An Audit of Structure-Based HIV Drug-Resistance Triage and Its LLM Explanations

Antiviral drug resistance forces medicinal chemists to triage candidate inhibitors by how well their binding survives known resistance mutations. We created ResistScope, a target-agnostic pipeline that docks a candidate against wildtype plus a panel of clinical resistance mutants, scores binding robustness ($\Delta\Delta G$), and explains each mutation's effect with an LLM. Both outputs have ground truth, so we test both on HIV-1 protease (6 PIs) and reverse transcriptase (4 NNRTIs). The docking score is beaten by its own geometry: ranking mutations by distance to the ligand, same receptor but no docking run, tracks measured fold-resistance better than $\Delta\Delta G$ on both targets. The energy earns its cost only at the top of the ranking in the rigid NNRTI pocket, where each drug's top ten holds 7-8 real resistance mutations against 3-4 for distance; on the flexible protease site that ordering reverses. An LLM judge marks 71-72% of the explanations as recovering the correct primary mechanism, and an attribution ablation over real, absent and corrupted structural context shows they condition on the pipeline, most strongly where the model's prior is weakest. A verification using no language model reproduces that ordering. Structural geometry carries the signal; the docking energy mostly does not.