Skip to main content
NCBI home page
ACS Omega logoLink to ACS Omega
. 2026 Sep 30;11(40):62050–62059. doi: 10.1021/acsomega.6c08355

Benchmarking Affinity-Based Docking for Irreversible Acetylcholinesterase Inhibitors: A Physics-Informed Reactivity Framework for Stereoselective Potency Prioritization

Jeongyun Kim †, Jin Yoo †,‡, Doo-Hee Lee ‡,*, Ku Kang ‡,*
PMCID: PMC13670796  PMID: 42890015

Abstract

Irreversible organophosphorus inhibitors of acetylcholinesterase (AChE) act by covalent phosphylation of the catalytic serine, yet their computational prioritization is often based on affinity scores from noncovalent docking. We benchmark that assumption against a curated set of measured human-AChE inhibition rate constants (n = 14; G-, V-, and organophosphate-pesticide classes, including a configuration-resolved VX pair). AutoDock Vina affinity (Spearman ρ = +0.12) and GNINA CNN scores (ρ = −0.40) do not provide statistically reliable potency rankings; explicit covalent docking also remains weak (ρ = +0.23, n = 11), with bootstrap confidence intervals crossing zero. Selecting covalent-compatible (near-attack) poses rather than top-ranked poses does not change this outcome. A quantum-chemical cluster scan supports a mechanistic interpretation: once the activated serine alkoxide/reactive complex is formed, P–O bond formation is downhill within the model, so the discriminating information is expected to reside in electrophilicity, leaving-group chemistry, and transport into reactive geometry rather than in a single docked pose. Consistently, mechanism-oriented molecular descriptors recover a modest but reproducible signal: TPSA alone correlates with potency (ρ = −0.62), and a Gaussian-process model used primarily for calibrated uncertainty yields leave-one-out ρ = +0.55 with permutation support (p ≈ 0.01) and a bootstrap CI whose lower bound lies near zero. A leave-both-VX-enantiomers-out test preserves the experimental S P > R P ordering. Given the small panel (n = 14), the descriptor correlations and stereochemical results are reported as exploratory and uncertainty-bounded rather than predictive. The resulting framework is presented as an uncertainty-bounded prioritization tool for data-scarce covalent AChE inhibitors, not as a substitute for kinetic measurement.

graphic file with name ao6c08355_0009.webp

graphic file with name ao6c08355_0007.webp

1. Introduction

Organophosphorus (OP) nerve agents and related pesticides inhibit AChE through covalent modification of the catalytic serine, producing acute cholinergic toxicity and motivating both countermeasure development and computational risk prioritization. − The structural basis of AChE catalysis and inhibition is well established: a buried active-site gorge positions substrates and inhibitors near the Ser203–His447–Glu334 catalytic triad, and phosphylation can be followed by aging reactions that reduce oxime reactivation. −

The computational tension is straightforward. Standard docking scores estimate reversible protein–ligand association, whereas the experimental end point for irreversible AChE inhibitors is an inactivation rate constant. Docking may still be useful for pose generation, active-site sanity checks, or identifying steric incompatibility, but whether an affinity score ranks covalent potency is an empirical question. This is especially important for stereogenic phosphorus centers, where enzymatic discrimination can span orders of magnitude. ,− The overall study design is summarized in Figure .

1.

1

Open in a new tab

Study design (schematic). The cartoon of the AChE active-site gorge and the small inset plots are schematic illustrations, not rendered structural or measured data; the quantitative results appear in Figures –. Irreversible AChE inhibition (measured human-AChE rate constants) is benchmarked against affinity-based docking scores and compared with reaction-aware molecular descriptors. Four inhibitors are drawn as representative examples only; all 14 kinetic anchors are used throughout and appear in the data figures.

1.1. Relation to Prior Work

That affinity-based docking scores rank covalent inhibitor potency poorly is not, in itself, a new observation. Systematic evaluations have documented the limited correlation between docking scores and covalent potency, and covalent-aware learning has been proposed to improve affinity prediction; dedicated covalent-docking methods have been developed for pose prediction and library-scale screening , and benchmarked against one another, and the scope and limitations of covalent docking have been reviewed in detail; , and the SCAR/SCARdock strategy showed that scoring the noncovalent poses that are compatible with covalent-bond formation can improve ranking in some systems. Our contribution is therefore deliberately delimited. It is not the qualitative claim that docking underperforms for covalent inhibitors, but (i) a curated, fully reproducible benchmark against measured human-AChE inactivation kinetics across G-, V-, and OP-pesticide classesincluding GNINA covalent mode and covalent-compatible (near-attack) pose scoring; (ii) an explicit low-data uncertainty treatment (label permutation, bootstrap intervals, applicability domain, and a leave-both-VX-enantiomers-out circularity control) rarely applied at n = 14; and (iii) a reaction-aware descriptor baseline that recovers a modest but reproducible signal where affinity scores do not.

Here, we ask a deliberately narrow question: in a human-AChE kinetic benchmark, do affinity-based docking scores contain enough signal to rank irreversible OP potency? We compare noncovalent Vina and GNINA scores with explicit covalent docking and with a small, mechanism-oriented descriptor model. The goal is not to claim a definitive toxicity predictor from n = 14 data points. Rather, it is to test a common modeling assumption, quantify its uncertainty, and provide a reproducible baseline for future covalent inhibitor studies in the spirit of current QSAR, computational toxicology, and model-validation guidance, including recent machine-learning work on OP and Novichok-related property prediction. −

2. Methods

2.1. Kinetic Anchor Set

The measured panel contains 14 human-AChE inhibition rate constants assembled from homogeneous literature sources: Worek et al. for structurally diverse OP inhibitors and Bester et al. for configuration-resolved VX kinetics. , Chemical classes include G-agents, V-agents, and OP-pesticides. Compounds requiring metabolic activation are retained in the benchmark only as labeled kinetic anchors and are treated cautiously in the covalent docking analysis. Four A-series structures are treated only as unlabeled projection targets; no claim of experimental validation is made for them.

2.2. Ligand Preparation and Conformer Handling

Structures were built in their neutral, as-drawn forms; the phosphyl center is nonionizable and amine centers were left neutral, consistent with the covalent-attack chemistry. All stereoisomers were enumerated explicitly with RDKit. For each stereoisomer, a 3D structure was embedded with the ETKDGv3 algorithm (fixed random seed) and optimized with the MMFF94 force field; the single lowest-energy conformer was carried into docking, and, for the quantum descriptors, further optimized at the DFT level below. Because docking and quantum descriptors can depend on the starting conformer, we tested this dependence directly for the four most flexible inhibitors (VX, VR, the diethyl tabun analogue, and fenamiphos): three distinct low-energy conformers each (ΔE MMFF ≤ 0.9 kcal mol–1) were reoptimized and their descriptors recomputed at the production level. The phosphorus partial charge varied by ≤0.007 e and the HOMO–LUMO gap by ≤0.06 eV for three of the four molecules (0.23 eV for VX), whereas the molecular dipole was more conformer-dependent; we therefore use the lowest-energy conformer consistently and report the full conformer-sensitivity table in the Supporting Information.

2.3. Docking Benchmark

Noncovalent docking used AutoDock Vina and GNINA against human AChE (PDB 4EY7), with the search box centered on Ser203 Oγ. ,− The box was 22.5 × 22.5 × 22.5 Å, and the reported score for each ligand is that of the top-ranked pose from the same engine (GNINA for the CNN scores, Vina for the affinity scores). Because the box is centered on Ser203 Oγ, we could additionally identify, for each ligand, the covalent-compatible (near-attack) pose whose phosphorus is closest to Ser203 Oγ and score it in the manner of the SCAR/SCARdock strategy; this analysis is reported in the Results. Robustness of the ranking result to the docking setup was confirmed with a deposited sensitivity protocol: across ±2 Å box translations (in x and y) and alternative random seeds (seven variants), the Vina rank correlation remained weak (Spearman ρ ∈ [−0.03, +0.17]) and the GNINA CNN correlation was weak-to-negative (ρ ∈ [−0.33, −0.10]), consistent with the primary result and never yielding a reliable potency rank (run_sensitivity.sh; SI). The influence of crystallographic active-site waters was tested by redocking with the 4EY7 waters near Ser203 Oγ retained; the result is reported in the Supporting Information. Explicit covalent docking in GNINA covalent mode was performed for directly phosphylating compounds only.

2.4. Quantum Chemistry and Descriptors

Ligand geometries were optimized at ωB97X-D3/def2-SVP with CPCM­(water) and RIJCOSX in ORCA, followed by def2-TZVP single-point calculations. − The extracted electronic descriptors included HOMO, LUMO, gap, dipole moment, and phosphorus partial charges. Structural descriptors included molecular weight, topological polar surface area, rotatable bond count, and leaving-group/connectivity flags. −

2.5. Reaction-Coordinate Scan

The quantum-cluster model is specified here in full, since its composition determines what the scan can and cannot show. It contains 70 atoms in four fragments and carries a net charge of −1 with a closed-shell singlet ground state: the complete VX molecule (42 atoms; C11H26NO2PS) together with side-chain models of the entire catalytic triad4-methylimidazole (C4H6N2) for His447, propionate (C3H5O2−) for Glu334, and methanol/methoxide (CH4O/CH3O–) for Ser203. The residues are truncated at the Cβ–Cγ level and enter as discrete side-chain fragments rather than as capped backbone units, so no peptide bond is present in the model. In the reactant complex, the serine model is the neutral alcohol and the histidine model is neutral imidazole; in the structure from which the scan is launched, the proton has already transferred, giving methoxide and imidazolium, so the charge decomposes as methoxide (−1) + imidazolium (+1) + propionate (−1). The scan therefore probes bond formation at an already activated nucleophile, with His447 present as the protonated general base rather than as a spectator; it does not model the proton transfer itself. A relaxed scan along the forming P···Oγ coordinate was run from 3.30 to 1.60 Å in 0.10 Å steps, all remaining degrees of freedom being optimized at each step, at ωB97X-D3/def2-SVP with CPCM­(water) and RIJCOSX in ORCA, following established cluster-model practice for enzymatic reaction interpretation. − This scan was computed for a single representative inhibitor (VX) and is used only as an illustrative relaxed approach coordinate; it is not a transition-state search and is not claimed to be quantitatively representative of all 14 inhibitors. This calculation is not presented as a full free-energy barrier for the complete general-base mechanism; it tests whether P–O bond formation itself introduces a large additional barrier after activation. No stationary point was located and no frequency analysis was performed, so no barrier height is reported from this model. We additionally attempted two larger clusters (88 atoms) that added three formamide molecules as minimal models of the oxyanion-hole backbone amides of Gly121, Gly122, and Ala204; neither produced a converged scan surface, and they are reported here only as unsuccessful attempts. Their inputs and logs are included in the deposited package so that the failure is inspectable rather than merely asserted.

2.6. Machine Learning, Circularity Control, and Uncertainty

Gaussian-process regression was used as a low-data uncertainty model, and a random forest served as a nonparametric baseline. − Primary validation used a leave-one-out cross-validation. To remove the most obvious circularity risk in stereochemical inference, both VX enantiomers were withheld simultaneously in a leave-both-VX-out test before predicting their rank order. Statistical support was estimated by 1,000 label permutations (each requiring a full leave-one-out refit) and 10,000 bootstrap resamples; the deposited script al_robustness.py reproduces both. The applicability domain (AD) was quantified from the k-nearest-neighbor distance (k = 3) in the standardized (z-scored) descriptor space used by the model, with the domain limit set at the training mean plus two standard deviations of the training k-NN distances, following QSAR best-practice recommendations; − , the descriptor list, scaling, and threshold are given in the Supporting Information.

3. Results and Discussion

3.1. Affinity-Based Docking Does Not Provide a Reliable Potency Rank

Docked poses were chemically plausible: phosphorus was generally oriented toward the catalytic serine and the ligands occupied the active-site gorge (Figure a). Across the whole docked panel, the Vina affinity spans only 3.26 kcal mol–1 (−7.01 to −3.75; Figure b), a narrow window within which the measured rate constants vary by more than seven orders of magnitude. Plausible poses, however, did not translate into reliable potency ranking. Vina affinity was weakly positive and statistically indistinguishable from zero (Spearman ρ = +0.12, p = 0.68, 95% bootstrap CI [−0.47, +0.64]). GNINA CNN scores trended negative (ρ = −0.40, p = 0.15, CI [−0.82, +0.21]). Explicit covalent docking did not resolve the problem (ρ = +0.23, p = 0.50, n = 11, CI [−0.57, +0.87]; Figure a–c). Figure d shows the failure concretely: fenamiphos, the weakest measured inhibitor of the panel, and butylsarin, among the strongest, occupy the same gorge with affinities that rank them in the wrong order. Thus, in this benchmark, the defensible conclusion is not that docking always antiranks potency, but that affinity scores do not provide a statistically reliable ranking signal for these irreversible inhibitors.

2.

2

Open in a new tab

Noncovalent docking of the organophosphorus panel against human AChE (PDB 4EY7). (a) Representative top-ranked GNINA poses in the active-site gorge, with the Ser203–His447–Glu334 catalytic triad in salmon and phosphorus oriented toward the catalytic serine; the search box was centered on Ser203 Oγ, 22.5 Å per side. The box and its placement are shown in Supporting Information Figure S2. (b) AutoDock Vina affinity for every docked structure, coloured by chemical class. Thirty-eight stereoisomers were docked in total: the 14 configuration-resolved kinetic anchors that carry a measured ki , together with 24 further stereoisomers and A-series projection targets for which no measured rate constant exists. Only the 14 anchors enter the rank correlations reported in the text; the remaining structures are plotted to show the narrow affinity window (−7.01 to −3.75 kcal mol–1) spanned by the entire panel. (c) Within-compound affinity difference between P-epimers, for the nine compounds docked as stereoisomer pairs or sets. The largest separation is 0.83 kcal mol–1 (A-234) and the smallest are 0.05 (VX) and 0.04 kcal mol–1 (A-242), so affinity scoring is close to degenerate for precisely the stereochemistry that the measured kinetics resolve over two orders of magnitude. The benchmark treats the docking score as a rank predictor and evaluates it directly against measured human-AChE inactivation rate constants ki .

3.

3

Open in a new tab

Docking score versus measured potency for the 14 kinetic anchors. (a) Measured logki against AutoDock Vina affinity (Spearman ρ = +0.12, not significant), colored by chemical class. (b) The same anchors against the GNINA CNN affinity (ρ = −0.40, n = 14); the two labelled extremes are butylsarin (BSAR) and fenamiphos (FEN). (c) Spearman ρ against logki with 95% bootstrap confidence intervals for all four scoring approaches. The three docking scoresVina affinity, GNINA CNN, and explicit GNINA covalent docking (ρ = +0.23, n = 11)all have intervals that cross zero. The physics-informed descriptor model is shown for comparison; its interval is reported in the text and in the Supporting Information, and its lower bound is close to zero. (d) The failure made concrete: fenamiphos (strong docking score, weakest measured inhibitor) and butylsarin (weaker docking score, among the strongest inhibitors) occupy the same gorge, so the affinity ordering inverts the measured one for the two extremes of the panel.

We also tested the covalent-compatible pose-selection strategy suggested by the SCAR/SCARdock work. Because the search box is centered on Ser203 Oγ, the near-attack pose (phosphorus closest to Ser203 Oγ) coincides with the top-ranked pose for all 14 inhibitors, so selecting covalent-compatible poses leaves the rank correlations unchanged (Vina ρ = +0.12, GNINA CNN ρ = −0.40); only 8 of the 14 inhibitors place any pose within 4.5 Å of Ser203 Oγ. Noncovalent scoring is also close to blind to configuration: within each compound, the two P-epimers differ by at most 0.83 kcal mol–1, and by 0.05 and 0.04 kcal mol–1 for VX and A-242, respectively (Figure c), whereas the measured VX pair differs by two orders of magnitude in k i (Figure c). Consistent with the SCAR/SCARdock premise, we therefore did evaluate covalent-compatible scoring, but for this organophosphorus panel, it does not recover a reliable ranking. The negative benchmark is likewise stable to the treatment of crystallographic active-site waters (Supporting Information).

4.

4

Open in a new tab

Covalent reactivity and stereoselectivity of VX at human acetylcholinesterase. (a) Relaxed potential-energy scan along the forming P···Oγ coordinate on a 70-atom quantum-cluster model of VX with the complete Ser203–His447–Glu334 catalytic triad; composition, charge, and level of theory are given in the Methods section and the Supporting Information. Energies are relative to the 3.30 Å point. This is an illustrative approach coordinate, not a transition-state search: no stationary point was located and no frequency analysis was performed. Two features are reported rather than smootheda discontinuous 5.3 kcal mol–1 drop between 3.00 and 2.90 Å (shaded), which reflects a geometric reorganization rather than progress along a continuous path, and a +1.1 kcal mol–1 rise between 2.90 and 2.50 Å. The profile is therefore not monotonically downhill and does not support describing the approach as barrier-free. Below 2.0 Å, it descends to a phosphoserine adduct 39.7 kcal mol–1 beneath the starting point. (b) Enzyme P-stereoselectivity for three classical agents, as log10 of the ratio of inhibition rate constants of the more-toxic to the less-toxic P-isomer, from values tabulated by Worek et al. (soman, sarin; ref ) and Bester et al. (VX, 1.4 × 108 / 1.2 × 106 M–1 min–1; ref ). The arrow on the soman bar denotes a lower bound. (c) Measured human-AChE inhibition rate constants for the configuration-resolved VX pair used as kinetic anchors here. (d) Out-of-sample stereochemical control. Both VX enantiomers were withheld and the Gaussian-process model refitted on the remaining 12 anchors. Bars are predicted logki with ±1 SD predictive uncertainty; open diamonds are the measured values from panel (c). The model recovers the experimental S P > R P ordering (8.50 versus 6.48) without having seen either enantiomer, although the R P interval is wide and the result is reported as exploratory. All panels were drawn by the authors with the deposited script make_fig4_r2.py; panels (b) and (c) replot rate constants tabulated in refs and . No graphic from any prior publication is reproduced or adapted.

3.2. Reaction-Coordinate Scan Rationalizes the Weak Affinity Signal

The cluster scan descends from the activated reactive complex to the phosphoserine adduct, with the final adduct 39.7 kcal mol–1 lower in the model relative to the 3.30 Å starting point (Figure a). Two features of the computed profile are reported explicitly rather than smoothed over, because both bear on how far the curve may be read. First, the energy drops by 5.3 kcal mol–1 across the single step from 3.00 to 2.90 Å. A relaxed scan that falls this steeply in one 0.10 Å increment has changed branch: the optimized geometry reorganizes at that point, so the segment is not a continuous path in the reaction coordinate. Second, between 2.90 and 2.50 Å, the energy rises by 1.1 kcal mol–1 before descending again. The profile is therefore not monotonically downhill, and we do not describe the approach as barrier-free; an earlier version of this figure carried such an annotation, which the data do not support and which has been removed. The interpretation drawn below is unaffected by either feature, because it concerns the bond-forming region under 2.0 Å, over which the coordinate descends smoothly by more than 28 kcal mol–1; the reorganization occurs at longer P···Oγ separation. The scan was run with ORCA’s LooseOpt convergence and with 28 of the 70 atoms held at fixed Cartesian positions; the remaining 42, including the hydrogen-bonded contacts between the alkoxide, imidazolium, and propionate fragments, relaxed at every step. Either the loose optimization threshold or a rearrangement among those unconstrained atoms can move the relaxed geometry to a different local minimum between successive steps, which is the most likely origin of the discontinuity. Because the calculation starts from an activated serine alkoxide, this result should not be read as the full enzymatic activation barrier. It instead supports a narrower interpretation: once the reactive complex has formed, P–O bond formation itself is not the step most likely to encode large rank differences. This makes a single affinity-ranked pose a limited proxy for potency and motivates descriptors tied to electrophilicity, leaving-group identity, polarity, and access to productive geometry.

We emphasize the limits of this single-ligand scan. It does not locate or validate a transition state, and it is not evidence of a barrier-free reaction. In particular, the general-base activation of Ser203 by His447 and the departure of the leaving group are plausibly rate-influencing and ligand-dependent steps: because the 14 inhibitors differ mainly in their leaving group, leaving-group electronicspartly captured here by the frontier-orbital and phosphorus-charge descriptorsare expected to matter for potency. An exploratory general-base QM-cluster model that included His447 did not yield a converged, transferable barrier, which reinforces our decision to report descriptor-level reactivity rather than absolute mechanistic barriers. A quantitative treatment of the leaving step and the complete general-base path, with explicit transition states for each chemotype, is the natural next study.

3.3. Mechanism-Oriented Descriptors Recover a Modest but Reproducible Signal

Single descriptors already reveal information absent from docking. TPSA correlates with measured logk i (Spearman ρ = −0.62, permutation p = 0.019, 95% bootstrap CI [−0.81, −0.16]), and leaving-group flags and phosphorus charge have mechanistically interpretable signs (Figure ; Figure b). A Gaussian-process model combining electronic and structural descriptors yields LOO–CV ρ = +0.55; a random-forest baseline gives ρ = +0.48 (Figure a). The GP result is supported by a label-permutation test (p = 0.011, 1,000 shuffles; the observed ρ lies above the 95th percentile of the permuted null, which is +0.34). The bootstrap interval is reported but not used to support significance: across random seeds its lower bound moves either side of zero at n = 14, so it does not robustly exclude zero, and we rest the claim on the permutation test instead. By contrast, the docking intervals extend far into negative territory (Vina [−0.47, +0.64], GNINA CNN [−0.82, +0.21]), which is the comparison that matters here. We therefore describe the model as a calibrated, exploratory prioritization tool rather than a high-accuracy predictor; the central result of this study is the negative docking benchmark, and the descriptor model is reported as a secondary, uncertainty-aware baseline.

5.

5

Open in a new tab

Mechanism-oriented descriptors for the 14 measured anchors and the A-series projection targets. Electronic descriptors (HOMO, LUMO, gap, dipole, phosphorus partial charge) from ωB97X-D3/def2-TZVP//def2-SVP with CPCM­(water), together with structural descriptors (molecular weight, TPSA, rotatable bonds, leaving-group flags). Anchors carry measured ki labels; A-series targets are shown as unlabeled projections.

6.

6

Open in a new tab

Physics-informed potency ranking. (a) Predicted versus measured logki for the Gaussian-process model (leave-one-out CV, ρ = + 0.55); the shaded band is the model predictive uncertainty. (b) Single-descriptor signs and importances, with TPSA and phosphorus charge dominant. The signal is modest but reproducible, and the model’s main added value is calibrated uncertainty rather than apparent point accuracy.

3.4. Stereochemical Test and A-Series Projection

The strongest circularity concern is that the only configuration-resolved VX pair could be memorized. We therefore withheld both VX enantiomers together, retrained on the remaining anchors, and predicted the VX pair out of sample. The model preserved the experimental S P > R P ordering (predicted logk i 8.50 versus 6.48, against measured 8.15 and 6.08; Figure c,d) and approximated the observed gap, whereas docking scores remained nearly degenerate for enantiomers (Figure c). The predictive interval on the R P isomer is wide (±2.33 compared with ±0.81 for S P), so the recovered ordering is reported as exploratory rather than as demonstrated stereochemical predictivity. This result is encouraging but not definitive: a broader set of resolved stereochemical kinetics would be required to establish general stereochemical predictivity.

Projection to A-series structures is reported as an uncertainty-bounded prior for experimental prioritization, in the same spirit as uncertainty-aware screening frameworks used elsewhere for hazard triage. Targets outside the descriptor-space AD are flagged rather than overinterpreted (quantified in the Supporting Information). No synthesis, production, or operational guidance is provided; the released materials contain only literature kinetic anchors, descriptor matrices, docking/QM inputs and outputs, and analysis scripts.

3.5. Scope, Robustness, and Limitations

Three limitations remain central. First, n = 14 is small; individual rank correlations therefore carry wide confidence intervals and limited statistical power, and with so few points, a single influential inhibitor can inflate or deflate an apparent correlation, so all claims are framed as exploratory, hypothesis-generating rank-signal comparisons with explicit uncertainty rather than as definitive predictive performance. This sample size is a structural constraint of the domain, in which stereospecific human-AChE kinetics for the most hazardous agents are largely restricted or unavailable. Second, most available anchors are not configuration-resolved, so stereochemical inference is supported by one externalized VX pair and by mechanistic plausibility rather than by a large resolved set. Third, the quantum-cluster result does not replace future QM/MM free-energy calculations of the complete general-base path, of the kind reported for soman phosphonylation of AChE. We also note that, whereas the phosphorus charge and frontier-orbital descriptors are robust to conformer choice, the molecular dipole is conformer-sensitive; the lowest-energy conformer is therefore used consistently (Supporting Information). These limitations cap the claims, but they do not change the benchmark observation: in the available human-AChE kinetic panel, affinity-based docking lacks a reliable potency-ranking signala finding that is stable across box translations, random seeds, active-site waters, and covalent-compatible pose selectionwhereas reaction-aware descriptors retain a modest, interpretable one.

4. Conclusions

For irreversible AChE inhibitors, affinity-based docking should not be assumed to rank the potency without kinetic validation. In this curated benchmark, Vina, GNINA CNN, and explicit covalent docking all have rank-correlation confidence intervals that cross zero. This negative benchmark is the more robust of our findings: it holds across box translations, random seeds, retained crystallographic waters, and covalent-compatible (near-attack) pose selections. Mechanism-oriented descriptors provide a small but reproducible signal, support an out-of-sample VX stereochemical test, and yield uncertainty-aware priors for data-scarce covalent inhibitors; because the panel is small (n = 14), this descriptor signal is presented as exploratory and should be retested on larger covalent datasets as they become available. The practical recommendation is conservative: use docking for pose generation and sanity checks but rank irreversible anticholinesterase potency with reaction-relevant descriptors, explicit uncertainty, and transparent applicability-domain flags.

Supplementary Material

ao6c08355_si_001.pdf (655.1KB, pdf)

Acknowledgments

This research was supported by the Korea Environmental Industry & Technology Institute (KEITI) through the “Field-Driven Environmental Technology Commercialization R&D Program” (RS-2026-25524751), funded by the Ministry of Climate, Energy and Environment of the Republic of Korea.

Data and Software Availability: All experimental anchors, quantum-chemical descriptors, docking inputs and outputs, quantum-cluster scan summaries, and analysis and figure-generation scripts are publicly available at https://github.com/bisu9082/covalent5 (a permanent Zenodo archive with a citable DOI will be minted from this repository on acceptance). A top-level README maps each reported statistic and figure to the script and command that regenerate it, and verify_repo_layout.py checks the deposited tree against that inventory and fails on any mismatch. The reproducibility package is intended for scientific validation only and excludes synthesis, production, dissemination, or weaponization information for chemical warfare agents.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.6c08355.

  • Dataset curation, including the complete table of the 14 kinetic anchors and a figure of their structures; ligand preparation and conformer sensitivity of the quantum descriptors; docking robustness checks, including a figure of the search box and near-attack geometry, box/seed sensitivity, and crystallographic active-site waters; covalent-compatible (near-attack) pose analysis with per-compound distances; the composition of the quantum-cluster model and the complete relaxed-scan surface; descriptor definitions; circularity-control tests; applicability-domain analysis; and reference/DOI verification notes (PDF)

J.K.: Methodology, Software, Formal analysis, Investigation, Visualization, Writingoriginal draft. J.Y.: Investigation, Data curation, Validation, Writingreview and editing. D.-H.L.: Investigation, Resources, Validation, Writingreview and editing. K.K.: Conceptualization, Supervision, Project administration, Writingreview and editing.

The authors declare no competing financial interest.

References

  1. Costanzi S., Machado J.-H., Mitchell M.. Nerve Agents: What They Are, How They Work, How to Counter Them. ACS Chem. Neurosci. 2018;9:873–885. doi: 10.1021/acschemneuro.8b00148. [DOI] [PubMed] [Google Scholar]
  2. Ganesan K., Raza S. K., Vijayaraghavan R.. Chemical warfare agents. J. Pharm. Bioallied Sci. 2010;2:166–178. doi: 10.4103/0975-7406.68498. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Sidell F. R., Borak J.. Chemical warfare agents: II. nerve agents. Ann. Emerg. Med. 1992;21:865–871. doi: 10.1016/S0196-0644(05)81036-4. [DOI] [PubMed] [Google Scholar]
  4. Bajgar J.. Organophosphates/nerve agent poisoning: mechanism of action, diagnosis, prophylaxis, and treatment. Adv. Clin. Chem. 2004;38:151–216. doi: 10.1016/s0065-2423(04)38006-6. [DOI] [PubMed] [Google Scholar]
  5. Munro N. B., Talmage S. S., Griffin G. D., Waters L. C., Watson A. P., King J. F., Hauschild V.. The sources, fate, and toxicity of chemical warfare agent degradation products. Environ. Health Perspect. 1999;107:933–974. doi: 10.1289/ehp.99107933. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Casida J. E., Durkin K. A.. Neuroactive insecticides: targets, selectivity, resistance, and secondary effects. Annu. Rev. Entomol. 2013;58:99–117. doi: 10.1146/annurev-ento-120811-153645. [DOI] [PubMed] [Google Scholar]
  7. Kim J., Kang K., Shin M., Kim S., Yoo J., Kim M.-K., Lee W. B., Lee D.-H.. Machine learning-enhanced colorimetric sensor array for rapid detection of nerve agents. J. Hazard. Mater. 2026;501:140703. doi: 10.1016/j.jhazmat.2025.140703. [DOI] [PubMed] [Google Scholar]
  8. Sussman J. L., Harel M., Frolow F., Oefner C., Goldman A., Toker L., Silman I.. Atomic structure of acetylcholinesterase from Torpedo californica: a prototypic acetylcholine-binding protein. Science. 1991;253:872–879. doi: 10.1126/science.1678899. [DOI] [PubMed] [Google Scholar]
  9. Cheung J., Rudolph M. J., Burshteyn F., Cassidy M. S., Gary E. N., Love J., Franklin M. C., Height J. J.. Structures of human acetylcholinesterase in complex with pharmacologically important ligands. J. Med. Chem. 2012;55:10282–10286. doi: 10.1021/jm300871x. [DOI] [PubMed] [Google Scholar]
  10. Elhanany E., Ordentlich A., Dgany O., Kaplan D., Segall Y., Barak R., Velan B., Shafferman A.. Resolving pathways of interaction of covalent inhibitors with the active site of acetylcholinesterases: MALDI-TOF/MS analysis of various nerve agent phosphyl adducts. Chem. Res. Toxicol. 2001;14:912–918. doi: 10.1021/tx0100542. [DOI] [PubMed] [Google Scholar]
  11. Barak D., Ordentlich A., Kaplan D., Barak R., Mizrahi D., Kronman C., Segall Y., Velan B., Shafferman A.. Evidence for P–N Bond Scission in Phosphoroamidate Nerve Agent Adducts of Human Acetylcholinesterase. Biochemistry. 2000;39:1156–1161. doi: 10.1021/bi992009n. [DOI] [PubMed] [Google Scholar]
  12. Worek F., Thiermann H., Szinicz L., Eyer P.. Kinetic analysis of interactions between human acetylcholinesterase, structurally different organophosphorus compounds and oximes. Biochem. Pharmacol. 2004;68:2237–2248. doi: 10.1016/j.bcp.2004.07.038. [DOI] [PubMed] [Google Scholar]
  13. Bester S. M., Guelta M. A., Cheung J., Winemiller M. D., Bae S. Y., Myslinski J., Pegan S. D., Height J. J.. Structural Insights of Stereospecific Inhibition of Human Acetylcholinesterase by VX and Subsequent Reactivation by HI-6. Chem. Res. Toxicol. 2018;31:1405–1417. doi: 10.1021/acs.chemrestox.8b00294. [DOI] [PubMed] [Google Scholar]
  14. Bassani D., Pavan M., Bolcato G., Sturlese M., Moro S.. Re-Exploring the Ability of Common Docking Programs to Correctly Reproduce the Binding Modes of Non-Covalent Inhibitors of SARS-CoV-2 Protease Mpro. Pharmaceuticals. 2022;15:180. doi: 10.3390/ph15020180. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Benschop H. P., De Jong L. P. A.. Nerve agent stereoisomers: analysis, isolation, and toxicology. Acc. Chem. Res. 1988;21:368–374. doi: 10.1021/ar00154a003. [DOI] [Google Scholar]
  16. Berman H. A., Decker M. M.. Chiral nature of covalent methylphosphonyl conjugates of acetylcholinesterase. J. Biol. Chem. 1989;264:3951–3956. doi: 10.1016/S0021-9258(19)84945-1. [DOI] [PubMed] [Google Scholar]
  17. Berman H. A., Leonard K.. Chiral reactions of acetylcholinesterase probed with enantiomeric methylphosphonothioates: noncovalent determinants of enzyme chirality. J. Biol. Chem. 1989;264:3942–3950. doi: 10.1016/S0021-9258(19)84944-X. [DOI] [PubMed] [Google Scholar]
  18. Leng J., Huang Z., Zheng L., Yang Y., Huang H., Gao Y., Huo L., Li Y., Sun Z., Zhang J. Z. H.. Revealing the limits of covalent docking and advancing affinity prediction with covalent-aware multi-task learning. Phys. Chem. Chem. Phys. 2026;28:4822–4834. doi: 10.1039/D5CP04981D. [DOI] [PubMed] [Google Scholar]
  19. Zhu K., Borrelli K. W., Greenwood J. R., Day T., Abel R., Farid R. S., Harder E.. Docking Covalent Inhibitors: A Parameter Free Approach To Pose Prediction and Scoring. J. Chem. Inf. Model. 2014;54:1932–1940. doi: 10.1021/ci500118s. [DOI] [PubMed] [Google Scholar]
  20. London N., Miller R. M., Krishnan S., Uchida K., Irwin J. J., Eidam O., Gibold L., Cimermančič P., Bonnet R., Shoichet B. K., Taunton J.. Covalent docking of large libraries for the discovery of chemical probes. Nat. Chem. Biol. 2014;10:1066–1072. doi: 10.1038/nchembio.1666. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Scarpino A., Ferenczy G. G., Keserü G. M.. Comparative Evaluation of Covalent Docking Tools. J. Chem. Inf. Model. 2018;58:1441–1458. doi: 10.1021/acs.jcim.8b00228. [DOI] [PubMed] [Google Scholar]
  22. Scarpino A., Ferenczy G. G., Keserü G. M.. Covalent Docking in Drug Discovery: Scope and Limitations. Curr. Pharm. Des. 2020;26:5684–5699. doi: 10.2174/1381612824999201105164942. [DOI] [PubMed] [Google Scholar]
  23. Sotriffer C.. Docking of Covalent Ligands: Challenges and Approaches. Mol. Inf. 2018;37:1800062. doi: 10.1002/minf.201800062. [DOI] [PubMed] [Google Scholar]
  24. Ai Y., Yu L., Tan X., Chai X., Liu S.. Discovery of Covalent Ligands via Noncovalent Docking by Dissecting Covalent Docking Based on a “Steric-Clashes Alleviating Receptor (SCAR)” Strategy. J. Chem. Inf. Model. 2016;56:1563–1575. doi: 10.1021/acs.jcim.6b00334. [DOI] [PubMed] [Google Scholar]
  25. Muratov E. N.. et al. QSAR without borders. Chem. Soc. Rev. 2020;49:3525–3564. doi: 10.1039/D0CS00098A. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Tropsha A.. Best Practices for QSAR Model Development, Validation, and Exploitation. Mol. Inf. 2010;29:476–488. doi: 10.1002/minf.201000061. [DOI] [PubMed] [Google Scholar]
  27. Sahigara F., Mansouri K., Ballabio D., Mauri A., Consonni V., Todeschini R.. Comparison of different approaches to define the applicability domain of QSAR models. Molecules. 2012;17:4791–4810. doi: 10.3390/molecules17054791. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Roy K., Das R. N., Ambure P., Aher R. B.. Be aware of error measures. Further studies on validation of predictive QSAR models. Chemom. Intell. Lab. Syst. 2016;152:18–33. doi: 10.1016/j.chemolab.2016.01.008. [DOI] [Google Scholar]
  29. Judson R., Richard A., Dix D. J., Houck K., Martin M., Kavlock R., Dellarco V., Henry T., Holderman T., Sayre P., Tan S., Carpenter T., Smith E.. The toxicity data landscape for environmental chemicals. Environ. Health Perspect. 2009;117:685–695. doi: 10.1289/ehp.0800168. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Richard A. M.. et al. ToxCast Chemical Landscape: Paving the Road to 21st Century Toxicology. Chem. Res. Toxicol. 2016;29:1225–1251. doi: 10.1021/acs.chemrestox.6b00135. [DOI] [PubMed] [Google Scholar]
  31. Thomas R. S., Paules R. S., Simeonov A., Fitzpatrick S. C., Crofton K. M., Casey W. M., Mendrick D. L.. The US Federal Tox21 Program: A strategic and operational plan for continued leadership. ALTEX. 2018;35:163–168. doi: 10.14573/altex.1803011. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Vignaux P. A., Lane T. R., Urbina F., Gerlach J., Puhl A. C., Snyder S. H., Ekins S.. Validation of Acetylcholinesterase Inhibition Machine Learning Models for Multiple Organophosphates and Carbamates. Chem. Res. Toxicol. 2023;36:188–201. doi: 10.1021/acs.chemrestox.2c00283. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Karim A., Mishra A., Newton M. A. H., Sattar A.. Efficient toxicity prediction via simple features using shallow neural networks and decision trees. ACS Omega. 2019;4:1874–1888. doi: 10.1021/acsomega.8b03173. [DOI] [Google Scholar]
  34. Zang Q., Rotroff D. M., Judson R. S.. Binary Classification of a Large Collection of Environmental Chemicals from Estrogen Receptor Assays by Quantitative Structure–Activity Relationship and Machine Learning Methods. J. Chem. Inf. Model. 2013;53:3244–3261. doi: 10.1021/ci400527b. [DOI] [PubMed] [Google Scholar]
  35. Jeong K., Lee J.-Y., Woo S., Kim D., Jeon Y., Ryu T. I., Hwang S.-R., Jeong W.-H.. Vapor Pressure and Toxicity Prediction for Novichok Agent Candidates Using Machine Learning Model: Preparation for Unascertained Nerve Agents after Chemical Weapons Convention Schedule 1 Update. Chem. Res. Toxicol. 2022;35:774–781. doi: 10.1021/acs.chemrestox.1c00410. [DOI] [PubMed] [Google Scholar]
  36. Jang Y., Kim J., Lee D.-H., Yoo J., Park J., Kang K.. Hybrid QSAR modeling of LD50 in organophosphorus nerve agents: a mechanistic approach using DFT and molecular docking. Comput. Toxicol. 2025;35:100363. doi: 10.1016/j.comtox.2025.100363. [DOI] [Google Scholar]
  37. Shin M., Kang K., Kim J., Kim S., Yoo J., Jang Y. J., Kim M.-K., Anslyn E. V., Lee D.-H.. Portable machine learning–integrated optical sensor array for rapid and selective detection of authentic Novichok nerve agents. Sens. Actuators, B. 2026;457:139706. doi: 10.1016/j.snb.2026.139706. [DOI] [Google Scholar]
  38. Trott O., Olson A. J.. AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. J. Comput. Chem. 2010;31:455–461. doi: 10.1002/jcc.21334. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Eberhardt J., Santos-Martins D., Tillack A. F., Forli S.. AutoDock Vina 1.2.0: New Docking Methods, Expanded Force Field, and Python Bindings. J. Chem. Inf. Model. 2021;61:3891–3898. doi: 10.1021/acs.jcim.1c00203. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. McNutt A. T., Francoeur P., Aggarwal R., Masuda T., Meli R., Ragoza M., Sunseri J., Koes D. R.. GNINA 1.0: molecular docking with deep learning. J. Cheminf. 2021;13:43. doi: 10.1186/s13321-021-00522-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Neese F.. The ORCA program system. WIREs Comput. Mol. Sci. 2012;2:73–78. doi: 10.1002/wcms.81. [DOI] [Google Scholar]
  42. Neese F., Wennmohs F., Becker U., Riplinger C.. The ORCA quantum chemistry program package. J. Chem. Phys. 2020;152:224108. doi: 10.1063/5.0004608. [DOI] [PubMed] [Google Scholar]
  43. Chai J.-D., Head-Gordon M.. Long-range corrected hybrid density functionals with damped atom-atom dispersion corrections. Phys. Chem. Chem. Phys. 2008;10:6615–6620. doi: 10.1039/b810189b. [DOI] [PubMed] [Google Scholar]
  44. Grimme S., Antony J., Ehrlich S., Krieg H.. A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H-Pu. J. Chem. Phys. 2010;132:154104. doi: 10.1063/1.3382344. [DOI] [PubMed] [Google Scholar]
  45. Weigend F., Ahlrichs R.. Balanced basis sets of split valence, triple zeta valence and quadruple zeta valence quality for H to Rn. Phys. Chem. Chem. Phys. 2005;7:3297–3305. doi: 10.1039/b508541a. [DOI] [PubMed] [Google Scholar]
  46. Barone V., Cossi M.. Quantum calculation of molecular energies and energy gradients in solution by a conductor solvent model. J. Phys. Chem. A. 1998;102:1995–2001. doi: 10.1021/jp9716997. [DOI] [Google Scholar]
  47. Rogers D., Hahn M.. Extended-connectivity fingerprints. J. Chem. Inf. Model. 2010;50:742–754. doi: 10.1021/ci100050t. [DOI] [PubMed] [Google Scholar]
  48. Ertl P., Rohde B., Selzer P.. Fast calculation of molecular polar surface area as a sum of fragment-based contributions and its application to the prediction of drug transport properties. J. Med. Chem. 2000;43:3714–3717. doi: 10.1021/jm000942e. [DOI] [PubMed] [Google Scholar]
  49. Wildman S. A., Crippen G. M.. Prediction of physicochemical parameters by atomic contributions. J. Chem. Inf. Comput. Sci. 1999;39:868–873. doi: 10.1021/ci990307l. [DOI] [Google Scholar]
  50. Williams A. J., Grulke C. M., Edwards J., McEachran A. D., Mansouri K., Baker N. C., Patlewicz G., Shah I., Wambaugh J. F., Judson R. S.. et al. The CompTox Chemistry Dashboard: a community data resource for environmental chemistry. J. Cheminf. 2017;9:61. doi: 10.1186/s13321-017-0247-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Siegbahn P. E. M., Himo F.. The quantum chemical cluster approach for modeling enzyme reactions. WIREs Comput. Mol. Sci. 2011;1:323–336. doi: 10.1002/wcms.13. [DOI] [Google Scholar]
  52. Himo F.. Recent Trends in Quantum Chemical Modeling of Enzymatic Reactions. J. Am. Chem. Soc. 2017;139:6780–6786. doi: 10.1021/jacs.7b02671. [DOI] [PubMed] [Google Scholar]
  53. Himo F., de Visser S. P.. Status report on the quantum chemical cluster approach for modeling enzyme reactions. Commun. Chem. 2022;5:29. doi: 10.1038/s42004-022-00642-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  54. Breiman L.. Random Forests. Mach. Learn. 2001;45:5–32. doi: 10.1023/A:1010933404324. [DOI] [Google Scholar]
  55. Lundberg S. M., Erion G., Chen H., DeGrave A., Prutkin J. M., Nair B., Katz R., Himmelfarb J., Bansal N., Lee S.-I.. From local explanations to global understanding with explainable AI for trees. Nat. Mach. Intell. 2020;2:56–67. doi: 10.1038/s42256-019-0138-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  56. Gawlikowski J., Tassi C. R. N., Ali M., Lee J., Humt M., Feng J., Kruspe A., Triebel R., Jung P., Roscher R., Shahzad M., Yang W., Bamler R., Zhu X. X.. A survey of uncertainty in deep neural networks. Artif. Intell. Rev. 2023;56:1513–1589. doi: 10.1007/s10462-023-10562-9. [DOI] [Google Scholar]
  57. Kang K., Kim J., Jang Y. J., Shin M., Kim S., Yoo J.. Open-set combustion-fingerprint classification with out-of-distribution alerting for environmental fire-hazard screening. J. Hazard. Mater. 2026;516:143199. doi: 10.1016/j.jhazmat.2026.143199. [DOI] [PubMed] [Google Scholar]
  58. Kim J., Bae S., Jang Y. J., Bae J., Lee D.-H., Lee D., Lee W. B., Kang K.. Integrated multi-endpoint hazard prioritization of unregulated industrial contaminants: A safety-first screening framework for emerging environmental threats. Ecotoxicol. Environ. Saf. 2026;319:120283. doi: 10.1016/j.ecoenv.2026.120283. [DOI] [PubMed] [Google Scholar]
  59. Sirin G. S., Zhang Y.. How Is Acetylcholinesterase Phosphonylated by Soman? An Ab Initio QM/MM Molecular Dynamics Study. J. Phys. Chem. A. 2014;118:9132–9139. doi: 10.1021/jp502712d. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

ao6c08355_si_001.pdf (655.1KB, pdf)

Data Availability Statement

Data and Software Availability: All experimental anchors, quantum-chemical descriptors, docking inputs and outputs, quantum-cluster scan summaries, and analysis and figure-generation scripts are publicly available at https://github.com/bisu9082/covalent5 (a permanent Zenodo archive with a citable DOI will be minted from this repository on acceptance). A top-level README maps each reported statistic and figure to the script and command that regenerate it, and verify_repo_layout.py checks the deposited tree against that inventory and fails on any mismatch. The reproducibility package is intended for scientific validation only and excludes synthesis, production, dissemination, or weaponization information for chemical warfare agents.

Articles from ACS Omega are provided here courtesy of American Chemical Society

RESOURCES