Method-, conformer-count-, and solvent-robust ranking of natural-product screening corpora via xtb semi-empirical quantum chemistry: a multi-axis robustness benchmark

Abstract

Tight-binding semi-empirical quantum chemistry (xtb-GFNn-xTB) has become a workhorse triage tool for ranking large compound libraries ahead of more expensive structure-based scoring. Whether the resulting ranks are robust to the methodological choices an analyst makes — the GFN parameter set, the conformer-ensemble size, and the implicit solvent model — is rarely tested at scale. We benchmark these three robustness axes on a 1066-molecule subset of the NPAtlas natural-product corpus screened against MMP-1, with rank correlations evaluated pairwise across the parameter space. Across the 432/480/528/576/720/816-conformer ladder, pairwise Spearman rank correlations are ρ = 1.0000 at four-decimal precision in every pair on both total energy and HOMO--LUMO gap channels. Switching the GFN parameter set from GFN2-xTB to GFN1-xTB on 432 conformers preserves total-energy ranks at ρ = +0.993 and gap-mean ranks at ρ = +0.978 (n = 1050 paired molecules); switching to GFN-FF gives only ρ = 0.91 and a top-50 intersection of 6/50, suggesting GFN-FF is not a substitute for GFN1/GFN2 at the top of the ranked list. Switching the implicit solvent model among nine ALPB dielectrics spanning ε ≈ 1.9--78 (hexane, chloroform, THF, octanol, ethanol, methanol, acetonitrile, DMSO, water) preserves gap-mean ranks at pairwise ρ ≥ 0.988 in all 36 pairs (median ρ = 0.997, n = 1050) and total-energy ranks at ρ = 1.0000 in every pair, with a 9-way top-50 intersection of 45/50 = 90%. We pair this robustness benchmark with an orthogonal cross-engine consistency check showing that Boltz-2 affinity prediction is dominated by molecular size and lipophilicity (Spearman ρ_MW = -0.65, ρ_logP = -0.54). Keywords: xtb, GFN-xTB, semi-empirical, natural products, MMP-1, screening robustness, conformer ensemble, ALPB, Boltz-2, multi-fidelity. --- Author: Cheongwoo Han (HanCheongWoo) | ORCID: 0009-0004-4805-8815 Affiliation: Genesis_Medicine Lab, Seoul, Republic of Korea / HAN PREDICT, Inc. / Recover Korean Medicine Clinic License: CC-BY 4.0 Funding: Self-funded by HAN PREDICT, Inc. Conflict of Interest: Author is founder of HAN PREDICT, Inc. and operator of Recover Korean Medicine Clinic. No commercial product is sold based on these results. Status: In silico only. No wet-lab data. IRB caveat does not apply (no human subjects). Prior history: Net-new manuscript drafted 2026-05-06 through 2026-05-08; not previously submitted to any venue. Direct OSF deposit as the canonical preprint record. Companion paper to Genesis paper #A (paper_A_zaff_abfe_limitations).