StateFlux: DFT-Validated Quantum Reactivity across Predicted Metabolic State Networks Shows Limited Incremental Value for Broad Toxicity Prediction
Year: 2026
Type: preprint
Abstract: Mechanistic molecular representations are increasingly used to supplement parent structure with protonation, metabolism, and quantum-chemical reactivity, but added physical detail does not necessarily provide predictive information beyond structure. StateFlux represents each drug as a network of chemical and predicted metabolic states with node-level GFN2-xTB properties and edge-resolved changes in ionization potential, electron affinity, electrophilicity, solvation, and local reactivity. On a locked DILIrank 2.0 development set (n = 376; 156 positive and 220 negative), 10 × 5 scaffold-grouped validation gave AUROC/AUPRC values of 0.8249/0.7727 for ECFP, 0.8049/0.7430 for ECFP + FluxCore, and 0.8010/0.7261 for ECFP + StateFluxCore. Strict temporal and scaffold-novel tests showed no consistent gain, and an independent 5,367-compound genotoxicity transfer benchmark reproduced the pattern (AUROC 0.7765 for ECFP versus 0.7732 for StateFluxCore). To separate limited predictive utility from inaccurate low-cost quantum chemistry, 24 parent-to-metabolite transformations were evaluated with 144 ωB97X-D/6-31+G*/SMD single-point calculations. GFN2-xTB closely tracked DFT changes in electrophilicity (Pearson r = 0.963; MAE = 0.408 eV; 79.2% sign agreement). Thus, transformation-induced quantum reactivity can be consistent with higher-level DFT without providing reliable incremental information for broad toxicity classification. Quantum-chemical fidelity and predictive utility should therefore be evaluated as separate claims in chemical machine learning.
Source: ChemRxiv
Author: Aris Tsai
Institution: Stanford University
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7214188939