å cryo-electron microscopy). Methods: A Claude Code agent curated 1,000 pChEMBL ≥5 FPR2 actives and 9,829 property-matched decoys, prepared ligand libraries under naive and skill-guided protocols, and screened them with 2 engines: Uni-Dock (graphics-processing-unit-accelerated AutoDock Vina) and DiffDock v1 (native confidence score). The evaluation used receiver operating characteristic (ROC) area under the curve (AUC), Boltzmann-enhanced discrimination of ROC, enrichment factors (EFs), and paired bias-corrected and accelerated bootstrap, DeLong, and permutation tests. Results: On Uni-Dock, the skill-guided protocol raised ROC AUC from 0.701 to 0.733 (paired ΔAUC = +0.020; DeLong P = 3.4 × 10^-3), while top-of-list enrichment regressed (EF@1% 2.67 → 1.23; EF@0.5% 3.33 → 0.00). Because the arms differ on 5 axes at once, this is a joint-protocol effect, not skill-file guidance alone. DiffDock v1 stayed near random (AUC 0.54 to 0.56) on this out-of-distribution G protein-coupled receptor. Conclusion: In this single-target case study, skill-file-guided protocol selection improved global ranking for a physics-based engine on FPR2 while degrading early enrichment, and it did not rescue a diffusion model outside its training distribution. The +0.020 AUC gain reflects a bundled preparation-and-docking-parameter intervention rather than the skill file alone, comes from a single run per protocol with no seed control, and awaits multitarget, multi-seed replication. Code, data, and the skill file are openly released.