Chemical neighborhood exploration for substrate discovery in biocatalysis

Significance

Computational methods in biocatalysis are mostly focused on biocatalytic matrix engineering, whereas identifying the most reactive substrates for a given catalyst remains much less developed. By quantitatively reproducing the experimentally observed activation barrier, our ab initio metadynamics simulations approach establishes a theoretical framework for predicting catalytically active substrates. This capability facilitates rational discovery of enzymatic reactivity across broad chemical libraries. Further, the direct stochastic computational analysis of substrate libraries is computationally demanding. Developed Subdate workflow introduces a prescreening strategy that enables efficient exploration of compound libraries by focusing resources on the most promising candidates.

Abstract

Predicting the substrate reactivity strength for a given biocatalyst remains a central challenge in computational biocatalysis. Here, we present Subdate, a modular workflow that combines descriptor-guided organization of substrate analogs with ab initio metadynamics simulations to prioritize reactive candidates. The workflow integrates i) substrate library construction, ii) conformer generation and descriptors set definition, iii) library clustering, iv) representative-substrate selection, and v) reaction-barrier quantitative prediction. Applied to selected biocatalysts (human butyrylcholinesterase and the catalytic antibody A17) sharing an SN2 reaction mechanism, Subdate quantitatively identifies reactivity trends that match experimental kinetic measurements. The developed workflow provides a mechanism-aware strategy for reactive substrate prioritization for efficient sampling through the chemical library in biocatalysis.

Identifying highly reactive substrates for a given biocatalyst remains a central challenge in computational biocatalysis, particularly when evaluating families of closely related analogs (1–3). While major advances have been made in enzyme engineering, computational prioritization of substrate analogs remains less developed, especially when reaction barriers rather than binding alone determine performance. Combinatorial approaches have greatly expanded the field through the use of chemical, DNA-encoded (4), and display libraries (5), enabling directed evolution of both substrates and catalytic protein templates (6). More recently, high-throughput screening technologies based on microfluidic droplet platforms have further accelerated the selection of biocatalysts with improved properties (7–9). A biocatalytic reaction can proceed efficiently when both partners, i.e., the catalytic template and the corresponding substrate, are suitable for each other. This depends on the thermodynamic parameters of reversible interactions, chemical steps, entropic factors, and dynamic factors (10). Accordingly, efforts to improve catalytic performance have focused largely on modification for the catalyst itself based on results of quantum mechanics/molecular mechanics (QM/MM) predictions (11, 12) or random mutagenesis (13, 14) and screening technologies (15). In addition, the discovery of catalytic antibodies and ribozymes broadened the landscape of possible biocatalytic templates (16–20). Yet despite major progress in expanding and optimizing catalytic scaffolds, methods for efficient prioritization of reactive substrate analogs remain limited.

Of particular interest are molecular systems involved in the catalytic metabolism of organophosphorus (OP) compounds, which are strongly linked to the toxicology of pesticides (21, 22) and other OP chemicals (23). As irreversible inhibitors of cholinesterases, these compounds remain an important public health concern (24). Over the past 70 y, mechanistic studies of OP inhibition and transformation by wild-type (WT) and engineered acetylcholinesterases (AChE) and butyrylcholinesterases (BChE) have established these enzymes as promising bioscavengers and antidotal agents (25, 26). In parallel, de novo engineered catalytic antibodies have shown efficient (11, 27) and stereoselective (28) covalent capture of OP substrates. These advances in catalytic template design highlight the growing need for computational strategies that can prioritize reactive substrate analogs for detailed mechanistic evaluation. Continued progress in computational chemistry and high-performance computing is expected to further support the development of biocatalysts for OP poisoning prophylaxis and post-exposure treatments (29). More broadly, the rational design of covalent and noncovalent modifiers continues to demonstrate therapeutic value in modern drug discovery, including selective inhibitors targeting the JAK2 pseudokinase domain (30, 31). Together, these advances provide a well-defined mechanistic setting for comparing the reactivity of related OP analogs across distinct biocatalytic templates and thus a suitable testbed for developing computational approaches to substrate prioritization.

Recent computational approaches for substrate design increasingly rely on rational, structure-guided strategies (32–34). For mechanistic evaluation, methods such as covalent docking, free-energy calculations, and enhanced-sampling simulations provide progressively richer descriptions but differ substantially in cost and resolution. Covalent docking is efficient for generating reactive binding poses yet offers limited treatment of solvation and barrier formation (35, 36). Free-energy methods achieve high accuracy for congeneric series but demand extensive sampling (37). Enhanced-sampling techniques access rare events but depend on collective variable choice (38, 39). Notably, novel neural network models can predict key kinetic parameters in seconds, though their predictive power remains strongly tied to the quality and relevance of the training data (40, 41).

In turn, ab initio metadynamics offers an attractive alternative by combining quantum mechanical (QM) potential energy surfaces with enhanced sampling, autonomously discovering reactive pathways and yielding activation barriers (ΔG‡) without empirical force fields, albeit at higher computational cost. Recent advances leverage tight-binding methods (e.g., DFTB, GFN-xTB) and theozyme-type strategies (42) to extend QM regions cost-effectively.

To prioritize substrates selection efficiently, we turn to physically grounded descriptors. Electron density distribution around rearranging bonds provides a mechanistically meaningful representation of reactivity. Descriptors derived from the Quantum Theory of Atoms in Molecules (QTAIM) (43)—such as kinetic and potential energy densities at bond critical points—capture key electronic differences across related substrates with low sensitivity to the choice of density functional theory (DFT) functional (44), offering a more reliable framework than empirical descriptors like Hammett constants.

Steric and polarity features are effectively captured by the COSMO solvent accessible surface representation (45), which yields physically meaningful descriptors of molecular shape and interaction potentials at low computational cost.

In this work, we combine QTAIM-derived and COSMO-based descriptors to enable targeted characterization of a library of closely related congeneric substrates, referred to as a chemical neighborhood, with the aim of efficiently identifying the most promising reactive candidates. Here, the notion of a chemical neighborhood is used in a local and task-specific sense: Rather than aiming at broad exploration of chemical space, we consider a set of descriptors computed for a predefined substrate family that share a common reactive scaffold but differ in substituents or electronic features influencing catalytic reactivity. In contrast to conventional high-dimensional representations of chemical neighborhood (46), the QTAIM-COSMO descriptor set is compact, physically grounded, and directly linked to electron density topology and solvation characteristics. As a result, the resulting projection admits a clear physicochemical interpretation and is particularly well-suited for resolving local variations within closely related substrates, enabling efficient and mechanistically meaningful prioritization of substrates for detailed analysis.

Here, we introduce Subdate (SUbstrate Bader Descriptors for biocATalytic reactivity Estimation), a computational workflow for iterative electron density-guided analysis and prioritization of biocatalytic substrates. The method combines rule-based generation of core substrate’s derivatives, DFT-based conformational, and descriptor analysis using QTAIM and COSMO surface parameters, hierarchical grouping, and targeted estimation of activation barriers in selected biocatalyst–substrate complexes by ab initio metadynamics. This multilevel design reduces the need for brute-force simulation of the entire chemical library by prioritizing the most informative candidates for higher-level evaluation. We applied the workflow to well-studied esterase systems of pharmacological and toxicological relevance, including recombinant human butyrylcholinesterase (BChE; catalytic residue Ser198) (26), its mutant BChE S198C (9), and a catalytic antibody with Tyr-L37 as the catalytic residue (27). Across more than 40 organophosphate substrates, we examined how related analogs behave toward distinct catalytic templates sharing the same reaction type. Experimental validation on 11 substrates spanning low to high predicted reactivity showed that the computational predictions captured major reactivity trends and identified previously unrecognized highly reactive substrates for BChE. Together, these results establish Subdate as an open-source, mechanism-aware framework for prioritizing reactive substrates within predefined analog series while reducing unnecessary computational cost.

Results

The Subdate Workflow.

Building upon predefined substrate generation rules and an initial seed molecule, our workflow systematically characterizes a chemical neighborhood to identify regions of optimal reactivity. This is achieved through iterative analysis combining ab initio metadynamic simulations and mapping between substrates descriptors. The workflow reduces the number of computationally expensive ΔG‡ evaluations by prioritizing structurally and chemically diverse compounds for sampling. The core computational pipeline (Fig. 1) comprises five modules:

• i.“Generator” constructs a congeneric substrate library based on predefined rules (conserved core structure and substituent composition);
• ii.“Descriptor” performs QM geometry optimization and extracts QTAIM electron density parameters together with COSMO steric/polarity descriptors for bonds/atoms related to reactivity;
• iii.“Clustering” applies hierarchical agglomerative clustering of the neighborhood;
• iv.“Walker” iteratively selects representative substrates for the Predictor module using the clustering and Predictor’s results;
• v.“Predictor” constructs pretransition state (preTS) complexes and calculates ΔG‡ via ab initio metadynamics (metaQD).

Fig. 1.

Subdate protocol pipeline scheme. Subdate protocol pipeline starts with the substrate library generation (Generator) followed by the DFT-based conformer optimization and QTAIM and COSMO characterization (Descriptor). Next module performs the substrate library hierarchical clusterization (Clustering). Then, the iterative loop of Walker and Predictor modules search for the most successful chemical compounds.

The comparison of ΔG‡ values for selected compounds is used by the Walker to rank promising vs. nonpromising branches and select new candidates for subsequent evaluation, minimizing computational resources spent on unpromising areas.

The workflow begins with defining substrate generation rules (modifiable sites and permissible substituents) followed by library generation. The Descriptor module first performs conformer screening using Conformer-RotamEr Sampling Tool (CREST) (47–49) and DFT optimization (Materials and Methods); then wavefunctions undergo QTAIM and COSMO analysis. QTAIM provides electron density descriptors for target atoms/bonds, and COSMO surfaces describe volume and integral charge of chosen groups.

Each compound was described by a 17-vector comprising 10 COSMO parameters (integral charge and VdW volume for five variable substituents R1–5) and seven QTAIM parameters: charge of phosphorus atom of the leaving group q(P), electron localization function ELF(C─O), bond ellipticity ϵ(P─O) and ϵ(C─O), Laplacian of electron density L(C─O) energy density H/ρ(P─O) and electron density H/ρ(C─O). These parameters have been shown to be adequate descriptors of reactivity and steric properties (50, 51). The Clustering module uses the combined QTAIM/COSMO matrix to perform agglomerative hierarchical clustering, producing a dendrogram that characterizes relative similarity among library compounds (52, 53). On the next step the Walker module exploits the dendrogram via a branch-and-bound search. For each cluster (branch), the medoid (the leaf minimizing sum of pairwise distances to all other leaves in that cluster) is selected as representative. Finally, the Predictor module calculates ΔG‡ for each chosen medoid (unless already obtained). Clusters are compared by medoid ΔG‡, and a successive halving strategy is applied: Among nonterminal clusters, branches are sorted by ascending medoid ΔG‡ and only the best branch is retained for further exploration. The retained branch is expanded into its two child nodes, and the process repeats. Terminal branches are not expanded; their medoid ΔG‡ contribute to the final candidate set. Predictor performs metaQD simulations with predefined collective variables to determine free energy profiles and reaction ΔG‡ (Fig. 2).

Fig. 2.

Schematic illustration of the clustering and Walker modules. The compounds were grouped using a combination of QTAIM + COSMO descriptors (a correlation threshold of 0.9), with hierarchical agglomerative clustering and Ward’s linkage in Euclidean space. The resulting dendrogram (A) displays compounds color-coded by ΔG‡ values only for evaluated medoids computed with Predictor module. The search procedure (B) employs a branch-and-bound with successive halving algorithm that starts from the root node, selects medoid compounds representing major clusters (C1 and C2), compares their ΔG‡, and follows the branch with the lower value, while the alternative branch is pruned (red cross). The process iteratively refines the search by comparing medoids of smaller subclusters (C11 and C12 in the second round, C121 and C122 in the third, etc.) until reaching the leaf level. The selected compound is indicated with a green checkmark. In the example shown, the search proceeded through four rounds, evaluating five compounds in total.

Substrate Chemical Neighborhood Exploration.

We applied the workflow to three biocatalysts with esterase activity: WT human butyrylcholinesterase (BChE), its mutant BChE S198C, and catalytic antibody A17. These were selected to validate applicability across diverse active site nucleophiles (serine, cysteine, tyrosine).

Library generation.

Starting from a phenolate organophosphate (Compound 14), we generated a library of 45 derivatives (SI Appendix, Table S1). Substituents included electron-donating (–OMe, –SMe, –Me, –tBu, –NMe₂, –SO₂Me, –Cl, –COSMe) and electron-withdrawing groups (CN, –NO₂, –SMe₂, –CO₂, –NMe₃, COOMe) placed at ortho, meta, or para positions, as well as disubstituted derivatives. Optimized geometries and frequency analyses are provided in the SI Appendix ZIP archive. The library showed a balanced distribution of steric and electronic properties, though compounds with two –SMe₂ or –NMe₃ groups (30, 31, 40, 41) exhibited altered QTAIM properties, and in compound 30 (ortho-, para-diSMe₂) the P–O bond became unstable.

Clustering strategy.

We compared principal component analysis (PCA) and hierarchical clustering for the descriptor-based grouping. Both gave similar results, but hierarchical clustering required fewer metaQD runs (SI Appendix, Table S6), so it was adopted for the main workflow. Systematic evaluation of 210 clustering schemes (three correlation thresholds, seven descriptor sets, 10 linkage/metric combinations) revealed: i) the framework identified the most active compounds in fewer than nine calculations; ii) no single clustering configuration was universally optimal, reflecting structural heterogeneity; iii) higher correlation thresholds (0.9, 0.95) improved predictive consistency; iv) the QTAIM + COSMO combination outperformed the expanded set including RDKIT, underscoring the value of high-quality descriptors and descriptor parsimony.

MetaQD simulations.

For each selected substrate, we prepared the enzyme–substrate system for metaQD as follows. Crystal structures were used: PDB 2XZC for A17 (27), PDB 7AMZ for BChE (54); BChE S198C was modeled by single-point mutation. Catalytic sites were isolated (residues within 6 Å of the nucleophile), termini capped, and a structurally conserved water molecule (from 7AMZ) was retained; a similar water was added to A17. Substrates were placed in preTS geometry with P–nucleophile distance 2.5 Å. For WT BChE and S198C, the P=O oxygen was oriented into the oxyanion hole; for A17, the catalytic tyrosine was modeled in its deprotonated (tyrosinate) form, as prior studies have shown this is essential for catalysis (28). MetaQD simulations were run to obtain multiple forward/reverse trajectories. Representative free energy surfaces are shown in Fig. 3. For the highly reactive 4-SMe₂-phenolate (Compound 12), we observed a low barrier of 9 kcal/mol with BChE, while the less reactive 4-OMe-phenolate (Compound 5) gave 13.5 kcal/mol. Barriers for the other two biocatalysts were substantially higher (>24 kcal/mol for BChE S198C and >30 kcal/mol for A17). For BChE S198C, for many substrates transition state formation occurred only in a single short episode, consistent with the weak nucleophilicity of cysteine. For A17, a semistable trigonal bipyramidal intermediate was observed—a feature not seen in BChE variants—likely due to stabilization of negative charge and redistribution to the leaving group oxygen or P=O bond. In reactive compounds this intermediate readily formed products; in less reactive cases, reversion to reactants via reprotonation of tyrosinate by water was common.

Fig. 3.

Comparative analysis of ab initio metadynamics for highly reactive (A) and low reactive (B) compounds: free energy surface heat-maps and CVs perturbations during trajectories. Red crosses indicate the presence of TS geometry.

For BChE, the most active substrates (ΔG‡ < 20 kcal/mol) contained –CN, –NO₂, –SMe₂, or –NMe₃ groups. Although these shared similar QTAIM profiles, barriers varied by up to 9 kcal/mol depending on substituent position: 2,5substituted derivatives consistently showed higher barriers than 2,4 analogs, indicating that steric constraints in the active site outweighed electronic effects. Compounds with –OMe, –SMe, –Me, –tBu, –NMe₂, –SO₂Me, –Cl, or –COSMe were predominantly moderate to low activity (ΔG‡ > 23 kcal/mol), as expected for an SN2 mechanism requiring leaving group stabilization. Chloro derivatives were exceptions, showing intermediate reactivity (22 to 25 kcal/mol) due to weak σ withdrawing capacity partially mitigating inherent electron-donating effects. Bulky substituents (–tBu, –SO₂Me, –COSMe) induced severe steric penalties, raising barriers by 2 to 5 kcal/mol relative to compounds with similar QTAIM properties. Full data are summarized in SI Appendix, Table S2.

Experimental Validation.

To validate the workflow, we performed in vitro experiments for all three catalysts. Ten compounds spanning the predicted activity range (9 to 16 kcal/mol) were selected alongside the initial substrate (Compound 14) (Fig. 4B).

Fig. 4.

Experimental validation of computationally generated compounds. (A) Representative progressive reaction curves of BChE inhibition by Compound 13 in the presence of chromogenic substrate. (B) Location of the experimentally tested compounds within the BChE-specific dendrogram for the chemical neighborhood. (C) Correlation plot between computed (ΔG‡ computational) and experimentally determined (ΔG‡ experimental) Gibbs free energy values for selected compounds reacting with native BChE.

Inhibition activity against WT BChE was measured using a progressive kinetics assay with chromogenic substrate (1 mM DTNB, 0.5 mM BTC) (Fig. 4A and SI Appendix, Fig. S2). For all tested compounds, the experimentally determined second-order rate constant k₂ was ≥ 6.5 × 10⁻⁴ M⁻¹·s⁻¹ (Compound 5). Most compounds showed k₂ one to two orders higher. Remarkably, 4-SMe₂-phenolate (Compound 12) gave k₂ = 7.2 × 10⁻¹ M⁻¹·s⁻¹, making it the most reactive compound in the dataset- even surpassing all NO₂-containing derivatives (e.g., paraoxon, Compound 16, k₂ = 7.7 × 10⁻³ M⁻¹·s⁻¹). The zwitterionic 2-NMe₃-phenolate (Compound 11) also showed promising activity (k₂ = 4.8 × 10⁻² M⁻¹·s⁻¹). In contrast, 4-CO₂-phenolate (Compound 13) and 4-SO₂Me-phenolate (Compound 8) were much less active (k₂ = 7.6 × 10⁻⁴ and 1.5 × 10⁻² M⁻¹·s⁻¹, respectively). Unexpectedly, 2,4,6-triCl-phenolate (Compound 23) showed no detectable inhibition despite a computed barrier (15.3 kcal/mol) similar to that of the initial substrate (15.5 kcal/mol). This discrepancy may be due to energetically unfavorable induced rearrangements of side chains Y332, F333, and F290, as suggested by minor steric clashes observed in modeling.

Overall, experimental results agreed well with computed barriers for BChE (R² ≈ 0.76; Fig. 4C and SI Appendix, Table S3). For the less efficient biocatalysts (A17 and BChE S198C), the method reliably characterized all tested substrates as inactive, with high computed barriers (SI Appendix, Tables S4 and S5). These results demonstrate that QTAIM/COSMO-based chemical neighborhood analysis effectively identifies and categorizes compounds with high chemical reactivity.

Predictor Module Evaluation.

We compared six computational approaches against experimental ΔG‡ for 11 substrates with native BChE (Table 1). Sequence-based machine learning models [CataPro (55), MPEK (56), RealKcat (57)] produced ΔG‡ values narrowly clustered between 14.8 to 17.5 kcal/mol, yielding near-zero correlation with experiment (R² ≤ 0.07), indicating they predict an average barrier with no substrate-specific discrimination.

Table 1.

Performance comparison of computational methods in predicting 11 substrate reactivity against native BChE

Compound	CataPro, kcal/mol	MPEK, kcal/mol	RealKcat, kcal/mol	OrbV3, ddG adsorption, kcal/mol	OrbV3, ddG reaction, kcal/mol	CovDock, kcal/mol	metaQD, kcal/mol	Experiment, kcal/mol
Compound 1	16.1	15.4	15.0–16.2	−61.1	−73.1	−59.5	11.3	14.7 ± 1.1
Compound 5	16.0	15.1	16.2–17.5	−71.0	−74.7	−55.5	13.5	17 ± 1.8
Compound 8	16.2	15.4	15.0–16.2	−70.6	−85.8	−61.1	13	16.1 ± 0.9
Compound 11	15.5	15.4	15.0–16.2	−146.2	−158.4	−55.6	9.8	14.4 ± 1.3
Compound 12	16.0	14.8	15.0–16.2	−158.0	−168.1	−65	9	12.8 ± 1.5
Compound 13	16,0	15.5	15.0–16.2	−48.7	−88.1	−57.2	13.3	16.9 ± 0.8
Compound 14	16.1	15.5	15.0–16.2	−64.3	60.6	−55.1	15.5	16.5 ± 1.4
Compound 16	15.6	15.4	16.2–17.5	−68.3	−85.3	−60.3	12.8	15.5 ± 1.0
Compound 17	16.0	15.8	15.0–16.2	−77.0	−84.9	−58	14.4	16.7 ± 1.9
Compound 19	16.2	16.2	16.2–17.5	−78.6	−98.9	−60.5	10.8	14.7 ± 0.8
Compound 23	16.6	15.8	16.2–17.5	−63.4	−74.5	−56.4	15.3	X
Correlation coef. R2	0.03	0.07	0.01	0.60	0.38	0.45	0.82

OrbV3 (58) adsorption energies ranged from −48.7 to −158.0 kcal/mol—orders of magnitude larger than typical noncovalent binding affinities. The most negative values corresponded to cationic substrates (Compound 11, 12), where positive charge in low dielectric implicit solvent generates strong Coulombic stabilization. Despite the absolute scale, the correlation with experimental barriers (R² = 0.60) suggests that adsorption energy captures some reactivity-relevant information, likely reflecting how tightly the substrate is positioned for catalysis. OrbV3’s fast inference (1.5 h per structure) enables preliminary screening of large libraries to flag extreme adsorption signatures for subsequent metaQD validation.

CovDock (59) scores ranged narrowly from −55 to −65 kcal/mol, with correlation R² = 0.45. More negative scores tended to correspond to more reactive substrates (e.g., Compound 12 scored −65.0 kcal/mol), but the relationship was not monotonic and could not resolve fine reactivity ordering. CovDock thus provides a coarse binary filter (active vs. inactive) rather than quantitative prediction.

metaQD achieved strong quantitative agreement (R² = 0.82), with computed barriers (9.0 to 15.5 kcal/mol) closely tracking experimental values (12.8 to 17.0 kcal/mol). The method correctly identified the most reactive (Compound 12) and least reactive (Compound 5) substrates, capturing the 4.2 kcal/mol spread distinguishing slow from fast substrates. Most predictions fell within 2 to 3 kcal/mol of experiment—accuracy sufficient for ranking and classification. For Compound 23, which showed no inhibition experimentally, metaQD predicted ΔG‡ = 15.3 kcal/mol, suggesting inactivity arises from factors preceding transition state formation rather than an inability to undergo phosphoryl transfer.

Discussion

The development of novel biocatalysts and their small molecule effectors has advanced drug discovery and biotechnological applications (54). Two complementary strategies have emerged: design of catalytic scaffolds and generation of tailored substrates. To make biocatalytic acts efficient, both the protein template and the substrate must be optimized—a “two-dimensional screening” concept we previously conceptualized (60).

Despite progress in computational enzyme design (12, 61–64), substrate mining remains comparatively underdeveloped (65–68). We developed Subdate to facilitate rational exploration and prioritization of congeneric substrate libraries (chemical neighborhoods) by combining computationally inexpensive hierarchical clustering of compounds’ descriptors with targeted ab initio metadynamics (metaQD). This iterative approach reduces the need for brute-force simulation across the entire chemical library.

We validated Subdate on well-characterized OP-metabolizing biocatalysts with available structural, kinetic, and mechanistic data (9, 69): WT butyrylcholinesterase (BChE), its mutant BChE S198C (70), and catalytic antibody A17 (27, 71). Application to this system identified highly reactive substrates for BChE, demonstrating a strategy that could be extended to pharmacological target optimization (72). Looking forward, Subdate is built as a modular pipeline, allowing easy replacement of components. For instance, a different metadynamics protocol or custom barrier estimation could be substituted. Clustering and diversity sampling approaches could be altered as well (73). Integrating machine learning into modules (iii–v) is a promising avenue: Accumulated metaQD data can train simple predictors to guide subsequent substrate selection. We previously observed that a boosting model trained on QTAIM vectors and metaQD barriers showed predictive power (42). With continued accumulation of high-quality structural and biophysical data, direct prediction of catalytic constants—a longstanding goal—may become feasible.

Alternative computational approaches showed clear limitations in this context. Covalent docking (CovDock) failed to reliably discriminate between substrates (R² = 0.45), yielding close scores for most compounds due to inadequate treatment of electron density variations. Blackbox neural network models (CataPro, MPEK, RealKcat) gave near-zero correlation (R² ≤ 0.07), confirming that sequence-based predictions remain inferior to structure-based approaches. Adsorption energies from MLIPs (OrbV3) showed moderate correlation (R² = 0.60) but are orders of magnitude larger than true activation barriers, making them unsuitable as direct proxies. However, these fast methods can serve as preliminary filters: OrbV3 and CovDock can rapidly screen large libraries to flag promising candidates before committing to metaQD, which alone achieves quantitative accuracy (R² = 0.82) by explicitly modeling bond-breaking events along the reaction coordinate.

Several limitations of the present approach should be noted. First, the accuracy of semiempirical methods used in metaQD typically reaches ~5 kcal/mol for organic substrates, but uncertainty increases for metalloenzymes due to inadequate metal-centered parameterization (74). This may limit application to metalloenzymes such as phosphotriesterases. Second, metaQD simulations are computationally intensive, especially for active sites with several hundred atoms. Third, multistep reactions require separate simulations for each elementary step, with distinct starting geometries and collective variables. Fourth, structurally different substrates can yield similar QTAIM descriptors but different enzymatic behavior (e.g., –diSMe₂ and –diNMe₃ derivatives show similar QTAIM but divergent barriers). These limitations are discussed further in the SI Appendix.

Despite these limitations, we anticipate that combining costly but precise methods like metaQD with physically grounded chemical neighborhood descriptors (QTAIM and COSMO) can accelerate discovery of highly reactive substrates.

Materials and Methods

Substrate Library Generation.

A congeneric library of 45 organophosphate substrates was generated by systematically varying substituents on a phenolate core (Compound 14). Substituents included electron-donating and electron-withdrawing groups at ortho, meta, and para positions, as well as disubstituted derivatives. Stable minima were confirmed by vibrational frequency analysis (no imaginary frequencies). Details are provided in SI Appendix, Materials and Methods.

Descriptor Calculation.

For each substrate, we performed conformer screening using CREST (47–49) with GFN2xTB, followed by DFT geometry optimization (ORCA 5.0.3, TPSS/ZORA-def2TZVP). QTAIM analysis (Multiwfn) provided electron density parameters for the P─O bond and adjacent atoms. COSMO surface analysis (MOPAC2016, PM7) yielded van der Waals volumes and integrated charges for substituents. After cross-correlation filtering (threshold < 0.9), a compact descriptor set of 7 QTAIM and 10 COSMO parameters was retained. Full descriptor definitions and computational settings are in SI Appendix, Materials and Methods.

Clustering and Iterative Selection.

Hierarchical agglomerative clustering (Ward linkage, Euclidean metric) was applied to the combined QTAIM/COSMO descriptor matrix, generating a dendrogram. A branch-and-bound search with successive halving was used to select representative medoids for metaQD evaluation. Details of the clustering and search procedure are provided in SI Appendix, Materials and Methods.

Ab Initio Metadynamics (metaQD).

Pretransition state complexes for WT BChE, BChE S198C, and catalytic antibody A17 were constructed from crystal structures (PDB: 7AMZ and 2XZC). For each complex, three independent 500 ps metadynamics trajectories were run at 298 K using DFTB+ with GBSA implicit solvation and PLUMED metadynamics module. Collective variables were defined as the nucleophile–P distance and P–leaving group distance. Activation free energies (ΔG‡) were extracted from the resulting free energy surfaces. Full simulation parameters, restraint schemes, and theozyme construction are described in SI Appendix, Materials and Methods.

Benchmarking.

For the 11 experimentally validated substrates, we compared metaQD predictions against: covalent docking (CovDock, Schrödinger Suite, OPLS3e force field), sequence-based ML models (CataPro, MPEK, RealKcat), and OrbV3 machine learning interatomic potential (adsorption and reaction energies). All benchmarking protocols are detailed in SI Appendix, Materials and Methods.

Experimental Validation.

Inhibition kinetics against WT BChE were measured using a chromogenic substrate (1 mM DTNB, 0.5 mM BTC) with progressive reaction assays. Second-order rate constants (k₂) were determined for 11 compounds and converted to ΔG‡ values. Detailed experimental procedures are provided in SI Appendix, Materials and Methods.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Data and code are available at GitHub repository Subdate-QTAIM/home (75). Other data are included in the manuscript and/or SI Appendix.