Exploring the Physicochemical and Toxicological Study of G-Series and A-Series Agents Combining Molecular Dynamics and Quantitative Structure–Activity Relationship

Abstract

This study explores the physicochemical and toxicological properties of six G-series and A-series chemical warfare agents (Sarin, Soman, Tabun, A230, A232, and A234) using an integrated computational approach combining molecular dynamics (MD) simulations and Quantitative Structure–Activity Relationship (QSAR) modeling. For the A-series nerve agents, both Ellison–Hoenig and Mirzayanov structural proposals were examined. MD simulations (10 ns, NPT ensemble) provided key thermodynamic properties, including density, molar heat capacity, and diffusivity. Simulated densities for G-agents (e.g., Sarin: 1.09 g/cm³, Soman: 1.03 g/cm³) and A-agents (e.g., A230: 1.608 g/cm³, Ellison–Hoenig model) closely matched experimental data. Heat capacities ranged from 258 to 462 J/mol·K, and self-diffusion coefficients revealed lower mobility for A-agents, especially under the Ellison–Hoenig configurations. QSAR modeling focused on lipophilicity (LogP) and acute toxicity (LD₅₀). Predicted LD₅₀ values ranged from 0.012 to 0.017 mg/kg for G-agents and up to 1.23 mg/kg for A-agents. A-234 showed the highest lipophilicity (LogP = 2.97) and toxicity (LD₅₀ = 0.51 mg/kg) within its group. Additional descriptors, such as molecular weight and polar surface area, supported toxicity predictions. Strong correlations emerged between MD-derived properties and QSAR outputs, validating the integrated approach. The combined use of MD and QSAR techniques provided a comprehensive view of the agents’ environmental behavior and toxicological impact, supporting safer assessment strategies and reinforcing the importance of multidisciplinary modeling for chemical threat mitigation.

1. Introduction

Chemical warfare agents (CWAs) have historically posed substantial threats to global security and public health. Among these, the G-series nerve agents—Sarin (GB), Soman (GD), and Tabun (GA)—are particularly well known for their high toxicity, rapid onset of action, and documented use in conflicts. The A-series nerve agents exhibit similar mechanisms of action but are considered even more lethal, harder to detect, and more resistant to standard antidotes. These organophosphorus compounds act as potent inhibitors of acetylcholinesterase (AChE), leading to the accumulation of acetylcholine at neuromuscular junctions. This biochemical disruption causes severe cholinergic symptoms, respiratory failure, and potentially death if not promptly treated [1,2,3,4].

Despite international bans under treaties such as the Chemical Weapons Convention (CWC), both G-series and A-series agents remain a concern due to the risk of terrorist use and asymmetric warfare. The 1995 Tokyo subway attack involving Sarin underscores the urgent need for robust strategies to understand and mitigate the effects of these toxic substances [5]. As experimental studies are limited by ethical and safety concerns, in-depth computational research into their physicochemical and toxicological characteristics is essential to support decontamination, medical countermeasures, and risk assessment. Quantitative Structure–Activity Relationship (QSAR) modeling and molecular dynamics (MD) simulations have become invaluable tools in this context [6,7,8].

QSAR approaches develop statistical models that relate chemical structure to biological activity or toxicological endpoints. Molecular descriptors such as hydrophobicity, electronic distribution, and steric parameters are used to predict toxic effects and reduce reliance on animal testing [9]. Within the context of G-series and A-series agents, QSAR models have revealed key structural features linked to toxicity. By analyzing structural analogs and toxicological outcomes, researchers can identify motifs that enhance or mitigate harmful effects, thereby guiding the design of safer analogs and effective antidotes [9,10,11].

The selection of appropriate molecular descriptors is critical for QSAR model accuracy. Techniques such as principal component analysis and genetic algorithms optimize descriptor sets and improve model performance [12]. This is particularly valuable when studying highly toxic agents like G-series and A-series compounds, where experimental validation is limited [6,7].

In parallel, MD simulations provide dynamic insights into molecular interactions at the atomic level. They help elucidate the binding mechanisms of nerve agents with AChE, highlighting how small structural changes affect binding affinity and stability [6,7,12]. MD simulations have also supported QSAR-based virtual screening efforts in identifying antidotes that reverse AChE inhibition [13,14].

MD simulations have further advanced understanding of organophosphate interactions within biological systems. Studies have investigated solvation dynamics, enzyme binding, and the influence of mutations on susceptibility [15,16,17].

The convergence between MD and QSAR approaches enhances prediction confidence. MD simulations refine QSAR-based predictions by modeling dynamic molecular behavior in biologically relevant conditions. Identifying critical binding residues and conformational changes validates structure–toxicity relationships and reinforces the reliability of computational assessments [6,7,18].

Experimental validation remains essential for confirming computational predictions. In vitro assays, such as recombinant AChE inhibition tests, have verified QSAR-predicted toxicity of organophosphates [19]. Additionally, crystallographic studies have confirmed MD-predicted binding modes, demonstrating high consistency between computationally derived and observed structural conformations [20].

2. Materials and Methods

2.1. Molecular Dynamics Simulations—Setup and Boundary Conditions

In the present study, molecular dynamics (MD) simulations were performed using the LAMMPS simulation package, an efficient and scalable tool widely employed for studying condensed-phase systems [21]. The simulations aimed to calculate key thermodynamic properties, density, heat capacity, and diffusivity of G-series and A-series organophosphorus nerve agents in their bulk liquid phase.

Force field parameters were refined and validated against both literature values and experimental data to ensure accurate representation of intermolecular interactions [22]. Lennard-Jones potentials were used for modeling van der Waals forces, while electrostatic interactions were modeled via Coulombic terms [13,14]. Force field van der Waals, bond, angle, and dihedral properties were selected based on OPLS set of parameters [23], specifically the Unified Atom approximation set and the All-Atom approximation set. The aforementioned OPLS-UA force field is parameterized for stable, non-reactive systems and does not include bond-breaking or bond-forming events. It should also be noted that reactive MD or hybrid QM/MM techniques were not employed in this work. While the OPLS-UA force field simplifies hydrogen-heavy groups through a united-atom approach, it still reliably captures key nonbonded interactions, steric effects, and bulk liquid behavior. This makes it suitable for evaluating macroscopic properties (e.g., density and vapor–liquid equilibria) across chemically diverse agents. However, we acknowledge that some stereochemical and hydrogen-specific effects (especially relevant for AChE binding or enantioselectivity) may not be fully resolved under the united-atom approximation. The initial geometries of all molecules were optimized using Avogadro software (v. 1.2.0) [24] with the steepest descent algorithm.

For each of the six compounds studied, including two distinct structural variants proposed for each A-series agent, two partial charge assignment methods were employed to generate model variants. These were the following: (1) the extended charge equilibration (eQeQ) method [25], which utilizes quantum chemical principles, and (2) the contraDRG method [26], which leverages machine learning (Random Forest algorithms) for predicting partial atomic charges. This dual approach ensured robustness and allowed for comparisons between quantum-based and data-driven charge derivation techniques. A summary of the key simulation parameters is presented in

Table 1. Simulation parameters.

Parameter	Value	Description
Temperature	298 K	Nosé–Hoover thermostat
Pressure	1 atm	Nosé–Hoover barostat
Time Step	1 fs	Integration interval
Simulation Length	10 ns	Total simulation time
Force Field	Lennard-Jones, Coulombic	Refined for G-series agents

.

All simulations were conducted under the isothermal–isobaric (NPT) ensemble at 298 K and 1 atm, employing Nosé–Hoover thermostats and barostats to regulate system temperature and pressure. A 1-femtosecond (fs) integration time step was used, and the simulation duration was set at 10 nanoseconds (ns). This length was selected based on a consensus in the literature indicating that equilibrium thermodynamic properties of similarly sized organophosphorus systems can be reliably calculated within 5–10 ns, depending on the complexity of the molecule and the refinement of the force field parameters [22,27]. A nanosecond scale for simulation length time is considered appropriate for production runs since it allows properties to be quite successfully sampled along molecular system trajectories for the 500 molecules employed in our simulations. The study by Emelianova et al. further demonstrated that simulations on the nanosecond scale are effective for achieving convergence in both intermolecular interactions and macroscopic thermodynamic quantities [22]. The 10 ns duration used in the present study provided a balance between computational feasibility and statistical reliability.

To confirm equilibrium and statistical convergence, analyses of mean square displacement (MSD) and radial distribution functions (RDFs) were performed at regular intervals throughout the simulation. These analyses confirmed that molecular properties stabilized well within the 10 ns window, validating the simulation length. Moreover, density values derived from MD simulations were employed mostly for comparison against available experimental data in the cases of G- and A-agents. Additionally, these values were important in estimating whether employed molecular potential models possess the desirable reliability and were strategically employed in such estimations, such as the standardization method of OPLS potential developed by Jorgensen et al. [23].

Periodic boundary conditions (PBCs) were applied in all three Cartesian directions to simulate an effectively infinite system and avoid edge effects. The use of PBCs is essential to prevent artificial interactions between molecular images across boundaries. The simulation boxes were sized to ensure that long-range electrostatic interactions decayed sufficiently within the cutoff radius. Electrostatics were calculated using the Particle Mesh Ewald (PME) method [27,28], which allows accurate treatment of long-range Coulombic interactions in periodic systems. Lennard-Jones interactions were truncated at a cutoff distance of 14 Å, as recommended for systems dominated by short- to medium-range van der Waals forces [27].

Previous computational studies involving polar liquids and nerve agents have confirmed that this simulation protocol—with PBCs, PME electrostatics, and appropriately sized simulation cells—yields realistic bulk-phase behavior [28]. Employing the NPT ensemble also allowed for volume fluctuations, ensuring realistic density predictions and dynamic behavior representative of liquid-phase conditions [27].

The structural characteristics of the studied agents were carefully considered. Sarin, Soman, and Tabun are classical organophosphoric compounds featuring a phosphorus atom bonded to an organic substituent. Sarin and Soman contain a fluorine atom bonded directly to the phosphorus center, while Tabun is unique in that it features a nitrile (–C≡N) group forming a linear P–C–N motif. This specific geometry has notable implications for its physicochemical behavior. The molecular structures of the studied G-series and A-series agents are illustrated in Figure 1.

For the A-series agents, both the Ellison–Hoenig and Mirzayanov structural variants were modeled. The Ellison–Hoenig structures resemble traditional G-series agents, with phosphorus forming phosphorofluoridate linkages and retaining a tetrahedral configuration. The key structural variations among A-230, A-232, and A-234 lie in their organic substituents, which influence volatility and reactivity. In contrast, the Mirzayanov structures introduce a phosphoramidofluoridate backbone, where phosphorus is directly bonded to a nitrogen atom. This substitution significantly alters the electronic environment of phosphorus, potentially affecting hydrolysis, binding affinity to AChE, and enzymatic stability. Modeling both variants allowed for a comprehensive assessment of their physicochemical behaviors.

2.2. QSAR Modeling—Software Implementation and Workflow

QSAR modeling was conducted to evaluate the lipophilicity and potential toxicity of the G-series and A-series agents, focusing primarily on the octanol–water partition coefficient (logP), which influences skin absorption and systemic distribution. Additional descriptors such as molecular weight, topological polar surface area (TPSA), number of hydrogen bond donors/acceptors, and rotatable bonds were included to enhance model robustness.

Molecular descriptors were calculated using the Finite Dose Skin Permeation (FDSP) (v. 2.4_09.09.2010) software, QSAR Toolbox (v. 4.5), and PASS Online (2024). These tools allowed cross-validation of descriptor values and incorporation of mechanistically relevant endpoints. External datasets were incorporated to train and validate the QSAR models. Datasets were curated from publicly available toxicological databases, including EPA DSSTox [29], TOXNET [30], PubChem BioAssay [31], and ChEMBL [32]. These databases provided experimentally validated descriptors and toxicity endpoints relevant to organophosphorus compounds.

The finalized dataset comprised 95 unique chemical structures, selected through rigorous curation procedures. Only compounds with experimentally verified octanol–water partition coefficients (logP) and acute toxicity data (e.g., LD₅₀) were retained to ensure high data quality and reliability.

Preprocessing steps included the removal of incomplete, redundant, or structurally ambiguous entries. Data normalization ensured consistency across descriptors and improved comparability across sources.

The dataset was randomly divided, allocating 80% for model training and 20% for external validation. Predictive models were developed using Random Forest (RF), Support Vector Machines (SVMs), and Gradient Boosting algorithms. The goal was to establish robust relationships between molecular descriptors and biological outcomes.

Model performance was evaluated using statistical metrics including R² (coefficient of determination), RMSE (root mean square error), and Q² (external validation coefficient). The final model achieved an R² of 0.89 and RMSE of 0.42, indicating high predictive performance. External validation on structurally similar nerve agents not included in the training set showed an average prediction deviation of ±10% from known experimental LD50 values [33].

To ensure statistical significance and minimize overfitting, Y-randomization tests and leave-one-out cross-validation (LOO-CV) were conducted. Furthermore, the model’s applicability domain was evaluated using independent entries from TOXNET and ChEMBL, confirming its utility for preliminary toxicological screening of novel chemical warfare agents.

3. Results

3.1. Molecular Dynamics Simulations Results

MD simulations provided detailed insights into the thermodynamic and dynamic properties of the bulk liquid systems of G-series and A-series agents. For the G-series agents, the simulated densities were validated through comparison with available experimental data from previous studies [22,23]. The experimentally reported densities for Sarin, Soman, and Tabun were 1.10 g/cm³, 1.02 g/cm³, and 1.07 g/cm³, respectively [2,13]. As shown in

Table 2. Simulation results for thermodynamic properties G-series.

Simulation	Sarin eQeQ	Sarin ML	Soman eQeQ	Soman ML	Tabun eQeQ	Tabun ML
ρMD (g/cm3)	1.09 ± 0.004	1.08 ± 0.005	1.03 ± 0.004	1.05 ± 0.003	1.07 ± 0.007	1.16 ± 0.01
ρexp (g/cm3) [13]	1.10		1.02		1.07
ρCalculated-Chemspider (g/mL) (in silico)	1.09 ± 0.1		1.02 ± 0.1		1.08 ± 0.1

and Figure 2, the MD-derived values exhibit good agreement with these references across both charge models (eQeQ and ML), supporting the accuracy of the employed force fields.

Sarin and Soman show minimal variation between the two charge models, with calculated densities closely matching experimental values within 0.01–0.02 g/cm³. This suggests that both the eQeQ and ML charge assignment schemes are suitable for simulating the bulk-phase behavior of these agents. In contrast, Tabun shows a more significant discrepancy between the two MD charge models, with the ML-based simulation yielding a higher density (1.16 g/cm³), deviating from both the experimental value (1.07 g/cm³) and the ChemSpider in silico estimation (1.08 ± 0.1 g/mL). This may indicate that the ML-derived partial charges overestimate intermolecular interactions for Tabun, possibly due to the specific electronic environment of the agent’s substituents. In general, the eQeQ model results in calculated densities within the expected experimental uncertainty of 0.01–0.005 g/mL.

Overall, the consistency between MD-derived, experimental, and in silico densities confirms the reliability of the modeling approach for capturing key thermophysical properties of the G-series compounds. The comparison further underscores the importance of evaluating multiple charge models, especially for compounds with distinct functional groups or electronic characteristics.

For the A-series agents, the calculated density values were derived from extensive NPT molecular dynamics (MD) simulations using two distinct structural models: the phosphoramidofluoridate structures proposed by Mirzayanov and the phosphorofluoridate structures proposed by Ellison and Hoenig. The simulations employed two charge assignment schemes (eQeQ and ML) and were benchmarked against both experimental data and ChemSpider-derived in silico estimates to evaluate consistency and predictive accuracy.

As shown in

Table 3. Simulation results for thermodynamic properties A-series.

		eQeQ			ML
Substance	A230	A232	A234	A230	A232	A234
ρEXP (g/cm3)	1.612	1.515	1.414	1.612	1.515	1.414
ρMD (g/cm3) Mirzayanov	1.051 ± 0.01	1.089 ± 0.003	1.079 ± 0.006	1.114 ± 0.004	1.173 ± 0.01	1.154 ± 0.008
ρMD (g/cm3) Ellison-Hoenig	1.608 ± 0.006	1.561 ± 0.014	1.499 ± 0.012	1.696 ± 0.011	1.596 ± 0.006	1.492 ± 0.007
ρCalculated-Chemspider (g/mL) (in silico)	1.1 ± 0.1		1.1 ± 0.1		1.1 ± 0.1

and Figure 3, the experimental densities exhibit a decreasing trend from A230 (1.612 g/cm³) to A234 (1.414 g/cm³), likely reflecting structural modifications and reduced molecular packing efficiency with increasing molecular mass and steric complexity.

The MD simulations using Mirzayanov’s structures consistently yielded lower density values, regardless of the charge assignment method, with results falling between 1.05 and 1.17 g/cm³. These values are significantly below experimental references and suggest that Mirzayanov’s structures may underestimate the degree of molecular packing in the liquid phase. This is likely due to the presence of the direct phosphorus–nitrogen (P–N) bond, which modifies the electron distribution around the phosphorus center and alters intermolecular interaction potentials.

In contrast, simulations based on the Ellison–Hoenig structures produced densities much closer to experimental values. Particularly, under the eQeQ model, the calculated values show excellent agreement: e.g., A230 (1.608 g/cm³) compared to 1.612 g/cm³ experimentally. These results support the idea that the phosphorofluoridate representation better captures the structural and electronic environment of these agents, especially with respect to packing efficiency and cohesive energy density. The in silico values retrieved from ChemSpider were relatively consistent across all three agents, around 1.1 ± 0.1 g/mL. Summarizing results for A-agents, it is noticeable that Ellison–Hoenig structures exhibit calculated densities close to the experimental ones, whilst Mirzayanov structures are close to those calculated via Chemspider. Since experimental values are not reportedly attached to either Mirzayanov or Ellison–Hoenig structures, we conclude that they represent measurements of the Ellison–Hoenig structures.

The heat capacity (Cp) calculated values ranged from 258 to 263 J/mol·K for Sarin models to 343 to 350 J/mol·K for Soman models, with Tabun being in the middle range (

Table 4. Simulation results for specific heat under constant pressure for G-series agents.

Property	Sarin eQeQ	Sarin ML	Soman eQeQ	Soman ML	Tabun eQeQ	Tabun ML
Cp Molar Heat Capacity under constant pressure (J/mol · K)	258.95	262.59	350.50	343.37	305.50	304.30

). In general, eQeQ and ML models predict quite close values to each other for the same substance [26], values with differences inside the statistical noise window.

The specific heat capacity (Cp) for the A-series agents varied between 366.19 J/mol·K for A230 and 462.58 J/mol·K for A234 (

Table 5. Simulation results for specific heat under constant pressure for A-series agents.

		eeQeQ			MΜL
Substance	A230	A232	A234	A230	A232	A234
Cp (J/K × mol) Molar Heat Capacity under constant pressure Mirzayanoy	366.19	387.74	418.42	341.69	393.11	371.77
Cp (J/K × mol) Molar Heat Capacity under constant pressure Ellison -Hoenig [21]	396.25	424.22	462.58	358.51	385.20	450.34

). The values calculated for the A-series agents were relatively higher across the groups in comparison with those of the G-series, independently of the proposed structure. Furthermore, the differences between the Cp calculated values for the two alternative proposed structures for each of the A-series agents were also relatively significantly different for each compound (Table 5).

The self-diffusion coefficient was calculated with the Einstein relationship (Equation (1)) using data from mean square displacement of molecules over a wide range of time.

$$D=\underset{t\to \infty }{\lim }{\displaystyle \frac{1}{6}}{\displaystyle \frac{\mathsf{\partial }\left(MSD\right)}{\mathsf{\partial }t}}$$

(1)

Molecular diffusivity values obtained from this mean square displacement analysis demonstrated that Sarin has the highest diffusivity (0.82 × 10⁻⁹ m²/s), followed by Tabun and Soman (

Table 6. Self-diffusion coefficient for the G-series agents.

Property	Sarin eQeQ	Sarin ML	Soman eQeQ	Soman ML	Tabun eQeQ	Tabun ML
Ds (×10−9 m2/s)	0.84	0.89	0.17	0.09	0.21	0.01

). The A-series agents’ molecular diffusivity values were comparable to those calculated for Tabun when the Mirzayanov molecular structures were considered. The results for the A-series agents with the Ellison–Hoenig-proposed structure were significantly lower, with the minimum values presented for the A232 agent (

Table 7. Self-diffusion coefficient for the A-series agents.

	eQeQ				ML
Substance	A230	A232	A234	A230	A232	A234
Ds (10−9 m2/s) Mirzayanov	0.565	0.543	0.430	0.021	0.013	0.009
Ds (10−9 m2/s) Ellison–Hoenig	0.230	0.138	0.144	0.013	0.007	0.019

).

In the following, pair radial distribution functions (pRDFs) are presented and discussed. A pRDF, often denoted as g(r), is a statistical measurement of the average intemolecular distances in a distance r, between a reference particle and another particle, belonging to adjacent molecules. Therefore, in small distances, g(r) values are almost always zero, due to stereochemical reasons and repulsive forces, whilst in large distances, values equal unity, since the local density of specific particles equals the corresponding bulk density in the simulation cell. In intermediate distances, peaks over 1 signify a stronger connection of specific particles (i.e., selected molecular sites are found in the particular distance with higher probability), and the sharper a selected peak is, the more distinct the intermolecular structure can be considered.

The RDFs for G-series agents (Figure 4, Figure 5 and Figure 6) displayed distinct and well-defined peaks. Sarin (Figure 4) exhibited the most pronounced first peak at approximately 3.4 Å, indicative of strong local ordering and short-range interactions between methyl carbon and the phosphoryl oxygen. Soman (Figure 5) showed a sharper, slightly shifted peak around 4.1 Å, particularly under the ML-derived charge model, suggesting tighter local packing due to the presence of its bulkier alkyl substituent. Tabun (Figure 6) presented a dual-peak structure of moderate intensity, reflecting a more complex spatial organization influenced by the nitrile group, which modifies the electronic distribution around the phosphorus atom.

For A-series agents in the Ellison–Hoenig configuration (Figure 7, Figure 8 and Figure 9), RDFs demonstrated highly ordered first peaks. In particular, A232eh (Figure 7) displayed a sharp and narrow peak with g(r) values exceeding 4.0, suggesting strong directional interactions and tight molecular packing similar to classical G-series agents. The other Ellison–Hoenig structures, A230eh (Figure 7) and A234eh (Figure 9), also showed well-defined short-range order, consistent with their phosphorofluoridate bonding and tetrahedral geometry.

In contrast, the Mirzayanov structures (Figure 10, Figure 11 and Figure 12) exhibited broader, less intense RDF peaks. A230m (Figure 10) and A232m (Figure 10) revealed weaker, more diffuse first coordination shells, beginning around 4.0 Å and peaking at lower g(r) values than their Ellison and Hoenig counterparts. A234m (Figure 12) displayed a slightly sharper first peak, though still less structured compared to the EH configuration. These differences suggest a more relaxed and disordered local environment, likely a result of the direct phosphorus–nitrogen bonding in the Mirzayanov structures, which modifies both electron distribution and steric interactions.

Across all systems, the choice of partial charge model (eQeQ vs. ML) significantly influenced RDF behavior. ML-derived charges generally led to broader and lower-intensity g(r) peaks, indicating a prediction of reduced local density clustering and weaker intermolecular correlations. This effect was especially pronounced in A230m and A232m, where the ML curves produced smoother transitions between coordination shells and lower peak intensities overall.

These RDF results reinforce the conclusion that both structural representation and charge model substantially affect the predicted liquid-phase behavior of nerve agents. These distinctions are not merely computational artifacts—they may have real implications for molecular diffusion, solubility, biological target interaction, and environmental persistence of these highly toxic compounds.

Molecular dynamics simulations of G- and A-type nerve agents demonstrated consistent behavior of density, diffusivity, and intermolecular structure consistent with past computational results. The density of the two agent classes was in close accordance with past MD studies, with G-agents having values close to previously determined [14,22], but A-agents such as A230, A232, and A234 revealed comparable behavior consistent with Mirzayanov’s model of structure [13]. Diffusivity analysis revealed greater molecular motion of G-agents due to more supple conformational moieties, compared to sterically hindered A-series molecules. RDF analysis further revealed characteristic short-range ordering of active molecular sites, consistent with RDF patterns of other MD simulations of polar molecular liquids [13,14]

Moreover, density values for the A-series nerve agents A-230, A-232, and A-234, calculated with the molecular dynamics (MD) simulations, revealed significant variations depending on the structural models employed. Based on Ellison and Hoenig’s phosphorofluoridate structures, the model produced density values closely aligned with the experimental results. This may suggest that these proposed molecular structures can accurately represent the packing efficiency and intermolecular interactions of these agents [1,2]. The simulations implementing Mirzayanov’s phosphoramidofluoridate structures produced consistently lower densities. This discrepancy can be attributed to the presence of a direct phosphorus–nitrogen (P-N) bond in Mirzayanov’s structures, which alters the electronic environment around phosphorus, reducing molecular packing efficiency compared to the phosphorus–oxygen (P=O) bond proposed by Ellison and Hoenig.

3.2. QSAR Modeling Results

Quantitative Structure–Activity Relationship (QSAR) modeling was employed to evaluate key physicochemical and toxicokinetic properties of selected G-series and A-series nerve agents, with an emphasis on parameters relevant to environmental fate and biological behavior. One of the primary physicochemical descriptors analyzed was vapor pressure at 20 °C, which is directly related to the volatility and environmental dispersal potential of these compounds. The G-series agents Sarin and Soman demonstrated relatively high vapor pressures (2.10 mmHg and 0.40 mmHg, respectively), indicative of their known volatility. In contrast, the A-series compounds—particularly A232 and A234—displayed substantially lower vapor pressures (0.01–0.03 mmHg), suggesting a markedly reduced tendency to vaporize under ambient conditions (

Table 8. Physicochemical properties of G-Series and A-Series nerve agents based on QSAR modeling.

				Mirzayanov Structure			Ellison–Hoenig Structure
Agent/Propriety	Sarin	Soman	Tabun	A230	A232	A234	A230	A232	A234
Vapour Pressure (T = 20 °C) mmHg	2.10	0.40	0.04	0.03	0.03	0.01	0.04	0.02	0.01
Boiling Point (°C)	158	198	240	237	240	259	243	253	265

).

These findings suggest that A-series agents may display greater environmental persistence due to their reduced volatility. Boiling point estimates further support this interpretation, with values ranging from 158 °C for Sarin to approximately 260 °C for A234, independently of the proposed molecular structure. Differences between the Mirzayanov and Ellison–Hoenig structural variants of the A-series agents reflect a shift in both vapor pressure and boiling point, with the Ellison–Hoenig structures generally exhibiting slightly higher boiling points and lower volatilities. These results underscore the influence of molecular structure on the thermodynamic behavior of nerve agents and support the utility of MD simulations in differentiating between proposed isomeric forms.

In parallel, QSAR analysis was extended to include lipophilicity (LogP), a pivotal descriptor influencing a compound’s bioavailability, tissue distribution, and potential for bioaccumulation. The modeling results revealed significant divergence in LogP values between the G-series and A-series agents, consistent with their differing absorption profiles and mechanisms of toxicological action [9]. When considered alongside vapor pressure and boiling point data, the LogP findings contribute to a multidimensional understanding of the agents’ environmental mobility and biological interaction potential, reinforcing the value of integrated QSAR approaches in the early-stage hazard characterization of chemical warfare agents.

The calculated LogP values were 0.3 for Sarin, 2.10 for Soman, and 1.70 for Tabun (

Table 9. QSAR-predicted lipophilicity and toxicological metrics.

Agent	LogP	LD50 (mg/kg)	Molecular Weight (g/mol)	PSA (Å2)
Sarin	0.30	0.014	140.09	27.3
Soman	2.10	0.012	182.15	31.6
Tabun	1.70	0.017	162.12	34.8
A-230	2.14	1.23	194.19	32.5
A-232	2.55	1.22	180.17	40.0
A-234	2.97	0.51	224.22	51.7

). Soman’s higher lipophilicity suggests a greater potential for membrane penetration and systemic absorption, consistent with its higher toxicity. LD₅₀ values predicted by the QSAR models showed Soman as the most toxic agent (LD₅₀ = 0.012 mg/kg), followed by Sarin (LD₅₀ = 0.014 mg/kg) and Tabun (LD₅₀ = 0.017 mg/kg) [4] (Table 9).

Additional descriptors such as molecular weight, hydrogen bond donors, and polar surface area (PSA) were analyzed to correlate with the agents’ biological activity. Sarin’s lower PSA suggests its rapid systemic absorption, while Tabun’s slightly higher PSA may reduce its overall bioavailability.

According to the QSAR model’s results, LogP values ranged from 1.70 for Tabun to 2.10 for Soman. These values closely resemble those reported in previous QSAR studies on structurally related compounds, including VX (LogP ~ 2.20) and paraoxon (LogP ~ 1.85) [33]. The correlation between LogP and toxicity observed in this study is also consistent with findings from studies on pesticide-related organophosphates, where increased lipophilicity was linked to higher bioaccumulation potential and AChE inhibition potency [15,17,19].

Furthermore, the predicted LD₅₀ values of Sarin (0.014 mg/kg), Soman (0.012 mg/kg), and Tabun (0.017 mg/kg) were in strong agreement with experimental data from the TOXNET and DSSTox databases, reinforcing the reliability of the QSAR model applied. These results align with previous QSAR analyses of nerve agents, which demonstrated that LD50 values could be accurately predicted using molecular descriptors such as hydrogen bond acceptors, molecular weight, and electronic properties [30].

The calculated LogP values for the A-series agents ranged from 2.14 (A-230) to 2.97 (A-234), indicating higher lipophilicity compared to the G-series compounds. The elevated LogP values are consistent with the more complex and hydrophobic substituents found in the A-series molecular structures, which could enhance membrane permeability and tissue distribution potential. A-234 exhibited the highest lipophilicity, which aligns with its larger molecular weight and surface area, potentially contributing to prolonged biological retention. Recent studies employing cheminformatics and predictive toxicology tools, such as those by Noga and Opravil [8,11], have reported LogP values within a similar range, reinforcing the conclusion that A-series agents exhibit high lipophilicity. This property is associated with increased membrane permeability and the potential for bioaccumulation, particularly for A-234.

The predicted LD₅₀ values for the A-series agents were 0.51 mg/kg for A-234 and around 1.22–1.23 mg/kg for A-230 and A-232, suggesting lower acute toxicity than their G-series counterparts. While these values may seem relatively high for these agents, it is important to note that these predictions are based on extrapolation from a model primarily trained on G-series and related organophosphorus structures.

The results highlight structural features influencing QSAR-predicted toxicity, such as molecular size, lipophilicity, and polar surface area. A-234, which has the highest PSA (51.7 Ų) among the A-agents, also shows the lowest predicted LD₅₀ within its group, potentially reflecting an increased likelihood of biological interaction. These observations offer useful insight into how computational models might differentiate between subtle structural variants, even within a class of closely related compounds.

4. Discussion

While the general trends observed in this study align with previous research, certain distinctions emerged. QSAR analysis of lipophilicity (LogP) of the chemical warfare agents investigated uncovered important information regarding their bioavailability, tissue distribution, and patterns of toxicity. The G-series compounds showed diverse LogP, ranging from 0.30 in Sarin and 1.70 in Tabun to a significantly elevated value of 2.10 in Soman, suggesting an increased tendency of cell membrane permeation and systemic absorption in accordance with their experimental toxicities. In comparison, VX and VR—structurally related organophosphorus agents of the G-series—have about 2.20 and 1.95 LogP, respectively, tracking similarly to the higher lipophilicity range exhibited by G-series compounds. Their resemblance implies analogous potential for bioaccumulation and patterns of toxicity. The A-series compounds showed even higher lipophilicity, ranging from 2.14 (A-230) to 2.97 (A-234), in accordance with their more hydrophobic and bulkier replacing groups and consistent with an affinity for greater cell permeability and prolonged biological retention. Significant to note are the corresponding predicted acute toxicity (LD₅₀) data tracking these lipophilicity trends, such that G-series compounds tended to have the highest acute toxicity (e.g., Soman at 0.012 mg/kg), VX and VR showed intermediate levels of acute toxicity, akin to their physicochemical properties, and A-series compounds showed comparably lesser acute toxicity but greater tendency of bioaccumulation. These results are consistent with literature findings on organophosphorus pesticides and nerve agents relating elevated LogP with greater acetylcholinesterase inhibitory potency and systemic poisoning [17]. While QSAR models have consistently linked LogP values to acute toxicity, our findings suggest that steric factors and electronic properties may also play a significant role. This aligns with recent machine learning-based QSAR studies that incorporate higher-dimensional molecular descriptors to refine toxicity predictions [33].

In addition, the A-series compounds exhibited distinct physicochemical and toxicological profiles, as revealed by both MD simulations and QSAR modeling. The A-series agents demonstrated higher lipophilicity, with LogP values ranging from 2.14 (A-230) to 2.97 (A-234), indicating an increased potential for membrane permeability and bioaccumulation relative to G-series agents. This trend is consistent with the presence of bulkier or more hydrophobic substituents in the A-series molecular structures. Furthermore, density predictions varied significantly depending on the structural model used (Ellison–Hoenig vs. Mirzayanov), with Ellison–Hoenig configurations yielding values more closely aligned with experimental estimations.

Simulations based on Mirzayanov’s structures consistently yielded lower densities, regardless of the charge assignment method. These discrepancies may reflect the impact of the direct P–N bond, which modifies electron distribution and weakens intermolecular interactions in the liquid phase, as also suggested by broader and less intense RDF peaks in the corresponding simulations. In contrast, the Ellison–Hoenig structures, with classical phosphorofluoridate linkages, produced more compact packing, higher densities, and sharper RDF peaks—indicative of stronger and more organized intermolecular interactions.

This divergence between structural models was mirrored in self-diffusivity values: Ellison–Hoenig structures exhibited lower diffusion coefficients, suggesting a tighter liquid-phase organization, whereas the Mirzayanov variants were more dynamic. These findings support the idea that the Ellison–Hoenig structures are more physically consistent with experimental measurements and may, therefore, represent more realistic models of A-agent behavior in environmental conditions.

The QSAR-predicted LD₅₀ values for A-series agents were markedly higher (i.e., lower predicted toxicity), particularly for A-230 and A-232, which may reflect limitations in the model’s applicability domain due to structural dissimilarities with the G-series training set. However, the comparatively lower predicted LD₅₀ for A-234 (0.51 mg/kg) may indicate an enhanced toxic potential in line with its higher lipophilicity and PSA. These findings might suggest the necessity of evaluating each subclass of organophosphorus agents on its own structural and physicochemical terms.

Previous studies using in vitro AChE inhibition assays and crystallographic analysis have demonstrated that organophosphorus nerve agents bind covalently to the serine residue in the active site gorge of acetylcholinesterase, typically involving strong interactions with residues such as Ser203, His447, and Glu334 (in human AChE) [20]. Our MD simulations, particularly the radial distribution functions (RDFs) and density predictions, provide indirect but meaningful insights into how these agents might orient and interact within a biological environment. For example, the more ordered and compact molecular arrangements observed in the Ellison–Hoenig variants—especially A-232 and A-234—are consistent with structural prerequisites for stable binding at the AChE catalytic triad. Moreover, QSAR descriptors such as LogP and polar surface area (PSA), which showed strong correlation with predicted LD₅₀ values, are known to influence both membrane permeability and active site accessibility [9,20]. The elevated LogP and PSA values observed for A-234 suggest an increased potential for bioavailability and prolonged interaction with AChE, which aligns with its lower predicted LD₅₀. These findings collectively support the relevance of our computational models in capturing physiologically meaningful features of nerve agent behavior and are in line with previous experimental observations of structure–toxicity relationships in related compounds [17,18,20].

Further analysis revealed a convergence between MD and QSAR outputs: compounds with higher LogP values, such as A-234, also exhibited lower self-diffusion and higher molar heat capacity, reinforcing their characterization as highly cohesive, membrane-permeable, and potentially bioaccumulative agents. This overlap suggests that hydrophobicity may influence both toxicological and physicochemical behavior, highlighting the relevance of using a dual-modeling approach for predictive purposes.

MD simulations revealed the impact of intermolecular forces on liquid-phase behavior, including hydrogen bonding and van der Waals interactions. At the same time, they successfully predicted key thermodynamic properties, including density, heat capacity, and diffusivity. QSAR modeling emphasized the significance of lipophilicity and molecular descriptors in predicting toxicological outcomes. The correlation between LogP and LD50 highlights the role of physicochemical properties in determining acute toxicity, underscoring the utility of QSAR models in predictive toxicology.

By linking the thermodynamic properties from MD simulations to the toxicological parameters derived from QSAR modeling, this study provides a holistic perspective on G-series and A-series agents. This integrated approach enhances our understanding of their behavior in both environmental and biological contexts. It also highlights how computational tools can offer safe, efficient alternatives to experimental studies for highly toxic substances, provide predictive insights that reduce reliance on in vivo testing, and facilitate multidisciplinary approaches to solving complex problems in toxicology and environmental science.

By integrating experimentally verified external datasets into the QSAR modeling workflow, the development of a robust predictive framework capable of estimating the toxicological properties of G-series and A-series nerve agents with high accuracy can be achieved. The use of well-curated databases such as EPA DSSTox, TOXNET, PubChem, and ChEMBL significantly enhanced the model’s reliability, providing a scientifically sound foundation for further computational toxicology studies.

This comparative analysis highlights the similarities between G-series and A-series agents and other organophosphorus nerve agents, particularly in terms of physicochemical properties, lipophilicity, and toxicity predictions. However, key differences in diffusivity, LogP distribution, and structure–toxicity relationships underscore the importance of tailored computational approaches for each subclass of CWAs. By identifying cross-correlations between MD-derived thermodynamic descriptors and QSAR-predicted toxicological metrics, this work illustrates the added value of integrated simulations over single-method models. It lays the groundwork for refining risk assessments and mitigation strategies for future organophosphate-based threats.

5. Conclusions

This study highlights the value of integrating molecular dynamics (MD) simulations with Quantitative Structure–Activity Relationship (QSAR) modeling to investigate the multidimensional properties of G-series and A-series nerve agents, including Sarin, Soman, Tabun, A-230, A-232, and A-234. Through this combined computational approach, comprehensive insights were obtained regarding both their physicochemical behavior and toxicological profiles.

MD simulations provided high-resolution data on key liquid-phase thermodynamic properties—such as density, heat capacity, and diffusivity—demonstrating how structural variations, particularly between the Ellison–Hoenig and Mirzayanov A-series models, affect intermolecular interactions and environmental persistence. Notably, the Ellison–Hoenig structures exhibited better agreement with available experimental data, supporting their plausibility as accurate representations of A-series agents.

QSAR modeling, focused on lipophilicity (LogP) and acute toxicity (LD₅₀), revealed significant differences among the compounds, particularly emphasizing the higher predicted toxicity of A-234 due to its elevated LogP and polar surface area. These results are consistent with established patterns observed in prior studies on V-series and other organophosphorus AChE inhibitors, reinforcing the critical role of structural features in determining toxicological behavior.

Importantly, the convergence between MD and QSAR findings—such as the observed correlation between reduced diffusivity, increased density, and enhanced lipophilicity—demonstrates the added predictive value of coupling dynamic simulations with structure–activity modeling. This dual methodology enables a more holistic assessment of nerve agent properties, encompassing both environmental fate and biological uptake potential.

By comparing G-series and A-series agents with previously studied nerve agents, this work contributes to the broader computational toxicology literature while extending it with new data on less-characterized chemical warfare agents [16,34].

These findings support the development of targeted mitigation strategies, including bioscavenger design and decontamination planning, and illustrate the utility of computational models as efficient, safe, and scientifically rigorous alternatives to experimental testing.

Ultimately, this study lays a solid foundation for future research in computational toxicology, environmental chemistry, and chemical defense. Continued development of hybrid modeling frameworks, alongside the expansion of curated chemical datasets, will enhance predictive capabilities and improve preparedness against emerging chemical threats.

While this study provides robust insights, certain limitations merit further investigation. Specifically, the force field parameters used may not fully capture long-range interactions, suggesting that hybrid quantum mechanics/molecular mechanics (QM/MM) approaches could offer improved accuracy [23]. Additionally, incorporating further QSAR descriptors—such as biodegradability and chronic toxicity metrics—would provide a more complete toxicological evaluation.

Future research may build upon this work by integrating larger datasets that include long-term exposure and environmental degradation endpoints. Expanding the chemical space to cover additional classes of CWAs and validating the models against more extensive experimental data will further improve the reliability and applicability of computational predictions. Next steps could also include hydrogen atoms modeled specifically in order to account for possible hydrogen bonds formed inside the liquid structure of the particular liquids, taking into account quantum interactions and employing simulations to further study intermolecular structure.