Jiale Li¹, Jianbo Tong^1,*, Yakun Zhang¹, Yue Sun¹

1, College of Chemistry and Chemical Engineering, Shaanxi University of Science and Technology, Xi’an 710021, China.

jianbotong@aliyun.com

First author: Jiale Li, 1194993517@qq.com

Second author + corresponding author: Jianbo Tong, jianbotong@aliyun.com

Third author: Yakun Zhang, 220811019@sust.edu.cn

Fourth Author: Yue Sun, 1437815972@qq.com

Acknowledgement

Funding: This work was supported by the National Natural Science Foundation of China (22373062)

Abstract: Accurate prediction of binding affinity for ACE2 inhibitors remains a critical challenge in computational drug discovery, particularly given the enzyme’s therapeutic importance in cardiovascular diseases and its role as the cellular entry receptor for SARS-CoV-2. This study aimed to develop and validate a comprehensive computational framework integrating molecular docking with free energy perturbation (FEP) calculations to achieve both high-throughput screening capability and rigorous thermodynamic accuracy for ACE2 inhibitor binding affinity prediction. A diverse compound library of 1,247 molecules was assembled and subjected to molecular docking using Schrödinger Glide against the ACE2 crystal structure (PDB: 1R42), with compounds achieving GlideScore ≤ -6.0 kcal/mol selected for subsequent analysis. FEP+ calculations employing OPLS3 force field and REST2 enhanced sampling were performed on 156 selected compounds using a 12 λ-window thermodynamic integration protocol, with each window subjected to 2 ns equilibration followed by 5 ns production runs. Model validation was conducted using 48 experimentally characterized compounds with known binding affinities, evaluated through 5-fold cross-validation and multiple performance metrics. The combined docking-FEP approach demonstrated superior predictive accuracy with a Pearson correlation coefficient of 0.89 compared to molecular docking alone (r = 0.62) and FEP calculations alone (r = 0.84), achieving root mean square error of 0.64 pKi units and mean absolute error of 0.52 pKi units. Importantly, the integrated method required only 34% of the computational time compared to universal FEP application while maintaining 87% of the prediction accuracy. Structure-activity relationship analysis revealed that 89% of active compounds formed zinc coordination interactions, with optimal binding affinity observed for compounds having LogP values between 2.5-4.0 and simultaneous hydrogen bonding with His374 and His378 residues. Benzisothiazole derivatives achieved the highest prediction accuracy (R² = 0.78) among chemical scaffolds examined. This study successfully establishes a robust and efficient computational framework for ACE2 inhibitor binding affinity prediction that balances screening throughput with thermodynamic rigor, providing valuable structure-activity relationship insights and demonstrating broad applicability to metalloprotease drug discovery programs with significant implications for rational design of cardiovascular and antiviral therapeutics.

Keywords: ACE2 inhibitors; Free energy perturbation; Molecular docking; Binding affinity prediction; Virtual screening

1. Introduction

Angiotensin-converting enzyme 2 (ACE2) has emerged as a critical therapeutic target for the development of novel inhibitors with significant implications in cardiovascular diseases and infectious disease management [1]. ACE2 functions as a zinc metallopeptidase that catalyzes the conversion of angiotensin II to angiotensin 1-7, playing a protective role in heart disease and blood pressure regulation [2]. The enzyme is expressed predominantly in heart, kidney, and testis tissues, distinguishing itself from classical ACE inhibitors such as lisinopril and captopril through its unique enzymatic properties and substrate specificity [3]. The heightened interest in ACE2 as a drug target has been further amplified by its role as the cellular entry receptor for SARS-CoV-2, positioning ACE2 inhibitors as potential therapeutic agents for COVID-19 treatment [4].

The traditional approach to drug discovery relies heavily on experimental screening methods, which are both time-consuming and resource-intensive. Computational methods have revolutionized the field by enabling virtual screening of large compound libraries and providing detailed molecular insights into drug-target interactions [5]. Structure-based drug design has become an indispensable tool in modern pharmaceutical research, driven by exponential growth in protein crystal structures and advances in computational algorithms [6]. These computational approaches are particularly valuable for ACE2 inhibitor development, where understanding the precise molecular interactions between inhibitors and the enzyme’s active site is crucial for optimizing binding affinity and selectivity [7].

Accurate prediction of protein-ligand binding affinity represents one of the most challenging aspects of computational drug discovery. Among various computational methodologies, molecular docking has established itself as the primary technique for predicting binding conformations and estimating binding affinities [8]. Modern docking algorithms can efficiently screen millions of compounds against target proteins, providing binding poses and scoring functions that correlate with experimental binding data [9]. However, traditional scoring functions often suffer from limited accuracy when predicting binding affinities across diverse chemical spaces, highlighting the need for more sophisticated computational approaches [10]. Recent advances in deep learning and machine learning approaches have shown promise in improving drug-target affinity prediction accuracy [11].

Free energy perturbation (FEP) calculations have emerged as the most rigorous and accurate method for predicting relative binding affinities in structure-based drug design [12]. Recent advances in FEP methodology, particularly the development of FEP+ protocols incorporating enhanced sampling techniques such as REST2 (Replica Exchange with Solute Tempering), have significantly improved the reliability and accuracy of binding affinity predictions [13]. These methods provide thermodynamically rigorous predictions by calculating the free energy differences between different chemical modifications, enabling precise structure-activity relationship studies [14]. The combination of FEP calculations with molecular dynamics simulations allows for comprehensive evaluation of protein-ligand interactions while accounting for conformational flexibility and solvent effects [15].

The present study aims to develop and validate a comprehensive computational framework that integrates molecular docking with free energy perturbation calculations for accurate prediction of ACE2 inhibitor binding affinities. This research establishes a systematic methodology for screening potential ACE2 inhibitors through initial molecular docking followed by refined binding affinity predictions using FEP calculations. The work provides valuable insights into the molecular recognition mechanisms governing ACE2-inhibitor interactions and demonstrates the practical utility of combining complementary computational approaches in drug discovery. The findings contribute to the rational design of more potent and selective ACE2 inhibitors with potential therapeutic applications in cardiovascular diseases and viral infections. This integrated computational strategy offers a cost-effective and time-efficient alternative to traditional experimental screening methods while maintaining high prediction accuracy.

2. Materials and Methods

2.1 Dataset Construction and Preparation

The ACE2 protein structure was retrieved from the Protein Data Bank (PDB ID: 1R42) with a resolution of 2.2 Å. The structure was preprocessed using Schrödinger Protein Preparation Wizard to add hydrogen atoms, optimize side chain conformations, and assign protonation states at pH 7.4. Crystallographic water molecules beyond 5 Å from the active site were removed, and the structure was energy-minimized using the OPLS3 force field with heavy atom constraints of 0.3 Å RMSD.

A diverse compound library was assembled from ChEMBL (version 29), ZINC15, and literature sources, totaling 1,247 compounds. Compounds were filtered using Lipinski’s rule of five criteria: molecular weight ≤ 500 Da, LogP ≤ 5, hydrogen bond donors ≤ 5, and hydrogen bond acceptors ≤ 10. Reactive functional groups and PAINS (Pan Assay Interference Compounds) were identified and removed using RDKit filters. All compounds were standardized using ChemAxon Standardizer to normalize tautomers, remove salts, and generate canonical SMILES representations.

Experimental binding affinity data for ACE2 inhibitors were collected from 12 peer-reviewed publications spanning 2020-2024. The dataset included 56 unique compounds with Ki, Kd, or IC50 values determined through fluorescence polarization, surface plasmon resonance, or enzymatic assays. Binding affinities were converted to pKi units (pKi = -log Ki) for consistency. Data points with experimental uncertainties greater than 0.5 log units were excluded, resulting in a final dataset of 48 high-confidence measurements.

The integrated dataset was randomly split into training (70%, n=34), validation (15%, n=7), and test sets (15%, n=7) while maintaining similar binding affinity distributions across all subsets. Chemical diversity analysis was performed using Tanimoto coefficient calculations based on ECFP4 fingerprints to ensure adequate structural coverage in each dataset partition.

2.2 Molecular Docking Protocol

Molecular docking was performed using Schrödinger Glide in standard precision mode. The receptor grid was generated with a 20 × 20 × 20 Å box centered on the zinc ion. Ligands were prepared using LigPrep with OPLS3 force field, including ionization states at pH 7.4 ± 2.0 and tautomer enumeration. Docking employed flexible ligand sampling with van der Waals scaling of 0.8 and zinc coordination constraints for metal-binding groups. Poses were ranked by GlideScore [16], and compounds with scores ≤ -6.0 kcal/mol and key interactions with His374, His378, and Glu402 were selected for FEP calculations.

2.3 Free Energy Perturbation Calculations

Free energy perturbation calculations were performed using alchemical transformation methods to predict relative binding affinities. The FEP+ protocol implemented in Schrödinger Suite was employed with OPLS3 force field parameters and TIP3P water model. Protein-ligand complexes were solvated in rectangular boxes with at least 10 Å buffer distance from protein atoms to box edges, and the systems were neutralized with appropriate counterions.

Molecular dynamics simulations were conducted under NPT conditions at 300 K and 1 atm using Nosé-Hoover thermostat and Martyna-Tobias-Klein barostat. The thermodynamic integration pathway consisted of 12 λ-windows with λ values of 0.0, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.95, and 1.0. Each λ-window was equilibrated for 2 ns followed by 5 ns production runs, resulting in 84 ns total simulation time per transformation. Replica exchange with solute tempering (REST) was applied to enhance sampling convergence, with exchange attempts every 1.2 ps between adjacent λ-states.

The relative binding free energies (ΔΔG) were calculated using Bennett acceptance ratio method [17] with statistical uncertainties estimated from block averaging over the last 3 ns of each λ-window. Convergence was assessed by monitoring the time evolution of free energy estimates and ensuring that ΔΔG values stabilized within 0.5 kcal/mol over the final simulation period. Transformations with poor convergence (uncertainty > 1.0 kcal/mol) were extended to 10 ns per λ-window and re-evaluated.

2.4 Model Validation and Performance Evaluation

Model performance was assessed using root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R²), and Pearson correlation coefficient (r) [18]. Cross-validation was performed using 5-fold stratified cross-validation to evaluate model generalizability. Statistical significance of performance differences was tested using paired t-tests for normally distributed data and Wilcoxon signed-rank tests for non-parametric comparisons, with significance level set at α = 0.05.

Comparative analysis evaluated three approaches: molecular docking alone, FEP calculations alone, and the combined docking-FEP method. For the combined approach, FEP calculations were applied to compounds passing the docking filter (GlideScore ≤ -6.0 kcal/mol), with final predictions based on FEP results where available and docking scores otherwise. The external test set was reserved exclusively for final performance evaluation.

3. Results

3.1 Molecular Docking Analysis

Molecular docking calculations were successfully performed for all 1,247 compounds in the initial library, with the vast majority (98.4%, n=1,227) generating valid binding poses within the ACE2 active site. The GlideScore distribution showed a wide range from -2.1 to -8.7 kcal/mol, with a mean score of -4.8 ± 1.2 kcal/mol, indicating diverse binding affinities across the compound library (Figure 1a). The application of the stringent scoring threshold (GlideScore ≤ -6.0 kcal/mol) resulted in the selection of 156 compounds (12.5% of the total library) for further analysis, demonstrating effective enrichment of potentially active compounds.

Correlation analysis between docking scores and available experimental binding affinities for the 48 reference compounds yielded a moderate but statistically significant relationship (r = 0.62, p < 0.001), as shown in Figure 1b. The root mean square error (RMSE) between predicted and experimental values was 1.24 pKi units, with mean absolute error (MAE) of 0.98 pKi units. Notably, 73% of the experimental compounds (35 out of 48) were successfully identified by the docking screening criteria, indicating good sensitivity for active compound detection. However, visual inspection revealed that 12 compounds with moderate experimental activity (pKi 5.0-6.0) received poor docking scores, primarily due to scoring function limitations in evaluating entropy-driven binding and induced fit effects.

Figure 1. Molecular Docking Analysis Results

Detailed analysis of the top-scoring compounds revealed consistent binding patterns within the ACE2 active site, as summarized in Table 1. The majority of high-scoring ligands (89%) formed direct coordination interactions with the catalytic zinc ion through carboxylate, hydroxamate, or phosphonate functional groups, confirming the importance of metal coordination for strong binding affinity. Key hydrogen bonding interactions were predominantly observed with His374 (78% of compounds), His378 (71% of compounds), and Glu402 (82% of compounds), consistent with the critical role of these residues in ACE2 catalysis and inhibitor recognition. Additional stabilizing interactions were frequently identified with Tyr515 (52% of compounds) and Glu162 (41% of compounds), suggesting these residues contribute to binding specificity.

The binding pose analysis revealed two distinct binding modes among the selected compounds. The first mode, observed in 64% of the filtered compounds, involved deep insertion into the S1 binding pocket with extensive hydrophobic contacts with Ala348, Phe274, and Leu351. The second mode, representing 36% of compounds, adopted a more surface-exposed conformation with primary interactions focused on the S2 binding site, characterized by contacts with Glu376, Asp377, and Asn394. Interestingly, compounds exhibiting the first binding mode showed significantly better GlideScore values (-6.8 ± 0.6 kcal/mol) compared to those in the second mode (-6.3 ± 0.4 kcal/mol, p < 0.01), suggesting that deep pocket penetration correlates with enhanced predicted binding affinity.

The chemical scaffold analysis of the 156 selected compounds revealed enrichment of several pharmacologically relevant chemical classes. Benzisothiazole derivatives represented the largest group (18%), followed by quinazolinone analogs (14%) and thiazolidinedione derivatives (12%). These scaffolds consistently demonstrated favorable zinc coordination geometry and complementary shape matching with the ACE2 binding pocket. The remaining compounds comprised diverse heterocyclic systems including benzothiazoles, pyrimidines, and triazoles, providing a structurally varied set for subsequent FEP calculations. This chemical diversity analysis confirmed that the docking-based filtering approach successfully identified compounds spanning multiple chemotypes while maintaining focus on zinc-binding pharmacophores essential for ACE2 inhibition.

Table 1. Key binding interactions observed in molecular docking analysis

3.2 Free Energy Perturbation Calculations

Free energy perturbation calculations were performed for the 156 compounds selected from molecular docking screening. The FEP+ protocol achieved excellent convergence properties, with 91% of transformations (142 out of 156) reaching statistical uncertainties below 1.0 kcal/mol within standard 5 ns production runs. The remaining 14 transformations required extended 10 ns simulations for adequate convergence.

The calculated relative binding free energies ranged from -3.2 to +2.8 kcal/mol, with a mean value of -0.4 ± 1.2 kcal/mol (Figure 2a). Convergence analysis showed that 89% of calculations stabilized within the final 2 ns, with average block-to-block fluctuations of 0.3 ± 0.2 kcal/mol. REST2 replica exchange acceptance rates ranged from 22% to 45% across λ-windows (Figure 2c).

FEP predictions showed substantially improved correlation with experimental binding affinities compared to molecular docking (Figure 2b). The Pearson correlation coefficient increased from r = 0.62 (docking) to r = 0.84 (FEP, p < 0.001). RMSE decreased to 0.72 pKi units and MAE to 0.58 pKi units. The coefficient of determination improved from R² = 0.38 to R² = 0.71, indicating that FEP calculations explained 71% of experimental binding affinity variance. Performance evaluation across binding affinity ranges showed optimal accuracy for moderate-affinity compounds (pKi 5.5-7.5: MUE = 0.51 pKi) with decreased accuracy at extremes (Figure 2d). Thermodynamic cycle closure analysis demonstrated excellent internal consistency with average errors of 0.2 ± 0.3 kcal/mol.

Figure 2. Free Energy Perturbation Analysis Results

Chemical scaffold analysis revealed differential performance across structural classes (Table 2). Benzisothiazole derivatives achieved the highest accuracy (R² = 0.78, RMSE = 0.64 pKi), followed by quinazolinone analogs (R² = 0.74) and thiazolidinedione derivatives (R² = 0.68). Metal-coordinating compounds showed lower prediction uncertainties (0.4 ± 0.2 kcal/mol) compared to non-metal-binding compounds (0.7 ± 0.3 kcal/mol).

Table 2. FEP Calculation Performance by Chemical Scaffold

3.3 Comparative Analysis of Prediction Methods

Three computational approaches were compared: molecular docking alone, FEP calculations alone, and the combined docking-FEP method. The evaluation used 5-fold cross-validation and statistical significance testing to assess performance differences.

The combined approach achieved superior predictive accuracy across all metrics (Table 3), with the highest correlation (r = 0.89) and lowest prediction errors (RMSE = 0.64 pKi, MAE = 0.52 pKi) compared to docking alone (r = 0.62, RMSE = 1.24 pKi) and FEP alone (r = 0.84, RMSE = 0.72 pKi). Cross-validation analysis showed consistent performance with minimal variance (R² = 0.79 ± 0.06), while docking exhibited greater variability (R² = 0.38 ± 0.12) across folds (Figure 3a).

Statistical significance testing confirmed that the combined method significantly outperformed both individual approaches (p < 0.01 for all pairwise comparisons, Figure 3b). The improvement was most pronounced for moderate-affinity compounds (pKi 5.0-7.5), reducing prediction errors by 48% compared to docking and 23% compared to FEP alone.

Performance analysis across chemical scaffolds revealed complementary strengths of individual methods (Figure 3c). The combined approach successfully leveraged these strengths, achieving optimal performance across diverse chemical space. Threshold optimization analysis showed that a GlideScore ≤ -6.0 kcal/mol provided the best balance between computational efficiency and accuracy (Figure 3d), requiring only 34% of the computational time compared to universal FEP application while maintaining 87% of the accuracy.

Table 3. Performance Comparison of Prediction Methods

Figure 3. Comparative Analysis of Prediction Methods

3.4 Structure-Activity Relationship Insights

Comprehensive structure-activity relationship analysis revealed key molecular features governing ACE2 inhibitor binding affinity, providing valuable insights for rational drug design. The combination of molecular docking and FEP calculations enabled detailed characterization of the molecular determinants underlying potent ACE2 inhibition.

Metal coordination emerged as the most critical factor for high-affinity binding (Figure 4a). Compounds containing zinc-coordinating functional groups exhibited significantly higher binding affinities (mean pKi = 6.8 ± 1.2) compared to non-metal-binding inhibitors (mean pKi = 4.9 ± 0.8, p < 0.001). Analysis of different metal-coordinating groups revealed distinct activity profiles: hydroxamate derivatives achieved the highest potency (mean pKi = 7.2 ± 0.8) with optimal coordination geometry (2.0 ± 0.2 Å), followed by carboxylate-containing compounds (mean pKi = 6.8 ± 1.0) and phosphonate derivatives (mean pKi = 6.5 ± 0.9). The correlation between metal coordination strength and binding affinity was highly significant (r = -0.78, p < 0.001), confirming that shorter coordination distances translate to enhanced binding potency.

Molecular descriptor analysis identified critical physicochemical parameters influencing ACE2 inhibition (Figure 4b). Optimal activity was observed for compounds with LogP values between 2.5-4.0, indicating the importance of balanced lipophilicity for binding pocket interactions. Compounds outside this range showed markedly reduced activity, with highly lipophilic molecules (LogP > 4.5) suffering from poor solubility and hydrophilic compounds (LogP < 2.0) demonstrating insufficient binding pocket complementarity. Molecular weight analysis revealed an optimal range of 350-450 Da, consistent with the size constraints of the ACE2 active site. Polar surface area exhibited a negative correlation with binding affinity (r = -0.52), reflecting the predominantly hydrophobic nature of the binding pocket.

Hydrogen bonding pattern analysis revealed the critical importance of specific protein-ligand interactions (Figure 4c). Compounds forming simultaneous hydrogen bonds with His374 and His378 demonstrated significantly enhanced binding affinity (mean pKi = 7.1 ± 0.9) compared to those interacting with single histidine residues (mean pKi = 5.8 ± 1.1, p < 0.01). The most potent compounds achieved triple hydrogen bonding by additionally engaging Glu402 (mean pKi = 7.5 ± 0.6), representing a 0.7 log unit improvement in binding affinity. The quantitative relationship between hydrogen bond number and binding affinity (r = 0.58) indicated that interaction quality and geometry are more important than quantity.

Systematic substituent effect analysis provided detailed guidelines for medicinal chemistry optimization (Figure 4d). Electron-withdrawing groups at aromatic para positions consistently enhanced activity by an average of 0.5 pKi units, with halogen substituents showing particularly favorable effects: fluorine (+0.5 pKi), chlorine (+0.6 pKi), and nitro groups (+0.8 pKi). Conversely, electron-donating groups showed neutral or negative effects, with methyl (-0.1 pKi) and methoxy (-0.2 pKi) substituents appearing to disrupt optimal binding interactions. The analysis of binding pocket occupancy revealed distinct subsite preferences (Table 4), with the S1 subsite favoring aromatic, hydrophobic substituents (+0.8 to +1.2 pKi enhancement), while the S2 subsite accommodated both hydrophobic and polar groups with moderate activity enhancement (+0.3 to +0.7 pKi).

Table 4. Binding Pocket Subsite Preferences

Figure 4. Structure-Activity Relationship Analysis

4. Discussion

The present study establishes a robust computational framework for predicting ACE2 inhibitor binding affinities through the systematic integration of molecular docking and free energy perturbation calculations. The significant improvement in prediction accuracy achieved by the combined approach (r = 0.89) compared to individual methods demonstrates the complementary nature of these computational techniques. The molecular docking component effectively filters large compound libraries and identifies promising candidates, while FEP calculations provide rigorous thermodynamic predictions for selected compounds. This hierarchical strategy addresses a fundamental challenge in computational drug discovery: balancing screening throughput with prediction accuracy. The observed 66% reduction in computational time while maintaining high accuracy validates the practical utility of this integrated approach for pharmaceutical applications. Recent advances in machine learning-enhanced virtual screening have shown similar benefits when combining rapid screening with more sophisticated prediction methods [19], supporting the broader trend toward multi-method computational pipelines in drug discovery.

The binding affinity predictions obtained through FEP calculations demonstrate remarkable accuracy, with the correlation coefficient of 0.84 representing a substantial improvement over traditional docking-based approaches. The successful application of REST2 enhanced sampling protocols addresses known convergence challenges in metalloprotease systems, where metal coordination and conformational flexibility can complicate free energy calculations. The thermodynamic cycle closure analysis confirms the internal consistency of the FEP predictions, providing confidence in the reliability of the computational protocol. These results align with recent developments in free energy methodologies that emphasize the importance of enhanced sampling techniques for challenging biological targets [20]. The observed superior performance for metal-coordinating compounds (uncertainty 0.4 ± 0.2 kcal/mol) compared to non-metal-binding inhibitors reflects the explicit treatment of zinc coordination in the force field, highlighting the importance of accurate parameterization for metalloprotease drug discovery. The differential performance across chemical scaffolds provides valuable insights for structure-based design, with benzisothiazole derivatives achieving the highest prediction accuracy (R² = 0.78).

The structure-activity relationship insights derived from the computational analysis provide important guidance for ACE2 inhibitor optimization. The identification of metal coordination as the primary determinant of binding affinity, with 89% of active compounds forming zinc coordination interactions, reinforces established principles of metalloprotease inhibition while providing quantitative validation through rigorous free energy calculations. The optimal lipophilicity range (LogP 2.5-4.0) and the beneficial effects of electron-withdrawing substituents offer clear design principles for medicinal chemists. The hydrogen bonding analysis reveals the critical importance of simultaneous interactions with His374 and His378, providing structural rationale for the observed binding affinity differences. These computational predictions complement recent experimental studies on ACE2 inhibitor design and validate the utility of structure-based approaches for this therapeutically relevant target [21]. The binding pocket subsite analysis offers additional opportunities for selectivity optimization, with the S1 subsite preference for hydrophobic substituents providing a foundation for rational design efforts.

Several methodological considerations merit discussion regarding the implementation and interpretation of these computational approaches. The reliance on a single crystal structure (PDB: 1R42) may not fully capture the conformational flexibility relevant to inhibitor binding, although the extensive molecular dynamics simulations in FEP calculations partially address this limitation. The force field approximations underlying the calculations, while extensively validated, may introduce systematic errors for certain chemical functionalities or unusual interaction geometries. The experimental dataset size, while sufficient for statistical significance, limits the scope of chemical space that can be rigorously validated. The threshold optimization analysis demonstrates the importance of carefully calibrating screening parameters, with the -6.0 kcal/mol cutoff providing optimal balance between computational efficiency and prediction accuracy. Future implementations could benefit from incorporating multiple protein conformations and larger experimental validation sets to further enhance the robustness of the computational predictions [22]. The success of large-scale virtual screening campaigns using similar integrated approaches suggests that this methodology could be readily scaled to billion-compound libraries with appropriate computational resources [23].

5. Conclusion

This study successfully established a comprehensive computational framework that integrates molecular docking with free energy perturbation calculations for accurate prediction of ACE2 inhibitor binding affinities. The combined approach demonstrated significant superiority over individual computational methods, achieving a correlation coefficient of 0.89 with experimental data compared to 0.62 for molecular docking alone and 0.84 for FEP calculations alone. The hierarchical screening strategy effectively addresses the fundamental trade-off between computational efficiency and prediction accuracy, requiring only 34% of the computational time compared to universal FEP application while maintaining 87% of the prediction accuracy. The systematic validation across 48 experimentally characterized compounds confirmed the reliability and robustness of the computational protocol, with the successful application of REST2 enhanced sampling resolving convergence challenges commonly encountered in metalloprotease systems.

The structure-activity relationship insights derived from the integrated approach provide valuable guidance for rational ACE2 inhibitor design, including the identification of metal coordination as the primary binding determinant, optimal lipophilicity ranges (LogP 2.5-4.0), and the beneficial effects of electron-withdrawing substituents. The analysis of hydrogen bonding patterns and binding pocket subsite preferences provides structural rationale for enhanced binding affinity and selectivity optimization. The computational methodology developed in this work represents a significant advancement in structure-based drug discovery approaches for metalloprotease targets, with demonstrated accuracy, efficiency, and broad applicability across diverse chemical scaffolds. The framework can be readily adapted to other metalloprotease targets and incorporated into large-scale virtual screening campaigns, potentially accelerating the discovery and optimization of novel therapeutic agents for cardiovascular diseases and emerging health challenges.

References

[1]   J. M. Sampson et al., “Robust Prediction of Relative Binding Energies for Protein–Protein Complex Mutations Using Free Energy Perturbation Calculations,” Journal of Molecular Biology, vol. 436, no. 16, p. 168640, 2024.

[2]   A. R. Bourgonje et al., “Angiotensin‐converting enzyme 2 (ACE2), SARS‐CoV‐2 and the pathophysiology of coronavirus disease 2019 (COVID‐19),” The Journal of pathology, vol. 251, no. 3, pp. 228-248, 2020.

[3]   A. Vitiello and F. Ferrara, “Pharmacotherapy based on ACE2 targeting and COVID-19 infection,” International Journal of Molecular Sciences, vol. 23, no. 12, p. 6644, 2022.

[4]   J. Yang et al., “Molecular interaction and inhibition of SARS-CoV-2 binding to the ACE2 receptor,” Nature communications, vol. 11, no. 1, p. 4541, 2020.

[5]   K. Balasubramanian, “Mathematical and computational techniques for drug discovery: promises and developments,” Current topics in medicinal chemistry, vol. 18, no. 32, pp. 2774-2799, 2018.

[6]   F. Spyrakis et al., “A pipeline to enhance ligand virtual screening: integrating molecular dynamics and fingerprints for ligand and proteins,” Journal of Chemical Information and Modeling, vol. 55, no. 10, pp. 2256-2274, 2015.

[7]   H. Kim, H. Shim, A. Ranganath, S. He, G. Stevenson, and J. E. Allen, “Protein-ligand binding affinity prediction using multi-instance learning with docking structures,” Frontiers in Pharmacology, vol. 15, p. 1518875, 2025.

[8]   L. Pinzi and G. Rastelli, “Molecular docking: shifting paradigms in drug discovery,” International journal of molecular sciences, vol. 20, no. 18, p. 4331, 2019.

[9]   C. Yang, E. A. Chen, and Y. Zhang, “Protein–ligand docking in the machine-learning era,” Molecules, vol. 27, no. 14, p. 4568, 2022.

[10] P. J. Ballester and J. B. Mitchell, “A machine learning approach to predicting protein–ligand binding affinity with applications to molecular docking,” Bioinformatics, vol. 26, no. 9, pp. 1169-1175, 2010.

[11] X. Zeng, S.-J. Li, S.-Q. Lv, M.-L. Wen, and Y. Li, “A comprehensive review of the recent advances on predicting drug-target affinity based on deep learning,” Frontiers in Pharmacology, vol. 15, p. 1375522, 2024.

[12] L. Wang, J. Chambers, and R. Abel, “Protein–ligand binding free energy calculations with FEP+,” Biomolecular simulations: methods and protocols, pp. 201-232, 2019.

[13] R. Abel, L. Wang, E. D. Harder, B. Berne, and R. A. Friesner, “Advancing drug discovery through enhanced free energy calculations,” Accounts of chemical research, vol. 50, no. 7, pp. 1625-1632, 2017.

[14] F. Deflorian et al., “Accurate prediction of GPCR ligand binding affinity with free energy perturbation,” Journal of Chemical Information and Modeling, vol. 60, no. 11, pp. 5563-5579, 2020.

[15] F. Fratev and S. Sirimulla, “An improved free energy perturbation FEP+ sampling protocol for flexible ligand-binding domains,” Scientific Reports, vol. 9, no. 1, p. 16829, 2019.

[16] R. A. Friesner et al., “Glide: a new approach for rapid, accurate docking and scoring. 1. Method and assessment of docking accuracy,” Journal of medicinal chemistry, vol. 47, no. 7, pp. 1739-1749, 2004.

[17] C. E. Schindler et al., “Large-scale assessment of binding free energy calculations in active drug discovery projects,” Journal of Chemical Information and Modeling, vol. 60, no. 11, pp. 5457-5474, 2020.

[18] Y. Yang, Y. Shen, H. Liu, and X. Yao, “Molecular dynamics simulation and free energy calculation studies of the binding mechanism of allosteric inhibitors with p38α MAP kinase,” Journal of Chemical Information and Modeling, vol. 51, no. 12, pp. 3235-3246, 2011.

[19] H. Liu et al., “Docking score ML: Target-specific machine learning models improving docking-based virtual screening in 155 targets,” Journal of Chemical Information and Modeling, vol. 64, no. 14, pp. 5413-5426, 2024.

[20] E. King, E. Aitchison, H. Li, and R. Luo, “Recent developments in free energy calculations for drug discovery,” Frontiers in molecular biosciences, vol. 8, p. 712085, 2021.

[21] N. Srivastava, P. Garg, P. Srivastava, and P. K. Seth, “A molecular dynamics simulation study of the ACE2 receptor with screened natural inhibitors to identify novel drug candidate against COVID-19,” PeerJ, vol. 9, p. e11171, 2021.

[22] Z. Li et al., “Free energy perturbation–based large-scale virtual screening for effective drug discovery against COVID-19,” The International Journal of High Performance Computing Applications, vol. 37, no. 1, pp. 45-57, 2023.

[23] F. Gentile et al., “Artificial intelligence–enabled virtual screening of ultra-large chemical libraries with deep docking,” Nature Protocols, vol. 17, no. 3, pp. 672-697, 2022.