Dual-Site Acetylcholinesterase Inhibition and Multiscale Stability of Fused Quinoline Sulfonamides: A Chemoinformatic GA-MLR and Molecular Dynamics Study

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The primary reason for the manuscript's unsuitability for publication is inadequate validation; the most important problem is the total absence of experimental data. The external validation is based on just eight compounds, which is far too small to show true predictive reliability, even though the QSAR model shows good predictive value. More significantly, all published results, including docking scores, molecular dynamics stability, and ADMET predictions, are based solely on computational methods and have not been confirmed experimentally. There is no proof that the suggested compounds have biological activity in the absence of biochemical tests (e.g., AChE inhibition measurements). Since in silico methods cannot accurately predict real pharmacological properties, this lack of experimental validation is a serious flaw.  A "Lead" compound identified through in silico screening—such as those mentioned in the text—cannot be truly validated unless its inhibitory concentration is tested against actual enzymes in a laboratory setting.
Additionally, no effort was made to validate predictions against independent experimental datasets or previously reported inhibitors beyond superficial comparison.
 As a result, the claims of the lead compounds' superiority and therapeutic potential are not adequately supported -  at least, preliminary experimental validation is expected to substantiate computational predictions. Therefore, the lack of experimental data fundamentally undermines the study's scientific credibility and justifies rejection in its current form.

Author Response

Comment 1:  The primary reason for the manuscript's unsuitability for publication is inadequate validation; the most important problem is the total absence of experimental data. The external validation is based on just eight compounds, which is far too small to show true predictive reliability, even though the QSAR model shows good predictive value. More significantly, all published results, including docking scores, molecular dynamics stability, and ADMET predictions, are based solely on computational methods and have not been confirmed experimentally. There is no proof that the suggested compounds have biological activity in the absence of biochemical tests (e.g., AChE inhibition measurements). Since in silico methods cannot accurately predict real pharmacological properties, this lack of experimental validation is a serious flaw.  A "Lead" compound identified through in silico screening—such as those mentioned in the text—cannot be truly validated unless its inhibitory concentration is tested against actual enzymes in a laboratory setting.
Additionally, no effort was made to validate predictions against independent experimental datasets or previously reported inhibitors beyond superficial comparison.
 As a result, the claims of the lead compounds' superiority and therapeutic potential are not adequately supported -  at least, preliminary experimental validation is expected to substantiate computational predictions. Therefore, the lack of experimental data fundamentally undermines the study's scientific credibility and justifies rejection in its current form.

Response 1: We appreciate the reviewer’s perspective regarding experimental validation. However, we would like to clarify that this manuscript was submitted to the Special Issue 'Recent Research on Biomimetic Chromatography, QSAR and Chemoinformatics,' which invites "pure model submissions" within the Molecular Informatics section. The scope of this issue is to highlight in silico methods as a robust alternative to traditional testing to reduce time and costs in the "Logic of Discovery" pipeline.

While the study is strictly computational, it is not superficial. We have addressed the concerns regarding predictive reliability by stress-testing the model with four rigorous validation layers:

1) Using a Williams Plot (Figure 2G/4), we have defined the exact chemical space boundaries of our model. This proves that Leads 19 and 20 are not statistical outliers or extrapolated guesses, but fall within the high-precision zone of the training set.

2) We conducted a Y-randomization test (Figure 2I). The resulting drop in performance confirms that our R² and Q² values are driven by real structural-activity relationships, not random chance.

3) Our 100-ns MD simulations were analyzed via FEL and DCCM. These metrics move the discussion beyond snapshots to demonstrate that the binding at the 20-Å gorge is a thermodynamically stable event, providing a clear mechanistic rationale for future testing.

4) The model was verified against established AChE inhibitors to ensure it correctly identifies known activity trends before being used for lead prioritization.

By adhering to the reduce and optimize philosophy of this Special Issue, we present these leads not as proven drugs, but as the statistically highest-probability candidates for the next phase of drug development.

Reviewer 2 Report

Comments and Suggestions for Authors

Comments to Authors

1. The reported dataset composition (115 compounds, 81 training, 8 external) is internally inconsistent with the supplementary files, which contain differing numbers of entries across datasets. This lack of alignment prevents clear reconstruction of the modeling workflow and raises serious concerns regarding data integrity.

2. The manuscript reports inconsistent external validation values (R²ext = 0.8620 in the text vs 0.737 in Table 1), which represent a critical discrepancy in the core performance metric. Such inconsistency undermines confidence in the reported predictive ability of the QSAR model.

3. The authors do not provide a rigorous and reproducible description of dataset curation, including inclusion/exclusion criteria, data standardization, or handling of duplicates and experimental variability.

4. The QSAR dataset appears to combine IC₅₀ values from multiple literature sources without demonstrating comparability of assay conditions (enzyme source, substrate, pH, protocol). This introduces uncontrolled experimental variability and may significantly compromise model validity.

5. The external validation set consists of only eight compounds, which is insufficient to robustly assess model generalizability. The very small sample size makes the reported R²ext highly sensitive to individual data points and statistically unreliable.

6. The manuscript assigns detailed mechanistic and binding interpretations to abstract topological descriptors, which are inherently statistical constructs. This level of mechanistic inference is not supported by MLR-based QSAR models and should be substantially tempered.

7. The selection of Leads 19 and 20 is not clearly aligned with the QSAR ranking, raising questions about the internal consistency of the prioritization strategy. A transparent multi-criteria decision framework (QSAR, ADMET, docking, synthetic feasibility) is required but not adequately provided.

8. Docking scores suggest substantially stronger binding affinities than those implied by QSAR-predicted activities for the same compounds. This discrepancy is not addressed and indicates a lack of coherence across the computational pipeline.

9. The manuscript refers to “blind docking,” yet the analysis focuses exclusively on the known active-site gorge of AChE without demonstrating unbiased global sampling. This suggests that docking was likely site-directed, and the terminology should be corrected or justified.

10. The docking protocol does not include validation steps such as redocking of the co-crystallized ligand or comparison of alternative binding sites. Without such validation, the reliability of the docking poses remains uncertain.

11. The MD results, particularly ligand RMSD behavior, indicate structural deviations that are interpreted as favorable without sufficient justification. A more balanced discussion acknowledging possible instability or pose rearrangement is required.

12. The study relies on a single 100 ns trajectory per system without independent replicates or alternative starting conformations. This is insufficient to establish convergence or robust dynamical conclusions.

13. The manuscript references apo simulations as a baseline for dynamic correlation analysis, but corresponding methodological details and results are not fully documented. This weakens the validity of comparative dynamical interpretations.

14. The predicted ADMET properties are presented as strongly favorable despite values that are only moderate or borderline (e.g., solubility and permeability metrics). The discussion should be revised to reflect the limitations of in silico predictions.

15. The manuscript makes claims regarding disease-modifying potential and dual-site inhibition without any experimental validation. Such claims should be significantly moderated or supported by experimental evidence.

16. Biomimetic chromatography (IAM/HSA) is proposed as a future validation strategy but is not included in the present study. Therefore, it should not be framed as part of the current validation framework, as this creates a mismatch between the scope of the work and the extent of the conclusions drawn.

17. Figure 1E does not clearly depict sulfonamide structures despite being described as such in the text. This inconsistency reduces clarity and may mislead readers.

18. Figures presenting model performance do not clearly distinguish between training, internal validation, and external validation datasets. This lack of clarity complicates interpretation of model performance.

19. The study relies on Dragon software for descriptor calculation but does not provide sufficient detail to ensure reproducibility. Given that the software is proprietary, additional transparency is required.

20. The reported RMSE and MAE values correspond to moderate predictive accuracy rather than exceptional performance. The language used in the manuscript exaggerates the model’s predictive strength.

Based on the above-mentioned concerns, the manuscript requires major revision. The study necessitates substantial methodological corrections and full clarification of dataset integrity and validation procedures before it can be considered for publication.

Author Response

Comment 1: The reported dataset composition (115 compounds, 81 training, 8 external) is internally inconsistent with the supplementary files, which contain differing numbers of entries across datasets. This lack of alignment prevents clear reconstruction of the modeling workflow and raises serious concerns regarding data integrity.

Response 1: We appreciate the reviewer pointing out this discrepancy, which resulted from an oversight in the descriptive text rather than a flaw in the data itself. To clarify the modeling workflow: the primary dataset consists of 115 compounds, which were partitioned into a training set (n = 81) and an internal test set (n = 34) for model development and initial validation. The 8 compounds mentioned separately represent a strictly independent external validation set sourced from different literature [30–33], intended as a final structural checkpoint. We have revised Section 3.2 to define the 81/34 split and the independent n = 8 set, ensuring the manuscript now perfectly aligns with the totals provided in the supplementary files.

Comment 2: The manuscript reports inconsistent external validation values (R²ext = 0.8620 in the text vs 0.737 in Table 1), which represent a critical discrepancy in the core performance metric. Such inconsistency undermines confidence in the reported predictive ability of the QSAR model.’

Response 2: We appreciate the opportunity to clarify these metrics. The value of 0.737 in Table 1 represents the predictive coefficient (R²_pred) for the internal test set (n = 34), which was used to evaluate the model's stability during the descriptor selection phase. In contrast, the R²_ext of 0.8620 reported in Section 2.3 and Figure 3 specifically refers to the final performance on the independent external set (n = 8) sourced from separate literature. We have updated Table 1 and the corresponding text to clearly distinguish between R²_pred (internal test) and R²_ext (independent external), ensuring the nomenclature is consistent with standard QSAR protocols.

Comment 3: The authors do not provide a rigorous and reproducible description of dataset curation, including inclusion/exclusion criteria, data standardization, or handling of duplicates and experimental variability.

Response 3: We agree that a transparent curation protocol is essential for QSAR reproducibility. Accordingly, we have expanded Section 3.1 to provide a step-by-step description of our data preparation workflow:

1. Inclusion/Exclusion Criteria: The dataset was restricted to tacrine-quinoline hybrids to ensure structural homogeneity. To minimize experimental "noise," we only included bioactivity data derived from the standard Ellman’s colorimetric assay, excluding studies with non-comparable inhibitory protocols.
2. Duplicate and Outlier Handling: We implemented a rigorous structural deduplication protocol. Where identical structures were reported across multiple sources, we either utilized the mean pIC₅₀ (if values were within 0.3 log units) or excluded the entry if the discrepancy suggested high experimental variability.
3. Chemical Standardization: All structures were converted to canonical SMILES, neutralized, and stripped of salts. We utilized the OCHEM standardization pipeline to ensure consistent tautomer representation before descriptor calculation.

These additions provide the necessary transparency to replicate our data-gathering process and address the concerns regarding data integrity.

Comment 4: The QSAR dataset appears to combine IC₅₀ values from multiple literature sources without demonstrating comparability of assay conditions (enzyme source, substrate, pH, protocol). This introduces uncontrolled experimental variability and may significantly compromise model validity.

Response 4: We acknowledge that experimental heterogeneity is a common challenge in multi-source QSAR studies. To mitigate this, we implemented strict curation filters, now detailed in Section 3.1:

1. We restricted the dataset to studies utilizing the standardized Ellman’s colorimetric assay. This ensures consistency in the substrate (acetylthiocholine iodide), chromogenic reagent (DTNB), and buffering conditions (pH 8.0), effectively minimizing protocol-induced variance.
2. the dataset includes human (hAChE), bovine (bAChE), and electric eel (EeAChE) sources, the 20-Å active site gorge, particularly the catalytic triad and the peripheral anionic site (PAS) is highly conserved across these species. Since our GA-MLR model utilizes 2D constitutional and topological descriptors like π-conjugation length and mass distribution) rather than species-specific 3D atomic coordinates, it captures the universal structural requirements for dual-site binding.
3. The model’s ability to accurately predict the potencies of an independent external set (R²_ext = 0.8620) confirms that the underlying training data retains a high signal-to-noise ratio. If the biological data were critically compromised by experimental noise, such high external predictive power would be mathematically unattainable.

We have updated the manuscript to state these comparability criteria, providing a transparent rationale for our data aggregation.

Comment 5: The external validation set consists of only eight compounds, which is insufficient to robustly assess model generalizability. The very small sample size makes the reported R²ext highly sensitive to individual data points and statistically unreliable.

Response 5: We acknowledge that an n = 8 set, in isolation, would be insufficient for a generalized assessment. However, our validation protocol is a two-tiered strategy where the n = 8 set serves as an independent stress test rather than the sole metric of reliability. As clarified in our response to Comment 1, the primary assessment of generalizability was conducted using an internal predictive set of 34 compounds (~30% of the primary dataset). The additional 8-compound external set was utilized as a secondary challenge using entirely unseen chemical scaffolds from independent literature [30–33]. Despite the small sample size, these strategically selected 8 compounds span an exceptionally wide potency range (from pIC₅₀ 4.8 to 10.7). Achieving a high correlation (R²_ext = 0.8620) across such a vast activity gradient, especially with structurally distinct scaffolds like pyrazolo-pyrano-pyridines, strongly suggests that the model is capturing fundamental structure-activity physics rather than over-fitting. We have updated Section 2.3 to frame the n = 8 set as a high-variance stress test that complements the larger 34-compound predictive validation.

Comment 6: The manuscript assigns detailed mechanistic and binding interpretations to abstract topological descriptors, which are inherently statistical constructs. This level of mechanistic inference is not supported by MLR-based QSAR models and should be substantially tempered.

Response 6: We agree that 2D-QSAR models identify statistical correlations rather than deterministic, physics-based interactions. Assigning residue-specific binding actions like direct interactions with Trp86, solely to topological descriptors is indeed a leap that overextends the statistical reach of MLR. Consequently, we have revised Section 2.2 to appropriately temper these interpretations:

1) We have removed definitive claims linking individual descriptors to specific amino acid residues.

2) Descriptors are now correctly framed as capturing macroscopic physicochemical trends, such as π-conjugation length or longitudinal mass distribution, that statistically correlate with inhibitory potency.

3) We have clarified that the QSAR model serves to generate a structural hypothesis, while the rigorous mechanistic and binding evidence is strictly deferred to the structure-based molecular docking and 100-ns molecular dynamics (MD) simulations presented later in the study.

This realignment ensures the manuscript respects the statistical boundaries of 2D chemoinformatics while maintaining a logical transition to the atomistic insights provided by the MD trajectories.

Comment 7: The selection of Leads 19 and 20 is not clearly aligned with the QSAR ranking, raising questions about the internal consistency of the prioritization strategy. A transparent multi-criteria decision framework (QSAR, ADMET, docking, synthetic feasibility) is required but not adequately provided.

Response 7: The reviewer is correct that Leads 19 and 20 are not the highest-ranked compounds based strictly on predicted pIC₅₀. However, our selection process utilized a Multi-Parameter Optimization (MPO) strategy designed to balance potent inhibition with drug-likeness and synthetic viability, avoiding the common pitfall of prioritizing maximal affinity at the expense of developability. To provide the requested transparency, we have expanded Section 2.4 to define our decision framework:

1) Leads were required to maintain a robust baseline of predicted activity (mid-to-low micromolar) to ensure biological relevance.

2) Compounds 19 and 20 achieved superior synthetic accessibility scores (SA ≈ 2.2) via the sulfonamide linker, ensuring straightforward laboratory validation compared to more complex top-ranked structures.

3) We prioritized leads that maintain optimal lipophilicity (log P) and BBB permeability, deliberately bypassing top-ranked QSAR candidates 14 and 22 that exhibited excessive molecular weight and poor pharmacokinetic profiles.

4) Leads 19 and 20 demonstrated significantly reduced risks for cardiotoxicity and neurotoxicity in our in silico screens compared to current clinical benchmarks.

By detailing this MPO framework in the revised manuscript, we demonstrate that the prioritization of Leads 19 and 20 was a deliberate strategic choice aimed at identifying viable drug candidates rather than merely reporting the highest statistical outliers.

Comment 8: Docking scores suggest substantially stronger binding affinities than those implied by QSAR-predicted activities for the same compounds. This discrepancy is not addressed and indicates a lack of coherence across the computational pipeline.

Response 8: We respectfully submit that this observation stems from a fundamental methodological distinction rather than a lack of pipeline coherence. The QSAR model is trained on empirical IC₅₀ data, which inherently incorporates complex wet-lab variables such as desolvation penalties, substrate competition, and assay-specific kinetics. In contrast, molecular docking scoring functions, designed primarily for pose prediction and relative ranking, evaluate idealized binding enthalpies in a simplified solvent environment. The tendency of docking functions to overestimate absolute macroscopic affinities by neglecting entropic and bulk-solvent contributions is a well-documented phenomenon in molecular informatics. We have updated Section 2.6 to define this thermodynamic scaling difference, reinforcing that our pipeline utilizes docking for structural orientation while relying on the validated QSAR model for activity prediction.

Comment 9: The manuscript refers to “blind docking,” yet the analysis focuses exclusively on the known active-site gorge of AChE without demonstrating unbiased global sampling. This suggests that docking was likely site-directed, and the terminology should be corrected or justified.

Response 9: We agree that blind docking is technically inaccurate here, as our analysis was intentionally focused on the well-characterized 20-Å catalytic gorge and peripheral anionic site (PAS). While the CB-Dock2 algorithm was utilized for automated cavity detection, our subsequent docking and MD simulations were directed toward these established inhibitory sites rather than an unbiased global search for allosteric pockets. Consequently, we have revised the Abstract, Introduction, and Figure 1 to replace blind docking with the more technically appropriate term cavity-directed docking. This adjustment ensures the nomenclature is strictly aligned with the site-specific methodology detailed in Section 3.4.

Comment 10: The docking protocol does not include validation steps such as redocking of the co-crystallized ligand or comparison of alternative binding sites. Without such validation, the reliability of the docking poses remains uncertain.

Response 10: We would like to direct the reviewer to Section 3.5 (previously 3.4), where this validation was documented. The protocol was validated via the self-docking of the co-crystallized ligand (Tacrine), resulting in an RMSD of 0.326 Å between the native and docked poses. This value is significantly below the widely accepted 2.0 Å threshold, confirming the high geometric fidelity and predictive reliability of the docking parameters used throughout this study.

Comment 11: The MD results, particularly ligand RMSD behavior, indicate structural deviations that are interpreted as favorable without sufficient justification. A more balanced discussion acknowledging possible instability or pose rearrangement is required.

Response 11: We agree that the observed 4.2 Å ligand RMSD shift warrants a more nuanced interpretation. Rather than viewing this strictly as a favorable adaptation, we have revised Section 2.7 to acknowledge this deviation as a significant structural rearrangement. This shift indicates that the initial docked pose, constrained by the static nature of the docking grid, underwent a dynamic refinement within the 20-Å gorge to reach a more energetically favorable, alternative minimum. Importantly, the ligand RMSD achieves a clear plateau for the final 60 ns of the simulation, suggesting that while the initial pose was metastable, the subsequent rearranged state is robust and equilibrated. We have updated the discussion to provide a balanced view of this conformational transition and its implications for dual-site binding.

Comment 12: The study relies on a single 100 ns trajectory per system without independent replicates or alternative starting conformations. This is insufficient to establish convergence or robust dynamical conclusions.

Response 12: We agree that multiple independent replicates are the important for exhaustive thermodynamic mapping. However, in the context of this hit-to-lead prioritization study, the 100-ns simulations were not intended to sample the global folding landscape of the enzyme. Instead, they were utilized as a comparative refinement tool to validate the local stability of the dual-site binding hypotheses generated by the QSAR model and to filter out metastable docking poses. To ensure the reliability of these individual trajectories, we rigorously assessed internal convergence using FEL clustering and DCCM analyses. These metrics provided a clear differentiation in the dynamic stability of Lead 19 versus Compound 20, supporting our prioritization logic. We have updated the Conclusion of Section 2.7 to frame these simulations as a comparative validation step and acknowledge that future detailed mechanistic studies will incorporate multi-replicate ensembles for broader sampling.

Comment 13: The manuscript references apo simulations as a baseline for dynamic correlation analysis, but corresponding methodological details and results are not fully documented. This weakens the validity of comparative dynamical interpretations.

Response 13: We acknowledge that while the results of the apo simulation were utilized as a baseline (specifically in the DCCM analysis in Figure 8C and discussed in Section 2.9), the methodological parameters for this trajectory were inadvertently omitted from the Methods section. To ensure complete reproducibility, we have updated Section 3.4 to clarify that the ligand-free apo system was prepared and simulated using the identical 100-ns protocol, force field (AMBER14SB), and thermodynamic ensemble (NPT) as the ligand-bound complexes. This addition provides the necessary methodological transparency to support our comparative dynamical interpretations.

Comment 14: The predicted ADMET properties are presented as strongly favorable despite values that are only moderate or borderline (e.g., solubility and permeability metrics). The discussion should be revised to reflect the limitations of in silico predictions.

Response 14: We agree that characterizing borderline ADMET metrics as strongly favorable was an overstatement. In silico predictions are best utilized as early-stage comparative guides for lead prioritization rather than absolute physiological guarantees. Accordingly, we have revised Section 2.5 to provide a more balanced interpretation of the pharmacokinetic profile. The text now acknowledges that the predicted aqueous solubility (log S) and Caco-2 permeability for Leads 19 and 20 are moderate. We have framed these values not as prohibitive failures, but as known pharmacokinetic hurdles that will require strategic formulation approaches and rigorous monitoring during future in vitro and in vivo validation stages. This adjustment aligns the manuscript with the inherent uncertainties of predictive ADMET modeling.

Comment 15: The manuscript makes claims regarding disease-modifying potential and dual-site inhibition without any experimental validation. Such claims should be significantly moderated or supported by experimental evidence.

Response 15: We agree that establishing a compound as a disease-modifying agent or a confirmed dual-site inhibitor requires rigorous in vitro and in vivo biological validation. Since the current scope of this study is focused on the in silico development of a computational rationale, these claims must be appropriately framed as predictive structural hypotheses. To address this, we have systematically moderated the language throughout the manuscript:

1) We have replaced definitive terms such as mechanistic proof and confirmed dual-site inhibition with computational rationale and simulated dual-site anchoring hypothesis.

2) The concluding remarks in the Results and the General Conclusion (Section 5) have been rewritten to state that the therapeutic potential of Lead 19 represents a structural model that requires future experimental confirmation.

3) We have clarified that our study provides the molecular logic and lead prioritization necessary to guide subsequent experimental assays, rather than claiming final pharmacological proof.

Comment 16: Biomimetic chromatography (IAM/HSA) is proposed as a future validation strategy but is not included in the present study. Therefore, it should not be framed as part of the current validation framework, as this creates a mismatch between the scope of the work and the extent of the conclusions drawn.

Response 16: We agree that the inclusion of biomimetic chromatography (IAM/HSA) in the Abstract and its proximity to the current results created an unintended mismatch between the study’s scope and its conclusions. To rectify this and clarify the boundaries of the present work:

1) We have removed all specific mentions of IAM and HSA chromatography from the Abstract, replacing them with a generalized reference to "future experimental validation."

2) We have rewritten the final paragraph of the Conclusion to explicitly categorize biomimetic chromatography as an external prospective strategy. We now state that these analyses are earmarked for subsequent experimental workflows to validate the current 100-ns MD-derived pharmacokinetic hypotheses. This adjustment ensures that the manuscript remains strictly focused on the computational lead prioritization while preserving a clear roadmap for the next phase of development.

Comment 17: Figure 1E does not clearly depict sulfonamide structures despite being described as such in the text. This inconsistency reduces clarity and may mislead readers.

Response 17: We appreciate the reviewer's attention to the alignment between our introductory text and the visual representation in Figure 1E. We acknowledge that while the text introduces the fused quinoline sulfonamide system, the figure utilizes a generalized box to represent the functional group diversity. To ensure absolute clarity and internal consistency: We have revised the phrasing in the Introduction to clarify that Figure 1E depicts the general pharmacophore template.

Comment 18: Figures presenting model performance do not clearly distinguish between training, internal validation, and external validation datasets. This lack of clarity complicates interpretation of model performance.

Response 18: We have updated Figure 2 to include an on-plot legend explicitly distinguishing the Training Set and External Test Set using distinct marker symbols and colors. Furthermore, the caption for Figure 2 has been revised to include the specific n-values for each subset.

Comment 19: The study relies on Dragon software for descriptor calculation but does not provide sufficient detail to ensure reproducibility. Given that the software is proprietary, additional transparency is required.

Response 19: We would like to clarify that the descriptors were not calculated via a standalone local license but through the Online Chemical Database (OCHEM) platform, which provides a standardized, version-controlled environment for Dragon7 and alvaDesc engines. To ensure full reproducibility and address the reviewer’s concern:

We have revised Section 3.1 to explicitly state the use of the OCHEM web-service and specify the descriptor versions used. We have confirmed that the complete set of calculated descriptors for all molecules in the dataset is included in the Supplementary Materials (Table S2). This allows any researcher to replicate our QSAR modeling and statistical validation regardless of their access to the Dragon software.

Comment 20: The reported RMSE and MAE values correspond to moderate predictive accuracy rather than exceptional performance. The language used in the manuscript exaggerates the model’s predictive strength.

Response 20: We agree that the terminology used to describe the model's predictive strength should strictly align with the calculated statistical parameters. While the RMSE and MAE values demonstrate high internal consistency and external reach, characterizing the performance as "exceptional" may obscure the inherent margins of error in 2D QSAR modeling. To ensure a more balanced and scientifically grounded presentation: We have systematically audited Section 2.3 and the Conclusions to replace superlative descriptors with objective terms such as "robust," "statistically significant," and "reliable for lead prioritization." We also have added a brief statement acknowledging that while the model provides a high-confidence filtering mechanism for the 20-Å AChE gorge, it serves as a predictive guide for hit-to-lead optimization rather than an absolute pharmacological replacement. This shift in tone ensures that our claims are fully supported by the reported statistical metrics.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The changes in version 2 do not sufficiently address the main criticism. Rather than significant gains in validation, the modifications are primarily rhetorical, cosmetic, or reframing. The lack of experimental validation, the primary concern, remains unresolved.

Reviewer 2 Report

Comments and Suggestions for Authors

Comments to the Authors

The authors have addressed the majority of the reviewer’s concerns and improved the overall clarity and rigor of the manuscript. The revisions have enhanced the consistency of the methodology and interpretation. While some limitations remain, they are acknowledged and do not preclude publication. The manuscript can be considered for acceptance.