Beyond Pose Accuracy: Reproducibility-Oriented Framework For Interpreting Web-Based Molecular Docking Outputs Using Decoupled Localization and Pose Fidelity Metrics

Beyond Pose Accuracy: Reproducibility-Oriented Framework For Interpreting Web-Based Molecular Docking Outputs Using Decoupled Localization and Pose Fidelity Metrics | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Beyond Pose Accuracy: Reproducibility-Oriented Framework For Interpreting Web-Based Molecular Docking Outputs Using Decoupled Localization and Pose Fidelity Metrics Maniratnam Puli, Venkata Ramana Singamaneni, Nikitha Bennuri, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8908837/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 7 You are reading this latest preprint version Abstract Context Web-based molecular docking servers are widely used in structure-based modeling due to their accessibility, yet their outputs are commonly evaluated using single-run pose accuracy metrics that implicitly assume deterministic behavior. Such assumptions can obscure variability arising from stochastic sampling and ranking procedures, particularly in black-box usage scenarios where replicate validation is uncommon. In this study, we introduce a reproducibility-aware geometric evaluation framework designed to disentangle binding-site localization, fixed-frame pose accuracy, and run-to-run stability. Application of this framework to representative protein–ligand systems demonstrates that accurate pose generation, correct site localization, and reproducible pose ranking are separable properties. The results illustrate that pose accuracy alone does not fully characterize the practical reliability of web-based docking outputs and motivate inclusion of stability-aware metrics in docking validation workflows. Methods Five widely used web-based docking servers representing distinct algorithmic paradigms were evaluated, including a Vina-class reference implementation (MolModa), stochastic sampling platforms (GalaxyDock, SwissDock, PaRDOCK), and a geometry-driven rigid-body approach (PatchDock). Docking outputs were analyzed using complementary geometric metrics: centroid-based localization to assess binding-site placement, in-place root-mean-square deviation (RMSD) to evaluate fixed-frame pose accuracy, and aligned RMSD to quantify conformational agreement independent of spatial placement. Operational stability was assessed using triplicate docking runs under unbiased re-docking conditions. All geometric analyses were performed using custom Python script designed to ensure coordinate-frame consistency and symmetry-aware RMSD evaluation. Web-based docking servers Molecular docking Active site localization Pose accuracy Reproducibility Algorithmic stability Figures Figure 1 Figure 2 Figure 3 1. Introduction 1.1 Molecular Docking Benchmarks: From Algorithms to Web Servers Molecular docking is a cornerstone of structure-based drug design, enabling prediction of ligand binding modes and the virtual screening of large compound libraries [ 1 , 2 ]. Historically, rigorous benchmarking has focused on standalone docking engines such as DOCK, FlexX, GLIDE, and Surflex, where users retain granular control over sampling parameters, grid definitions, and scoring functions [ 3 ]. Advances in algorithmic efficiency and computational infrastructure have since driven a transition toward automated, web-based docking platforms that allow non-expert users to perform complex simulations without local software installations [ 2 ]. While this democratization has broadened access, it has also introduced a “black-box” challenge, as users often lack visibility into underlying search parameters, stochastic elements such as random seeds, and server-specific failure modes. As a result, the focus of validation shifts from evaluating theoretical algorithmic capability to assessing the operational reliability and reproducibility of docking services, which remains a persistent challenge in computational drug discovery and necessitates transparent workflows and reproducible evaluation strategies [ 4 ] 1.2 Lessons from Classical Benchmarks The evaluation of docking performance therefore requires standardized and carefully controlled benchmarking protocols to ensure meaningful methodological comparison. In a seminal study, Cross et al. (2009) established a rigorous framework for assessing pose accuracy and virtual screening accuracy across six major docking programs [ 5 ]. Their benchmark emphasized that the primary criterion for docking success is the ability to reproduce the crystallographic ligand pose within a Root-Mean-Square Deviation (RMSD) of 2.0 Å. Importantly, Cross et al. demonstrated that absolute docking scores are rarely comparable across different programs and frequently fail to correlate linearly with RMSD, indicating that internal pose ranking consistency is often a more reliable indicator of performance than raw energy values. Their methodology—leveraging high-resolution datasets such as the Astex diverse set and prioritizing geometric fidelity over binding-energy prediction—continues to serve as a foundational reference for docking validation strategies [ 6 ]. 1.3 Gaps in Current Web Docking Evaluations When classical benchmarking concepts such as those proposed by Cross et al. are applied to modern web-based docking servers, several methodological gaps become apparent. Most existing evaluations implicitly treat web servers as deterministic tools, reporting results from a single docking run. This practice overlooks the stochastic nature of many contemporary search algorithms, including simulated annealing and genetic-algorithm–based approaches, for which repeated submissions using identical inputs may yield divergent solutions due to random initialization or server-side execution variability [ 7 ]. In addition, web-based docking platforms differ substantially in how binding-site information is specified and internally interpreted, ranging from Cartesian grid definitions to reference-ligand guidance, residue-based constraints, or geometry-driven matching. Such heterogeneity introduces user-invisible sources of variability that may influence docking outcomes even when nominally equivalent binding-site inputs are provided [ 8 , 9 ]. Collectively, these considerations suggest that modern validation efforts should extend beyond conformational accuracy alone and incorporate explicit assessment of binding-site localization and algorithmic reproducibility across repeated docking runs. 1.4 Objectives of the Present Study In this study, we evaluate five distinct web-based docking servers—SwissDock, ParDOCK, GalaxyDock, PatchDock, and MolModa (representing a Vina-class reference implementation)—using the high-resolution Astex diverse set. By adapting the rigorous pose-recovery protocols of Cross et al. (2009) to the practical constraints of web-based services, we demonstrate a reproducibility-aware evaluation framework rather than a definitive performance ranking, we pursue two primary objectives: Aim 1 : To demonstrate a geometric evaluation strategy that decouples binding-site localization from pose fidelity by applying complementary distance-based measures (centroid distance, in-place RMSD, and aligned RMSD) to top-ranked docked structures relative to experimental co-crystal references. Aim 2 : To demonstrate a reproducibility-aware evaluation strategy by applying triplicate docking and inter-run RMSD analysis to characterize run-to-run variability and illustrate how algorithmic stability influences the interpretation of docking results. Reliable identification of ligand binding modes underpins mechanistic interpretation of protein–ligand recognition, functional annotation, and hypothesis generation in structural biology. Therefore, assessing the biological reliability of widely used docking servers is critical for downstream interpretation. The objective of this work is to introduce and demonstrate a reproducibility-aware evaluation framework for docking reliability, rather than to provide a definitive ranking of docking servers. 2. Materials and Methods 2.1. Protein–Ligand Dataset To provide a representative testbed for demonstrating the proposed evaluation framework, we utilized the Astex Diverse set [ 6 ], a curated collection of high-resolution protein–ligand complexes spanning major protein classes, including enzymes, ATP-binding proteins, and nuclear receptors, designed to minimize structural redundancy while maximizing chemical diversity From the Astex diverse set, protein-ligand complexes were selected according to the following criteria: Crystallographic resolution < 2.0Å, one nuclear receptor has 2.10 Å Absence of covalent protein-ligand interactions: No missing, unresolved residues and mutations within the binding site; Absence of metal-mediated covalent coordination to the ligand; and Availability of a single, well-defined bound ligand Inclusion of these varied architectures reduces bias toward specific binding pocket geometrics, such as deeply buried active sites versus shallow surface grooves. A total of ten protein-ligand complexes were selected for this evaluation to balance chemical and structural diversity with methodological control. Selected Complexes are listed in Table 1 . Table 1 10 X-ray Complexes Used for Cognate Ligand Docking S.no Protein class PDB ID Ligand Resolution (Å) Reference for protein PDB 1. Enzyme 1ia1 TQ3 1.70 [ 10 ] 2. Enzyme 1s3v TQD 1.80 [ 11 ] 3. Enzyme 1t40 ID5 1.80 [ 12 ] 4. ATPase 1opk P16 1.80 [ 13 ] 5. ATPase 1t46 ST 1.60 [ 14 ] 6. ATPase 1ywr L19 1.95 [ 15 ] 7. Nuclear receptor 1s19 MC9 2.10 [ 16 ] 8. Nuclear receptor 1sj0 E4D 1.90 [ 17 ] 9. Nuclear receptor 1sqn NDR 1.45 [ 18 ] 10. Nuclear receptor 1z95 198 1.80 [ 19 ] 2.2. Protein–Ligand preparation Prior to docking, all complexes were prepared to generate a consistent apo-like receptor state. The crystallographic ligand was extracted, protein structures were prepared using UCSF ChimeraX Dock Prep by solvent removal, residue standardization, hydrogen addition, and charge assignment while keeping all heavy atoms fixed at their crystallographic positions. No energy minimization of protein heavy atoms was performed [ 20 ] to ensure appropriate electrostatic treatment. The crystallographic ligand extracted from each complex was saved as an independent file without additional energy minimization and was retained as the reference structure for RMSD calculations. To preserve atom ordering and chemical topology, the same crystallographic ligand was spatially displaced and reused as the docking ligand. ChimeraX was selected for preparation due to its deterministic pipeline, ensuring transparent handling of hydrogens and residue states while strictly preserving crystallographic coordinates. 2.3. Docking Servers Evaluated Docking serves were selected according to the following inclusion criteria Public accessibility , defined as free Web-based availability without the need for local software installation or proprietary licensing; Domain applicability , requiring explicit support for small-molecule protein-ligand docking; Active-site-guided sampling , Whereby the docking region can be constrained through user-specified Cartesian coordinates, binding-site residues, or reference-ligand guidance, thereby avoiding blind docking behaviour; Algorithmic diversity , such that the selected servers employ distinct docking paradigms to minimize methodological redundancy; Community adoption , Evidence by prior usage and citation in the computational chemistry literature; and Practical usability , requiring that individual docking jobs complete within a reasonable time frame compatible with high-throughput docking workflows. Five publicly accessible web-based docking platforms were evaluated, selected to represent a diverse range of docking algorithms and search paradigms. MolModa (Baseline) : A browser-based docking platform implementing the AutoDock Vina scoring function and search algorithm. Prepared protein and ligand structures were uploaded to the MolModa interface, where protonation of both receptor and ligand was performed using the server’s internal preparation routine. Docking was carried out using the compound docking option with default search parameters. The docking grid was defined based on the crystallographic binding site and centered on the reference ligand. Grid dimensions were set to encompass the maximal extent of the crystallographic ligand with an additional 10.0 Å buffer (5.0 Å per axis). Docking calculations were performed on the protonated receptor–ligand system, and the top-ranked pose returned by the server was retained for subsequent analysis [ 21 – 25 ]. SwissDock : SwissDock was used as a web-based docking platform based on the EADock DSS algorithm. Prepared protein and ligand structures were submitted to the SwissDock server using the Attracting Cavities mode, which restricts docking to energetically favorable regions of the protein surface. The docking search space was defined using explicit Cartesian boundaries identical to those employed for MolModa. The grid was centered on the crystallographic reference ligand and extended to encompass the maximal ligand dimensions with an additional 10.0 Å buffer (5.0 Å per axis) ensuring sufficient sampling while avoiding artificial confinement to the crystal pose. [ 26 – 28 ] ParDOCK : ParDOCK was used as a web-based protein–ligand docking server employing an all-atom energy-based Monte Carlo sampling protocol performing rigid docking. The protein–reference ligand complex, prepared in UCSF ChimeraX, was submitted to ParDOCK to implicitly define the binding site. In addition, the docking ligand, generated by displacing the crystallographic ligand in ChimeraX, was provided as a separate input for docking. Docking was performed with the reference ligand guiding localization of the Monte Carlo search to the correct binding cavity. [ 29 ] GalaxyDock : GalaxyDock was accessed via the GalaxyWEB server for protein–ligand docking. Receptors were provided in PDB format and ligands in MOL2 format, with docking restricted to user-specified binding-site residues derived from the crystallographic complex. Docking was performed using the GalaxyDock BP2 protocol, which employs conformational space annealing (CSA ) for global pose exploration while allowing limited protein side-chain flexibility. Ligands were treated as fully flexible during docking. An energy-driven, grid-based scoring function was used to evaluate poses within the defined binding site. [ 30 , 31 ]. PatchDock : PatchDock is a geometry-based molecular docking webserver that performs rigid-body docking based on molecular surface shape complementarity and surface patch complementarity. Docking was performed by submitting the prepared protein and ligand structures in PDB format, along with a receptor binding-site file in TXT format to restrict docking transformations to the specified active-site region. It defines the docking region through surface-patch matching constrained to interacting residues identified from the experimental complex The top-ranked docked complexes are reported by the server for further analysis. [ 32 , 33 ]. Each complex was subjected to triplicate docking across five independent web-based servers, followed by pose alignment and multi-metric RMSD evaluation. This design resulted in a total of 150 individual docking evaluations, each requiring verification of binding-site localization and coordinate-frame consistency. The dataset size was chosen to allow granular, case-by-case inspection of docking behavior and reproducibility, which is central to demonstrating the proposed evaluation framework. Substantially expanding the dataset would limit such manual oversight without necessarily improving interpretability in the context of illustrating methodological behavior across heterogeneous web-based platforms. Metric justification. Because binding-site specification strategies differ substantially among web-based docking servers, a binding-site localization metric (centroid distance) was incorporated to verify that these heterogeneous approaches resulted in comparable confinement of docked ligands within the intended binding pocket [ 9 , 34 ]. Server-specific exclusions. For the 1Z95 system [ 19 ], ParDOCK consistently failed to progress beyond the ligand preparation stage despite correct molecular geometry, neutral formal charge, and repeated submission using standard preparation protocols. As this behavior was reproducible and appeared to reflect a server-side parsing limitation rather than a chemical or structural issue, this system was excluded from ParDOCK-specific analyses while being retained for all other servers. 2.4. Docking Protocol All docking calculations were performed using a standardized protocol designed to reflect typical “black-box” usage of web-based docking servers. For each server, docking was conducted using default parameters, and only the single top-ranked pose returned by each run was retained for analysis. No post-docking rescoring or energy minimization was applied, ensuring that pose evaluation relied solely on the server’s internal scoring and ranking procedures. Triplicate strategy- To assess algorithmic stability and reproducibility, each protein–ligand pair was docked in three independent, consecutive runs for every server using identical input structures. Repeated runs capture stochastic variability inherent to many docking algorithms (e.g., Monte Carlo sampling or simulated annealing) and enable quantification of run-to-run variance in pose accuracy. Restricting analysis to the top-ranked pose across all runs further mitigates ranking bias arising from differences in the number of poses generated by individual servers, ensuring that performance comparisons reflect algorithmic consistency rather than sampling depth. 2.5 Pose Evaluation Metrics (Aim1) Docking accuracy was evaluated using three complementary metrics capturing binding-site localization and pose accuracy: centroid distance, In-place RMSD, and aligned heavy-atom RMSD. Together, these metrics decouple errors in spatial localization from conformational inaccuracies. Docked complexes returned by ParDOCK were provided in a global coordinate frame differing from the reference structures. To enable meaningful RMSD comparisons, ParDOCK complexes were rigidly aligned to their corresponding reference proteins using the MatchMaker tool in UCSF ChimeraX [ 20 ], based on protein backbone atoms only; ligand atoms were excluded from the alignment. No alignment was required for other servers, which preserved the original receptor coordinate frame. 2.5.1 Centroid Distance (Binding-Site Localization) Centroid distance measures the spatial displacement of the docked ligand relative to the crystallographic binding site without any structural superposition. It was calculated as the Euclidean distance between the geometric centroids of the docked and reference ligand poses, defined as the arithmetic mean of all heavy-atom Cartesian coordinates. Because no alignment is performed, this metric reflects translational displacement within the binding pocket and is independent of ligand orientation or internal conformation. Centroid distances ≤ 2.0 Å were considered indicative of correct binding-site localization, with ≤ 1.0 Å classified as excellent localization consistent with the center-of-mass tolerance thresholds applied in recent kinase benchmarking studies Tripathi et al., 2025 [ 35 ]. 2.5.2 In-Place RMSD (Conformational Accuracy) In-place RMSD quantifies conformational accuracy within a fixed binding-site frame. RMSD was computed directly between corresponding heavy atoms of the docked and reference ligands without ligand superposition, using the receptor-aligned coordinate frame. This metric isolates errors arising from incorrect ligand orientation or conformation after correct site localization has been achieved. 2.5.3 Aligned Heavy-Atom RMSD (Overall Pose Accuracy) Aligned heavy-atom RMSD assesses overall pose similarity following optimal rigid-body superposition of the docked ligand onto the crystallographic reference. RMSD was calculated over corresponding heavy atoms after least-squares alignment, consistent with the classical definition of docking success used in established benchmarks such as Cross et al. A threshold of 2.0 Å was used as the primary success criterion. While the same numerical cutoff was applied to In-place RMSD, the two metrics capture distinct aspects of docking performance: aligned RMSD reflects optimal pose recovery, whereas In-place RMSD provides a stricter evaluation within a fixed spatial frame. A graphical representation of the three metrics has been given in Fig. 1 . The evaluation criteria for the metrics are given in Table 2 . Table 2 Acceptance threshold Metric ≤ 1.0 Å ≤ 2.0 Å > 2.0 Å Centroid distance Excellent localization Correct localization Wrong site In-place RMSD Near-native conformation Acceptable conformation Distorted pose Aligned RMSD High-accuracy pose Correct pose Incorrect pose 2.5.4 RMSD script RMSD-based metrics were calculated using custom Python script implementing RDkit which performs three measures: (i) centroid distance between geometric centers of docked and crystallographic ligands; (ii) in-place RMSD computed in the fixed receptor coordinate frame; and (iii) aligned RMSD obtained after optimal heavy-atom superposition of ligand coordinates, with RMSD < 2.0 Å considered as threshold. Script is symmetry-aware, incorporate atom-mapping procedures to ensure consistency across differing atom orders, and robust to ligand rotations and conformational variability. The script and the usage instructions have been provided in data repository accessible at - https://data.mendeley.com/datasets/sv7jcdc4xj/1 2.6 Reproducibility and Robustness Metrics (Aim 2) 2.6.1. Pose Reproducibility and algorithmic stability (Primary Robustness Metric) Pose reproducibility quantifies the consistency of a docking server across independent executions. For each protein–ligand pair, docking was performed in three independent runs using identical inputs and default parameters, retaining only the top-ranked pose from each run. Run-to-run consistency was assessed by computing pairwise heavy-atom RMSDs between top-ranked poses from all run pairs using In-place RMSD (no superposition) between triplicate docking runs (Run1–Run2, Run1–Run3, Run2–Run3). For each protein–ligand system, the median inter-run RMSD was calculated and aggregated across proteins to obtain server-level reproducibility statistics. Low median inter-run RMSD indicates a deterministically stable docking process, whereas high variance indicates stochastic sampling behavior or convergence failure 2.6.2. Robustness of correct pose prediction Robustness of correct pose prediction evaluates whether a docking server can reproducibly generate a correct ligand binding pose within the crystallographic binding site across repeated docking runs. Correct predictions are defined using In-place heavy-atom RMSD ≤ 2.0 Å relative to the experimental ligand structure, ensuring simultaneous correctness of binding-site localization, orientation, and conformation. For each protein–server pair, triplicate outcomes were collapsed into protein-level categories reflecting consistent (3/3), partial (2/3), sporadic (1/3), or absent (0/3) native pose recovery. Results were summarized as counts per server, enabling evaluation of reproducibility of correct docking outcomes rather than average error magnitude. 2.6.3. Relative Robustness Against a Baseline Server To provide a reference context for reproducibility behavior, robustness metrics were interpreted relative to MolModa, which was included as a representative Vina-class implementation. This comparison illustrates differences between platforms that occasionally yield accurate poses but exhibit variable stability and those that display more consistent behavior, highlighting practical trade-offs in docking reproducibility rather than establishing performance rankings. 3. Results 3.1. Active-Site Localization and Pose Accuracy Results are interpreted as indicative trends within this curated dataset and are not intended as universal performance rankings. 3.1.1. Accuracy of Binding Site Localization As shown in Fig. 2 A and Table 3 , differences in binding-site localization behavior were observed across the evaluated servers. MolModa localized all targets within the centroid-distance success threshold (10/10), indicating consistent binding-site confinement within this dataset. SwissDock exhibited similar localization behavior, with a small number of isolated deviations resulting in a 90% localization success rate. In contrast, GalaxyDock displayed greater dispersion in centroid distances, including occasional large displacements. PatchDock showed the broadest centroid-distance distribution and the lowest localization success rate. ParDOCK exhibited intermediate localization behavior but failed to generate a pose for one target (1Z95), which was treated as a localization failure. Complete per-run centroid distance values are provided in Supplementary Table S1 . Collectively, these observations illustrate server-dependent differences in binding-site localization behavior under identical input conditions, highlighting the utility of centroid-based analysis within the proposed evaluation framework 3.1.2. In-place RMSD Overall pose accuracy across servers As shown in Fig. 2 B and Table 3 , MolModa produced consistently low In-place RMSD values across targets. Notably, these values closely tracked centroid distances, indicating minimal additional orientation or conformational error once the binding site was localized. GalaxyDock exhibited increased variability in In-place RMSD, with several targets exceeding the 2.0 Å threshold despite successful pocket localization. SwissDock showed a comparable pattern, suggesting reduced consistency in final pose placement relative to localization performance. Rigid-body–dominated approaches showed larger deviations: PatchDock and ParDOCK exhibited substantially higher In-place RMSD values, reflecting limited capacity to recover native ligand orientations and internal conformations under the evaluated conditions. Complete per-run data are provided in Supplementary Table S2 . 3.1.3. Aligned Pose RMSD MolModa maintained low aligned RMSD values across targets, consistent with its In-place RMSD behavior. GalaxyDock also demonstrated low aligned RMSD values for many targets despite increased variability in pose placement, indicating partial decoupling between conformational similarity and spatial accuracy for this server. SwissDock showed reduced conformational similarity relative to its localization behavior, reflecting an inverse trend between pocket identification and internal geometry reproduction. PatchDock produced near-zero aligned RMSD values with negligible dispersion, a direct consequence of its rigid-body protocol that preserves the input ligand conformation without flexible sampling. In contrast, ParDOCK exhibited higher aligned RMSD values, indicating reduced consistency in reproducing bioactive ligand geometries (Fig. 2 C; Table 3 ). Complete per-run aligned RMSD values are provided in Supplementary Table S3 . Table 3 Evaluation of metrics for binding site localization and pose accuracy across five docking servers. Values represent the median with the interquartile range (IQR) in parentheses. Success rate denotes the percentage of targets where the docked ligand centroid was within 2.0 Å of the crystallographic reference. Server Centroid distance Median (Å) (IQR) In-place rmsd (Å) Median (IQR) Aligned Pose rmsd (Å) Median (IQR) Success rate (%) MolModa 0.1925 (0.19) 0.4029 (0.23) 0.2304 (0.11) 100 GalaxyDock 0.366 (1.15) 0.7876 (1.94) 0.3649 (0.78) 80 SwissDock 0.292 (0.35) 0.6912 (1.59) 0.6258 (0.81) 90 PatchDock 1.225 (2.77) 3.4175 (5.07) 0.0005 (0.000075) 60 ParDock 0.899 (0.77) 3.7944 (2.07) 1.7407 (1.07) 70 3.1.4. Protein-class–dependent trends in binding-site localization and pose accuracy When docking behavior was stratified by protein class, systematic differences in binding-site localization and fixed-frame pose placement were observed across enzymes, ATPases, and nuclear receptors. For enzyme targets (1IA1, 1S3V, 1T40), centroid distances and In-place RMSD values were generally low for MolModa and SwissDock, with relatively narrow dispersion compared to other protein classes. In contrast, GalaxyDock and PatchDock exhibited increased variability for individual enzyme complexes, reflected by elevated centroid displacement and In-place RMSD values for specific targets, while ParDOCK showed intermediate deviations across the same set. For ATPase targets (1OPK, 1T46, 1YWR), broader dispersion in both centroid distance and In-place RMSD was observed across several servers. GalaxyDock and PatchDock displayed higher variability in fixed-frame pose placement for multiple ATPase complexes, whereas MolModa and SwissDock showed comparatively lower centroid displacement with reduced spread across targets. ParDOCK exhibited elevated In-place RMSD values for multiple ATPase structures. For nuclear receptor targets (1S19, 1SJ0, 1SQN, 1Z95), centroid distances were generally low for MolModa, GalaxyDock, and SwissDock, indicating consistent binding-site localization across these servers. However, In-place RMSD values showed wider distributions for GalaxyDock, PatchDock, and ParDOCK, reflecting increased variability in pose placement despite preserved localization to the binding site. Overall, stratification by protein class illustrates that docking behavior varies with target family and server type, reinforcing the utility of class-aware analysis within the proposed evaluation framework. 3.2. Algorithmic Stability and Pose Reproducibility 3.2.1. Inter-run pose consistency MolModa, PatchDock, and ParDOCK returned identical top-ranked poses across repeated executions for all evaluated systems, indicating deterministic pose output under identical input conditions. In contrast, GalaxyDock and SwissDock exhibited measurable inter-run variability, with several targets showing substantial differences between replicate docking runs, including cases with large deviations in fixed-frame pose placement (Fig. 3 ; Table 4 ). Table 4 Summary of inter-run pose reproducibility across docking servers. Median and interquartile range (IQR) of pairwise inter-run in-place RMSD values (Å) calculated from triplicate docking runs. Values close to zero indicate high pose reproducibility across independent runs. Server Median (IQR) MolModa 0 (0) GalaxyDock 0.4568 (0.61) SwissDock 0.4342 (1.05) PatchDock 0 (0) ParDock 0 (0) 3.2.2. Robustness of correct pose prediction MolModa exhibited perfect robustness, achieving consistent native pose recovery for all evaluated targets. In contrast, the stochastic docking servers GalaxyDock and SwissDock displayed reduced robustness, with consistent recovery observed for only a subset of targets and complete failure for others. PatchDock showed moderate robustness, recovering the native pose consistently for approximately half of the dataset, whereas ParDOCK demonstrated the poorest performance, with consistent native pose recovery observed for only a single target. Detailed categorical outcomes are summarized in Table 5 . Table 5 Consistency of native pose recovery across repeated docking runs. Number of protein–server pairs achieving successful native pose recovery (In-Place RMSD ≤ 2.0 Å) in all three runs (3/3), two runs (2/3), one run (1/3), or none of the three runs (0/3) across triplicate docking experiments. Server 3/3 Success 2/3 Success 1/3 Success 0/3 Success MolModa 10 0 0 0 GalaxyDock 7 0 0 3 SwissDock 7 0 2 1 PatchDock 5 0 0 5 ParDock 1 0 0 9 3.2.3. Relative Robustness Against a Baseline Server MolModa, which exhibited deterministic inter-run behavior and consistent native pose recovery across all evaluated targets, was used as a representative Vina-class reference implementation for contextualizing reproducibility behavior. Relative to this reference, GalaxyDock and SwissDock displayed reduced robustness, characterized by non-zero inter-run deviations and consistent native pose recovery for only a subset of targets. PatchDock showed near-deterministic inter-run behavior similar to the reference, but achieved consistent native pose recovery for only a portion of the dataset. ParDOCK exhibited minimal inter-run RMSD dispersion but failed to reproducibly recover the native pose for the majority of targets. Collectively, these reference-contextualized results illustrate server-dependent differences in operational stability and provide a reproducibility-focused perspective that is independent of pose accuracy. 4. Discussion The objective of this work is to illustrate the utility of reproducibility-aware evaluation metrics rather than to derive definitive global performance hierarchies across protein classes. From a biological perspective, unreliable binding-site localization or unstable pose recovery can lead to incorrect interpretation of protein–ligand interactions, residue contributions, and structure–function relationships. Although Cross et al. applied correlation and paired statistical tests to compare mean RMSDs across docking programs, those analyses were designed to assess inter-method agreement across distinct protein–ligand complexes. In the present study, RMSD metrics arise from repeated stochastic executions on the same complexes and therefore do not constitute independent samples suitable for inferential hypothesis testing. In this context, the term stochastic refers to docking platforms whose search and/or ranking procedures explicitly incorporate random sampling steps, resulting in non-identical pose rankings across independent executions, rather than simply to randomized initial coordinate placement. 4.1. Decoupling Pose Accuracy from Reproducibility Reveals Distinct Failure Modes The primary contribution of this study is the demonstration that pose accuracy and run-to-run reproducibility represent complementary and non-interchangeable dimensions of docking performance. Conventional single-run evaluations implicitly assume that a server producing an accurate pose once will do so reliably. The reproducibility-aware framework presented here shows that this assumption does not necessarily hold for web-based docking platforms. When localization, fixed-frame pose placement, and inter-run consistency metrics are considered together, multiple failure modes become apparent. Some docking platforms generate geometrically accurate ligand conformations, as reflected by low aligned RMSD values, yet exhibit substantial variability in pose ranking across independent executions. This behavior indicates that accurate conformations can be sampled but are not consistently selected as the top-ranked solution. Conversely, other platforms produce identical outputs across repeated runs but nevertheless fail to recover native-like poses for a subset of complexes, demonstrating that deterministic behavior alone does not guarantee correctness. These observations highlight that pose accuracy and reproducibility interrogate fundamentally different aspects of docking algorithms. Pose accuracy reflects the capacity of a search strategy and scoring function to identify a near-native pose within the energy landscape, whereas reproducibility reflects the stability of the ranking process with respect to stochastic sampling. A docking platform may therefore perform well by one criterion while performing poorly by the other. By explicitly separating these dimensions, the present framework exposes algorithm-dependent trade-offs between sampling breadth and ranking stability. Platforms incorporating extensive stochastic exploration may intermittently access correct poses but display variable convergence behavior, while more deterministic or geometry-driven approaches tend to yield stable outputs that may lack sufficient adaptability for diverse ligand–receptor geometries. Importantly, these findings emphasize that pose accuracy alone is insufficient for qualifying web-based docking results under black-box usage conditions. Stability-aware evaluation is required to determine whether an observed pose represents a reproducible prediction or a single-run artifact. The framework introduced here provides a practical means to make this distinction. 4.2. Limitations of Rigid and Geometric Docking Approaches The behavior of PatchDock illustrates methodological constraints associated with rigid-body, geometry-driven docking strategies under re-docking conditions. PatchDock consistently produced near-zero aligned heavy-atom RMSD values across the dataset. This behavior reflects the rigid-body nature of the algorithm, which preserves the internal geometry of the input ligand and does not perform conformational sampling. Because the docking ligand was derived directly from the crystallographic pose and displaced only by rigid-body translation, close geometric agreement following optimal superposition is expected and does not, by itself, indicate successful binding-site localization. Accordingly, centroid distance and In-place RMSD metrics were essential for distinguishing cases of true pocket localization from geometrically aligned poses located outside the binding site. In the absence of internal energy minimization or induced-fit modeling, rigid geometric approaches may be limited in their ability to accommodate subtle steric conflicts or conformational adjustments that contribute to native-like pose placement, even under re-docking conditions. ParDOCK exhibited a distinct failure mode characterized by limited reproducible native pose recovery across the evaluated systems. Although both PatchDock and ParDOCK represent ligands using rigid geometries, their underlying search strategies differ substantially. PatchDock relies on purely geometric matching, whereas ParDOCK applies energy-based Monte Carlo sampling with explicit ligand reorientation. Despite employing an all-atom, energy-based protocol, ParDOCK showed limited consistency in native pose recovery under the tested conditions. This discrepancy suggests sensitivity to factors such as sampling efficiency, force-field representation, or scoring discrimination rather than to ligand rigidity alone. An additional practical limitation was encountered for ParDOCK during docking of the nuclear receptor target PDB ID: 1Z95 [ 19 ]. Despite repeated submissions using standardized ligand preparation and geometry optimization protocols, ParDOCK consistently rejected the ligand input for this system, preventing pose generation. Consequently, this target was excluded from ParDOCK-specific analyses. While the precise cause of this behavior could not be determined from the web interface, it highlights a general constraint of web-based docking platforms, where input handling and error reporting are not always transparent to the user. 4.3. Implications for “Black-Box” Users of Web-Based Docking Servers A central practical implication of this study concerns how results from web-based docking platforms should be interpreted under routine “black-box” usage conditions. The increasing accessibility of such tools has contributed to an implicit assumption that single-submission docking outputs are deterministic and directly actionable. The reproducibilty-aware evaluation framework presented here demonstrates that this assumption is not universally valid. For docking platforms employing stochastic search or ranking procedures, single-run outputs cannot be assumed to represent stable predictions. Although such platforms may sample native-like binding modes, run-to-run variability indicates that individual submissions may yield alternative high-ranking poses. Within this context, absence of convergence across repeated runs should not be interpreted as lack of algorithmic capability, but rather as a manifestation of stochastic sampling behavior. Replicate docking therefore becomes a methodological requirement rather than an optional refinement step. By contrast, platforms exhibiting deterministic or near-deterministic behavior provide consistent pose outputs under identical input conditions. Such behavior supports their use in rapid screening or hypothesis-generation workflows, particularly when computational resources, queue limits, or user expertise restrict the feasibility of repeated submissions. However, deterministic behavior alone does not guarantee correct binding-site localization or accurate pose placement, underscoring the need for complementary geometric validation metrics. Rigid and geometry-driven docking strategies introduce an additional interpretive layer. Deterministic geometric matching can produce reproducible outputs, yet geometric similarity under optimal superposition does not necessarily imply successful localization within the intended binding pocket. Consequently, reliance on aligned RMSD alone is insufficient for evaluating docking success when using rigid or geometry-driven approaches, and localization-aware metrics are essential. Based on these considerations, the present framework supports a tiered interpretation strategy for web-based docking results. Deterministic platforms are well suited for rapid exploratory analyses, whereas stochastic platforms require replicate sampling and convergence assessment to establish confidence in predicted poses. Rigid or geometry-driven platforms may be useful for coarse shape-based exploration but should be complemented with localization and fixed-frame pose metrics when used in re-docking scenarios. Collectively, these observations emphasize that effective use of web-based docking servers depends not only on server selection but also on adoption of stability-aware evaluation practices. The framework introduced here provides a practical basis for such interpretation without assuming universal performance hierarchies across docking algorithms 4.4. Limitations and Future Directions The primary scope of this study is the introduction and demonstration of a reproducibility-aware geometric evaluation framework rather than exhaustive characterization of docking performance across large chemical or target spaces. Accordingly, the selected set of protein–ligand complexes was chosen to provide controlled test cases for illustrating distinct failure modes in localization, pose placement, and reproducibility, rather than to support statistical generalization across protein families. While the present work demonstrates the utility of decoupling localization, fixed-frame pose accuracy, and inter-run stability, additional validation on larger and more diverse datasets will be necessary to further refine recommended threshold values and to explore how these metrics behave across broader classes of targets and ligand chemotypes. Extension of the framework to systems exhibiting substantial backbone rearrangements, shallow or highly solvent-exposed binding sites, and allosteric pockets represents a particularly important direction. The framework currently focuses on pose geometry and reproducibility and does not incorporate binding affinity prediction or enrichment-based virtual screening metrics. Integration of reproducibility and stability-aware geometric evaluation with scoring-function benchmarking represents a natural future extension. Finally, computational efficiency and wall-clock runtime were not analyzed in this study. For web-based platforms, reported runtimes are strongly influenced by server load, queue policies, and backend hardware in addition to algorithmic complexity. As such, runtime benchmarking is better addressed through controlled local implementations or standardized cloud environments. 5. Conclusions This work introduces a reproducibilty-aware geometric evaluation framework for assessing web-based molecular docking outputs under typical black-box usage conditions. By explicitly decoupling binding-site localization, fixed-frame pose accuracy, and run-to-run reproducibility, the framework demonstrates that conventional single-run docking assessments can mask distinct and practically important failure modes. Application of this framework illustrates that accurate pose generation, correct binding-site localization, and reproducible ranking behavior are separable algorithmic properties. Consequently, pose accuracy alone is insufficient to characterize the practical reliability of docking predictions. Incorporation of reproducibility and localization-aware metrics provides essential contextual information for interpreting docking results generated by heterogeneous web-based platforms. Rather than establishing universal performance hierarchies, this study provides a generalizable methodology for diagnosing docking behavior and for identifying when replicate sampling and complementary geometric metrics are necessary. Adoption of such reproducibility-aware evaluation practices can improve confidence in docking-derived hypotheses and support more reliable use of web-based docking tools in structure-based drug design. Abbreviations RMSD, Root Mean Square Deviation Declarations Conflicts of interest statement Venkata ramana singamaneni is an employee of Cambrex, Charles city, Iowa, USA. This work was conducted outside the scope of his employment, and Cambrex had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The authors declare no other competing interests. Data Availability All input protein structures, reference ligands, docked pose files, per-run result files and per-run RMSD tables required to reproduce the analyses are provided in data repository, Analysis script used for centroid distance, in-place RMSD, and aligned RMSD calculations along with usage instructions is also included as plain-text files. All data are available without restriction in the Mendeley repository- https://data.mendeley.com/datasets/sv7jcdc4xj/1 Funding- The authors did not receive support from any organization for the submitted work. CRediT authorship contribution statement Maniratnam Puli – Supervision (lead); Conceptualisation (lead); Data curation (lead); writing – original draft (lead); formal analysis (lead); software (lead); visualisation (lead). Venkata Ramana Singamaneni - software (supporting), formal analysis (supporting). Nikitha Bennuri – Data curation (equal); formal analysis (equal); software (equal). Sravya Peethani - Data curation (equal); formal analysis (equal); software (equal). Hawanika Durgam - Data curation (equal); formal analysis (equal); software (equal). Sowjanya Varipelli - Data curation (equal); formal analysis (equal); software (equal). Acknowledgment- None References Kitchen DB, Decornez H, Furr JR, Bajorath J (2004) Docking and scoring in virtual screening for drug discovery: methods and applications. nat rev drug discov 3:935 – 49. https://doi.org/10.1038/nrd1549 Pinzi L, Rastelli G (2019) Molecular docking: shifting paradigms in drug discovery. Int j of mol sci 20:4331. Kellenberger E, Rodrigo J, Muller P, Rognan D (2004) Comparative evaluation of eight docking tools for docking and virtual screening accuracy. proteins: struct, funct, and Bioinform 57:225 – 42. Martis EA, Téletchéa S (2025) Ten quick tips to perform meaningful and reproducible molecular docking calculations. PLOS comput biol 21: e1013030. Cross JB, Thompson DC, Rai BK, Baber JC, Fan KY, Hu Y, Humblet C (2009) Comparison of several molecular docking programs: pose prediction and virtual screening accuracy. J chem inf model. 49:1455–74. Hartshorn MJ, Verdonk ML, Chessari G, Brewerton SC, Mooij WT, Mortenson PN, Murray CW (2007) Diverse, high-quality test set for the validation of protein– ligand docking performance. J med chem 50:726–41. Sahu MK, Nayak AK, Hailemeskel B, Eyupoglu OE (2024) Exploring recent updates on molecular docking: Types, method, application, limitation & future prospects. Int J Pharm Res Allied Sci 13:24–40. Feinstein WP, Brylinski M (2015) Calculating an optimal box size for ligand docking and virtual screening against experimental and predicted binding pockets. J cheminf 7:18. Plewczynski D, Łaźniewski M, Augustyniak R, Ginalski K (2011) Can we trust docking results? Evaluation of seven commonly used programs on PDBbind database. J comput chem 32:742–55. Whitlow M, Howard AJ, Stewart D, Hardman KD, Chan JH, Baccanari DP, Tansik RL, Hong JS, Kuyper LF (2001) X-ray Crystal Structures of Candida albicans Dihydrofolate Reductase: High Resolution Ternary Complexes in Which the Dihydronicotinamide Moiety of NADPH Is Displaced by an Inhibitor. J med chem 44:2928–32. Cody V, Luft JR, Pangborn W, Gangjee A, Queener SF. Structure determination of tetrahydroquinazoline antifolates in complex with human and Pneumocystis carinii dihydrofolate reductase: correlations between enzyme selectivity and stereochemistry. Biol Crystallogr 60:646–55. Ruiz F, Hazemann I, Mitschler A, Joachimiak A, Schneider T, Karplus M, Podjarny A (2004) The crystallographic structure of the aldose reductase–IDD552 complex shows direct proton donation from tyrosine 48. Biol Crystallogr 60:1347–54. Nagar B, Hantschel O, Young MA, Scheffzek K, Veach D, Bornmann W, Clarkson B, Superti-Furga G, Kuriyan J (2003) Structural basis for the autoinhibition of c-Abl tyrosine kinase. Cell 112:859–71. Mol CD, Dougan DR, Schneider TR, Skene RJ, Kraus ML, Scheibe DN, Snell GP, Zou H, Sang BC, Wilson KP (2004) Structural basis for the autoinhibition and STI-571 inhibition of c-Kit tyrosine kinase. J Biol Chem 279:31655–63. Golebiowski A, Townes JA, Laufersweiler MJ, Brugel TA, Clark MP, Clark CM, Djung JF, Laughlin SK, Sabat MP, Bookland RG, VanRens JC, De B, Hsieh LC, Janusz MJ, Walter RL, Webster ME, Mekel MJ (2005) The development of monocyclic pyrazolone based cytokine synthesis inhibitors. Bioorg Med Chem Lett 15:2285–9. Tocchini-Valentini G, Rochel N, Wurtz JM, Moras D (2004) Crystal structures of the vitamin D nuclear receptor liganded with the vitamin D side chain analogues calcipotriol and seocalcitol, receptor agonists of clinical importance. Insights into a structural basis for the switching of calcipotriol to a receptor antagonist by further side chain modification. J Med Chem 47:1956–61. Kim S, Wu JY, Birzin ET, Frisch K, Chan W, Pai LY, Yang YT, Mosley RT, Fitzgerald PM, Sharma N, Dahllund J, Thorsell AG, DiNinno F, Rohrer SP, Schaeffer JM, Hammond ML (2004) Estrogen receptor ligands. II. Discovery of benzoxathiins as potent, selective estrogen receptor alpha modulators. J Med Chem 47:2171–5. Madauss KP, Deng SJ, Austin RJ, Lambert MH, McLay I, Pritchard J, Short SA, Stewart EL, Uings IJ, Williams SP (2004) Progesterone receptor ligand binding pocket flexibility: crystal structures of the norethindrone and mometasone furoate complexes. J Med Chem. 47:3381–7. Bohl CE, Gao W, Miller DD, Bell CE, Dalton JT (2005) Structural basis for antagonism and resistance of bicalutamide in prostate cancer. Proc Natl Acad Sci U S A 102:6201–6. Pettersen EF, Goddard TD, Huang CC, Meng EC, Couch GS, Croll TI, et al (2021) UCSF ChimeraX: Structure visualization for researchers, educators, and developers. Protein Sci 30:70–82. Kochnev Y, Ahmed M, Maldonado AM, Durrant JD (2024) MolModa: accessible and secure molecular docking in a web browser. Nucleic Acids Res 52: W498–W506. O'Boyle NM, Banck M, James CA, Morley C, Vandermeersch T, Hutchison GR (2011) Open Babel: An open chemical toolbox. J cheminform 3:33. Word JM, Lovell SC, Richardson JS, Richardson DC (1999) Asparagine and glutamine: using hydrogen atom contacts in the choice of side-chain amide orientation. J Mol Biol 285:1735–47. Eberhardt J, Santos-Martins D, Tillack AF, Forli S (2021) AutoDock Vina 1.2. 0: new docking methods, expanded force field, and python bindings. J chem Inform Model 61:3891–8. Bugnon M, Röhrig UF, Goullieux M, Perez MA, Daina A, Michielin O, Zoete V (2024) SwissDock 2024: major enhancements for small-molecule docking with Attracting Cavities and AutoDock Vina. Nucleic Acids Res 52: W324-32. Grosdidier A, Zoete V, Michielin O (2011) SwissDock, a protein-small molecule docking web service based on EADock DSS. Nucleic Acids Res 39: W270-7. Zoete V, Schuepbach T, Bovigny C, Chaskar P, Daina A, Röhrig UF, Michielin O (2016) Attracting cavities for docking. Replacing the rough energy landscape of the protein by a smooth attracting landscape. J Comput Chem 37:437–47. Rohrig UF, Goullieux M, Bugnon M, Zoete V (2023) Attracting cavities 2.0: improving the flexibility and robustness for small-molecule docking. J Chem Inform Model 63:3925–40. Gupta A, Gandhimathi A, Sharma P, Jayaram B (2007) ParDOCK: an all atom energy based Monte Carlo docking protocol for protein-ligand complexes. Protein and peptide letters. 14:632–46. Shin WH, Kim JK, Kim DS, Seok C (2013) GalaxyDock2: Protein–ligand docking using beta-complex and global optimization. J Comput Chem 34:2647–56. Baek M, Shin WH, Chung HW, Seok C (2017) GalaxyDock BP2 score: a hybrid scoring function for accurate protein–ligand docking. J Comput Aided Mol Des 31:653–66. Duhovny D, Nussinov R, Wolfson HJ (2002) Efficient unbound docking of rigid molecules. In International workshop on algorithms in bioinformatics Sep 17 (pp. 185–200). Berlin, Heidelberg: Springer Berlin Heidelberg. Schneidman-Duhovny D, Inbar Y, Nussinov R, Wolfson HJ (2005) PatchDock and SymmDock: servers for rigid and symmetric docking. Nucleic Acids Res 33(suppl_2): W363-7. Volkamer A, Griewel A, Grombacher T, Rarey M (2010) Analyzing the topology of active sites: on the prediction of pockets and subpockets. J Chem Inform Model 50:2041–52. Tripathi A, Suri K, K S, Murugan NA (2025) Assessing the accuracy of binding pose prediction for kinase proteins and 7-azaindole inhibitors: a study with AutoDock4, Vina, DOCK 6, and GNINA 1.0. RSC Adv 15:47051–47065. Additional Declarations Competing interest reported. Venkata ramana singamaneni is an employee of Cambrex, Charles city, Iowa, USA. This work was conducted outside the scope of his employment, and Cambrex had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The authors declare no other competing interests. Supplementary Files SupplementaryMaterial1.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 09 May, 2026 Reviewers agreed at journal 02 May, 2026 Reviewers agreed at journal 26 Apr, 2026 Reviewers invited by journal 14 Apr, 2026 Editor assigned by journal 23 Feb, 2026 Submission checks completed at journal 23 Feb, 2026 First submitted to journal 18 Feb, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8908837","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":623761700,"identity":"dd87c5e9-31d5-4ad8-90b6-eb3e2abaa2a2","order_by":0,"name":"Maniratnam Puli","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA9UlEQVRIiWNgGAWjYFAC5oYDQJIHyrMBYsbGA/i1MMK0MIN4aXARvFpg1oGIw2AmXi3mEomNh27U3JExbz9/7MHPnPN2a9sPA22psYnGpcVyRmLD4Zxjz3hkziSzG/Zuu5287UwiUMuxtNwGHFoMboC0sB3mkWBIZpPgBWoxOwDUwthwmICWf0At/I/ZJP9uO5dsdv4hEVpy24BaJJLZpHm3HbAzu0HAFsueh0DZPpCWx2bSstuSE8xuAG1JwOMXc/bkw59zvh22l+BPfCb5dpudvdn59IcPPtTY4HYYukAiWGUCDuVYtdjjUTwKRsEoGAUjFAAAguFmh2+Y/v0AAAAASUVORK5CYII=","orcid":"","institution":"G Pulla Reddy College of Pharmacy","correspondingAuthor":true,"prefix":"","firstName":"Maniratnam","middleName":"","lastName":"Puli","suffix":""},{"id":623761701,"identity":"c1599464-226c-49e6-acd7-8badf2a95617","order_by":1,"name":"Venkata Ramana Singamaneni","email":"","orcid":"","institution":"Cambrex","correspondingAuthor":false,"prefix":"","firstName":"Venkata","middleName":"Ramana","lastName":"Singamaneni","suffix":""},{"id":623761702,"identity":"9d37a847-1093-430b-9a62-f364f8cacc54","order_by":2,"name":"Nikitha Bennuri","email":"","orcid":"","institution":"G Pulla Reddy College of Pharmacy","correspondingAuthor":false,"prefix":"","firstName":"Nikitha","middleName":"","lastName":"Bennuri","suffix":""},{"id":623761703,"identity":"ee8a94ec-3df9-493d-8474-708ce34edef5","order_by":3,"name":"Sravya Peethani","email":"","orcid":"","institution":"G Pulla Reddy College of Pharmacy","correspondingAuthor":false,"prefix":"","firstName":"Sravya","middleName":"","lastName":"Peethani","suffix":""},{"id":623761704,"identity":"96a73eff-9af5-4334-857d-d6369c6535e3","order_by":4,"name":"Hawanika Durgam","email":"","orcid":"","institution":"G Pulla Reddy College of Pharmacy","correspondingAuthor":false,"prefix":"","firstName":"Hawanika","middleName":"","lastName":"Durgam","suffix":""},{"id":623761705,"identity":"f9e450d0-e99a-4bba-a49f-b2237c25cd1c","order_by":5,"name":"Sowjanya Varipelli","email":"","orcid":"","institution":"G Pulla Reddy College of Pharmacy","correspondingAuthor":false,"prefix":"","firstName":"Sowjanya","middleName":"","lastName":"Varipelli","suffix":""}],"badges":[],"createdAt":"2026-02-18 11:24:38","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8908837/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8908837/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107570367,"identity":"5cd8ca99-893d-422d-8945-c81aaeef233e","added_by":"auto","created_at":"2026-04-22 18:10:09","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":408760,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGeometric metrics used to decouple binding-site localization, pose placement, and conformational fidelity in unbiased re-docking. (A)\u003c/strong\u003eReference state showing the crystallographic ligand and its geometric center within the binding pocket. \u003cstrong\u003e(B)\u003c/strong\u003e Centroid distance, which quantifies binding-site localization error as the Euclidean distance between the centroids of the docked and crystallographic ligand poses. \u003cstrong\u003e(C)\u003c/strong\u003eIn-place RMSD, measuring orientation and placement accuracy within the fixed coordinate frame of the receptor without prior superposition. \u003cstrong\u003e(D)\u003c/strong\u003eAligned RMSD, computed after optimal rigid-body alignment of docked and reference ligands, reporting conformational fidelity independent of binding-site placement\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8908837/v1/f1d45e9b1194f4a6f5fc2339.png"},{"id":107570370,"identity":"af5c2860-9633-437a-ba63-b28958834513","added_by":"auto","created_at":"2026-04-22 18:10:10","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":38244,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eInter-server comparison of binding site localization and pose accuracy across five docking servers.\u003c/strong\u003e (A) Distribution of centroid distances (ΔCentroid) between docked ligands and their co-crystallized references, assessing binding site localization accuracy. (B) In-place RMSD distributions quantifying overall pose accuracy, incorporating translational, rotational, and conformational deviations relative to the crystallographic pose. (C) Aligned pose RMSD distributions evaluating ligand conformational accuracy independent of spatial positioning. For all box plots, the center line denotes the median, boxes represent the interquartile range (IQR), whiskers extend to 1.5×IQR, and individual points indicate outliers. The × symbol marks the mean\u003c/p\u003e","description":"","filename":"groupimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8908837/v1/edfa7b360afa9a8226a0b736.jpeg"},{"id":107570368,"identity":"019332e9-0e3f-4bc5-a8b5-8f354dddfb4d","added_by":"auto","created_at":"2026-04-22 18:10:09","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":43821,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eInter-run pose consistency of web docking servers.\u003c/strong\u003eBox-and-whisker plots show the distribution of pairwise inter-run \u003cem\u003eIn-place\u003c/em\u003eRMSD values (Å) between best-ranked poses obtained from three independent docking runs for each server. Lower values indicate higher reproducibility. MolModa, PatchDock, and ParDOCK exhibit near-deterministic behavior with negligible inter-run deviation, whereas GalaxyDock and SwissDock display greater variability, reflecting stochastic sampling effects. Boxes represent the interquartile range (IQR), horizontal lines denote the median, whiskers indicate the data range, and crosses mark the mean\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8908837/v1/359c8650cae7a5ad03326cae.png"},{"id":107706711,"identity":"2574cf72-a8e9-4c16-8ee1-6dd003746036","added_by":"auto","created_at":"2026-04-24 09:18:36","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":866838,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8908837/v1/0d3337b7-1866-4ffd-8a9d-4f6bbc43561d.pdf"},{"id":107570366,"identity":"c23a085e-f7a7-4572-9dcc-0ecdb3f8ec4d","added_by":"auto","created_at":"2026-04-22 18:10:09","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":41856,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial1.docx","url":"https://assets-eu.researchsquare.com/files/rs-8908837/v1/d919849b6981c12bf41654b7.docx"}],"financialInterests":"Competing interest reported. Venkata ramana singamaneni is an employee of Cambrex, Charles city, Iowa, USA. This work was conducted outside the scope of his employment, and Cambrex had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The authors declare no other competing interests.","formattedTitle":"Beyond Pose Accuracy: Reproducibility-Oriented Framework For Interpreting Web-Based Molecular Docking Outputs Using Decoupled Localization and Pose Fidelity Metrics","fulltext":[{"header":"1. Introduction","content":"\u003cdiv id=\"Sec2\" class=\"Section2\"\u003e \u003ch2\u003e1.1 Molecular Docking Benchmarks: From Algorithms to Web Servers\u003c/h2\u003e \u003cp\u003eMolecular docking is a cornerstone of structure-based drug design, enabling prediction of ligand binding modes and the virtual screening of large compound libraries [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Historically, rigorous benchmarking has focused on standalone docking engines such as DOCK, FlexX, GLIDE, and Surflex, where users retain granular control over sampling parameters, grid definitions, and scoring functions [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Advances in algorithmic efficiency and computational infrastructure have since driven a transition toward automated, web-based docking platforms that allow non-expert users to perform complex simulations without local software installations [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. While this democratization has broadened access, it has also introduced a \u0026ldquo;black-box\u0026rdquo; challenge, as users often lack visibility into underlying search parameters, stochastic elements such as random seeds, and server-specific failure modes. As a result, the focus of validation shifts from evaluating theoretical algorithmic capability to assessing the operational reliability and reproducibility of docking services, which remains a persistent challenge in computational drug discovery and necessitates transparent workflows and reproducible evaluation strategies [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e1.2 Lessons from Classical Benchmarks\u003c/h2\u003e \u003cp\u003eThe evaluation of docking performance therefore requires standardized and carefully controlled benchmarking protocols to ensure meaningful methodological comparison. In a seminal study, Cross et al. (2009) established a rigorous framework for assessing pose accuracy and virtual screening accuracy across six major docking programs [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Their benchmark emphasized that the primary criterion for docking success is the ability to reproduce the crystallographic ligand pose within a Root-Mean-Square Deviation (RMSD) of 2.0 \u0026Aring;. Importantly, Cross et al. demonstrated that absolute docking scores are rarely comparable across different programs and frequently fail to correlate linearly with RMSD, indicating that internal pose ranking consistency is often a more reliable indicator of performance than raw energy values. Their methodology\u0026mdash;leveraging high-resolution datasets such as the Astex diverse set and prioritizing geometric fidelity over binding-energy prediction\u0026mdash;continues to serve as a foundational reference for docking validation strategies [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e1.3 Gaps in Current Web Docking Evaluations\u003c/h2\u003e \u003cp\u003eWhen classical benchmarking concepts such as those proposed by Cross et al. are applied to modern web-based docking servers, several methodological gaps become apparent. Most existing evaluations implicitly treat web servers as deterministic tools, reporting results from a single docking run. This practice overlooks the stochastic nature of many contemporary search algorithms, including simulated annealing and genetic-algorithm\u0026ndash;based approaches, for which repeated submissions using identical inputs may yield divergent solutions due to random initialization or server-side execution variability [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. In addition, web-based docking platforms differ substantially in how binding-site information is specified and internally interpreted, ranging from Cartesian grid definitions to reference-ligand guidance, residue-based constraints, or geometry-driven matching. Such heterogeneity introduces user-invisible sources of variability that may influence docking outcomes even when nominally equivalent binding-site inputs are provided [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Collectively, these considerations suggest that modern validation efforts should extend beyond conformational accuracy alone and incorporate explicit assessment of binding-site localization and algorithmic reproducibility across repeated docking runs.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e1.4 Objectives of the Present Study\u003c/h2\u003e \u003cp\u003eIn this study, we evaluate five distinct web-based docking servers\u0026mdash;SwissDock, ParDOCK, GalaxyDock, PatchDock, and MolModa (representing a Vina-class reference implementation)\u0026mdash;using the high-resolution Astex diverse set. By adapting the rigorous pose-recovery protocols of Cross et al. (2009) to the practical constraints of web-based services, we demonstrate a reproducibility-aware evaluation framework rather than a definitive performance ranking, we pursue two primary objectives:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAim 1\u003c/b\u003e: To demonstrate a geometric evaluation strategy that decouples binding-site localization from pose fidelity by applying complementary distance-based measures (centroid distance, in-place RMSD, and aligned RMSD) to top-ranked docked structures relative to experimental co-crystal references.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAim 2\u003c/b\u003e: To demonstrate a reproducibility-aware evaluation strategy by applying triplicate docking and inter-run RMSD analysis to characterize run-to-run variability and illustrate how algorithmic stability influences the interpretation of docking results.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eReliable identification of ligand binding modes underpins mechanistic interpretation of protein\u0026ndash;ligand recognition, functional annotation, and hypothesis generation in structural biology. Therefore, assessing the biological reliability of widely used docking servers is critical for downstream interpretation. The objective of this work is to introduce and demonstrate a reproducibility-aware evaluation framework for docking reliability, rather than to provide a definitive ranking of docking servers.\u003c/p\u003e \u003c/div\u003e"},{"header":"2. Materials and Methods","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Protein\u0026ndash;Ligand Dataset\u003c/h2\u003e \u003cp\u003eTo provide a representative testbed for demonstrating the proposed evaluation framework, we utilized the Astex Diverse set [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], a curated collection of high-resolution protein\u0026ndash;ligand complexes spanning major protein classes, including enzymes, ATP-binding proteins, and nuclear receptors, designed to minimize structural redundancy while maximizing chemical diversity From the Astex diverse set, protein-ligand complexes were selected according to the following criteria:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eCrystallographic resolution\u0026thinsp;\u0026lt;\u0026thinsp;2.0\u0026Aring;, one nuclear receptor has 2.10 \u0026Aring;\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eAbsence of covalent protein-ligand interactions:\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eNo missing, unresolved residues and mutations within the binding site;\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eAbsence of metal-mediated covalent coordination to the ligand; and\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eAvailability of a single, well-defined bound ligand\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eInclusion of these varied architectures reduces bias toward specific binding pocket geometrics, such as deeply buried active sites versus shallow surface grooves. A total of ten protein-ligand complexes were selected for this evaluation to balance chemical and structural diversity with methodological control. Selected Complexes are listed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e10 X-ray Complexes Used for Cognate Ligand Docking\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eS.no\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eProtein class\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePDB ID\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLigand\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eResolution (\u0026Aring;)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eReference for protein PDB\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEnzyme\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1ia1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTQ3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEnzyme\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1s3v\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTQD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEnzyme\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1t40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eID5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eATPase\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1opk\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eP16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eATPase\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1t46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eST\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eATPase\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1ywr\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eL19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNuclear receptor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1s19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMC9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e8.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNuclear receptor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1sj0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eE4D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e9.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNuclear receptor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1sqn\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNDR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e10.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNuclear receptor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1z95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e198\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Protein\u0026ndash;Ligand preparation\u003c/h2\u003e \u003cp\u003ePrior to docking, all complexes were prepared to generate a consistent apo-like receptor state. The crystallographic ligand was extracted, protein structures were prepared using UCSF ChimeraX Dock Prep by solvent removal, residue standardization, hydrogen addition, and charge assignment while keeping all heavy atoms fixed at their crystallographic positions. No energy minimization of protein heavy atoms was performed [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] to ensure appropriate electrostatic treatment. The crystallographic ligand extracted from each complex was saved as an independent file without additional energy minimization and was retained as the reference structure for RMSD calculations. To preserve atom ordering and chemical topology, the same crystallographic ligand was spatially displaced and reused as the docking ligand. ChimeraX was selected for preparation due to its deterministic pipeline, ensuring transparent handling of hydrogens and residue states while strictly preserving crystallographic coordinates.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Docking Servers Evaluated\u003c/h2\u003e \u003cp\u003eDocking serves were selected according to the following inclusion criteria\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003ePublic accessibility\u003c/b\u003e, defined as free Web-based availability without the need for local software installation or proprietary licensing;\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eDomain applicability\u003c/b\u003e, requiring explicit support for small-molecule protein-ligand docking;\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eActive-site-guided sampling\u003c/b\u003e, Whereby the docking region can be constrained through user-specified Cartesian coordinates, binding-site residues, or reference-ligand guidance, thereby avoiding blind docking behaviour;\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAlgorithmic diversity\u003c/b\u003e, such that the selected servers employ distinct docking paradigms to minimize methodological redundancy;\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eCommunity adoption\u003c/b\u003e, Evidence by prior usage and citation in the computational chemistry literature; and\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003ePractical usability\u003c/b\u003e, requiring that individual docking jobs complete within a reasonable time frame compatible with high-throughput docking workflows.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eFive publicly accessible web-based docking platforms were evaluated, selected to represent a diverse range of docking algorithms and search paradigms.\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eMolModa (Baseline)\u003c/b\u003e: A browser-based docking platform implementing the AutoDock Vina scoring function and search algorithm. Prepared protein and ligand structures were uploaded to the MolModa interface, where protonation of both receptor and ligand was performed using the server\u0026rsquo;s internal preparation routine. Docking was carried out using the compound docking option with default search parameters. The docking grid was defined based on the crystallographic binding site and centered on the reference ligand. Grid dimensions were set to encompass the maximal extent of the crystallographic ligand with an additional 10.0 \u0026Aring; buffer (5.0 \u0026Aring; per axis). Docking calculations were performed on the protonated receptor\u0026ndash;ligand system, and the top-ranked pose returned by the server was retained for subsequent analysis [\u003cspan additionalcitationids=\"CR22 CR23 CR24\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eSwissDock\u003c/b\u003e: SwissDock was used as a web-based docking platform based on the EADock DSS algorithm. Prepared protein and ligand structures were submitted to the SwissDock server using the Attracting Cavities mode, which restricts docking to energetically favorable regions of the protein surface. The docking search space was defined using explicit Cartesian boundaries identical to those employed for MolModa. The grid was centered on the crystallographic reference ligand and extended to encompass the maximal ligand dimensions with an additional 10.0 \u0026Aring; buffer (5.0 \u0026Aring; per axis) ensuring sufficient sampling while avoiding artificial confinement to the crystal pose. [\u003cspan additionalcitationids=\"CR27\" citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eParDOCK\u003c/b\u003e: ParDOCK was used as a web-based protein\u0026ndash;ligand docking server employing an all-atom energy-based Monte Carlo sampling protocol performing rigid docking. The protein\u0026ndash;reference ligand complex, prepared in UCSF ChimeraX, was submitted to ParDOCK to implicitly define the binding site. In addition, the docking ligand, generated by displacing the crystallographic ligand in ChimeraX, was provided as a separate input for docking. Docking was performed with the reference ligand guiding localization of the Monte Carlo search to the correct binding cavity. [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eGalaxyDock\u003c/b\u003e: GalaxyDock was accessed via the GalaxyWEB server for protein\u0026ndash;ligand docking. Receptors were provided in PDB format and ligands in MOL2 format, with docking restricted to user-specified binding-site residues derived from the crystallographic complex. Docking was performed using the GalaxyDock BP2 protocol, which employs conformational space annealing (CSA\u003cb\u003e)\u003c/b\u003e for global pose exploration while allowing limited protein side-chain flexibility. Ligands were treated as fully flexible during docking. An energy-driven, grid-based scoring function was used to evaluate poses within the defined binding site. [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003ePatchDock\u003c/b\u003e: PatchDock is a geometry-based molecular docking webserver that performs rigid-body docking based on molecular surface shape complementarity and surface patch complementarity. Docking was performed by submitting the prepared protein and ligand structures in PDB format, along with a receptor binding-site file in TXT format to restrict docking transformations to the specified active-site region. It defines the docking region through surface-patch matching constrained to interacting residues identified from the experimental complex The top-ranked docked complexes are reported by the server for further analysis. [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eEach complex was subjected to triplicate docking across five independent web-based servers, followed by pose alignment and multi-metric RMSD evaluation. This design resulted in a total of 150 individual docking evaluations, each requiring verification of binding-site localization and coordinate-frame consistency. The dataset size was chosen to allow granular, case-by-case inspection of docking behavior and reproducibility, which is central to demonstrating the proposed evaluation framework. Substantially expanding the dataset would limit such manual oversight without necessarily improving interpretability in the context of illustrating methodological behavior across heterogeneous web-based platforms.\u003c/p\u003e \u003cp\u003e \u003cb\u003eMetric justification.\u003c/b\u003e Because binding-site specification strategies differ substantially among web-based docking servers, a binding-site localization metric (centroid distance) was incorporated to verify that these heterogeneous approaches resulted in comparable confinement of docked ligands within the intended binding pocket [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003cb\u003eServer-specific exclusions.\u003c/b\u003e For the 1Z95 system [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], ParDOCK consistently failed to progress beyond the ligand preparation stage despite correct molecular geometry, neutral formal charge, and repeated submission using standard preparation protocols. As this behavior was reproducible and appeared to reflect a server-side parsing limitation rather than a chemical or structural issue, this system was excluded from ParDOCK-specific analyses while being retained for all other servers.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.4. Docking Protocol\u003c/h2\u003e \u003cp\u003eAll docking calculations were performed using a standardized protocol designed to reflect typical \u0026ldquo;black-box\u0026rdquo; usage of web-based docking servers. For each server, docking was conducted using default parameters, and only the single top-ranked pose returned by each run was retained for analysis. No post-docking rescoring or energy minimization was applied, ensuring that pose evaluation relied solely on the server\u0026rsquo;s internal scoring and ranking procedures.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTriplicate strategy-\u003c/b\u003e To assess algorithmic stability and reproducibility, each protein\u0026ndash;ligand pair was docked in three independent, consecutive runs for every server using identical input structures. Repeated runs capture stochastic variability inherent to many docking algorithms (e.g., Monte Carlo sampling or simulated annealing) and enable quantification of run-to-run variance in pose accuracy. Restricting analysis to the top-ranked pose across all runs further mitigates ranking bias arising from differences in the number of poses generated by individual servers, ensuring that performance comparisons reflect algorithmic consistency rather than sampling depth.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Pose Evaluation Metrics (Aim1)\u003c/h2\u003e \u003cp\u003eDocking accuracy was evaluated using three complementary metrics capturing binding-site localization and pose accuracy: centroid distance, In-place RMSD, and aligned heavy-atom RMSD. Together, these metrics decouple errors in spatial localization from conformational inaccuracies.\u003c/p\u003e \u003cp\u003eDocked complexes returned by ParDOCK were provided in a global coordinate frame differing from the reference structures. To enable meaningful RMSD comparisons, ParDOCK complexes were rigidly aligned to their corresponding reference proteins using the MatchMaker tool in UCSF ChimeraX [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], based on protein backbone atoms only; ligand atoms were excluded from the alignment. No alignment was required for other servers, which preserved the original receptor coordinate frame.\u003c/p\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e2.5.1 Centroid Distance (Binding-Site Localization)\u003c/h2\u003e \u003cp\u003eCentroid distance measures the spatial displacement of the docked ligand relative to the crystallographic binding site without any structural superposition. It was calculated as the Euclidean distance between the geometric centroids of the docked and reference ligand poses, defined as the arithmetic mean of all heavy-atom Cartesian coordinates. Because no alignment is performed, this metric reflects translational displacement within the binding pocket and is independent of ligand orientation or internal conformation. Centroid distances\u0026thinsp;\u0026le;\u0026thinsp;\u003cb\u003e2.0 \u0026Aring;\u003c/b\u003e were considered indicative of correct binding-site localization, with \u0026le;\u0026thinsp;\u003cb\u003e1.0 \u0026Aring;\u003c/b\u003e classified as excellent localization consistent with the center-of-mass tolerance thresholds applied in recent kinase benchmarking studies Tripathi et al., 2025 [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e2.5.2 In-Place RMSD (Conformational Accuracy)\u003c/h2\u003e \u003cp\u003eIn-place RMSD quantifies conformational accuracy within a fixed binding-site frame. RMSD was computed directly between corresponding heavy atoms of the docked and reference ligands without ligand superposition, using the receptor-aligned coordinate frame. This metric isolates errors arising from incorrect ligand orientation or conformation after correct site localization has been achieved.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e2.5.3 Aligned Heavy-Atom RMSD (Overall Pose Accuracy)\u003c/h2\u003e \u003cp\u003eAligned heavy-atom RMSD assesses overall pose similarity following optimal rigid-body superposition of the docked ligand onto the crystallographic reference. RMSD was calculated over corresponding heavy atoms after least-squares alignment, consistent with the classical definition of docking success used in established benchmarks such as Cross \u003cem\u003eet al.\u003c/em\u003e A threshold of \u003cb\u003e2.0 \u0026Aring;\u003c/b\u003e was used as the primary success criterion. While the same numerical cutoff was applied to In-place RMSD, the two metrics capture distinct aspects of docking performance: aligned RMSD reflects optimal pose recovery, whereas In-place RMSD provides a stricter evaluation within a fixed spatial frame. A graphical representation of the three metrics has been given in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The evaluation criteria for the metrics are given in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAcceptance threshold\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026le;\u0026thinsp;1.0 \u0026Aring;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026le;\u0026thinsp;2.0 \u0026Aring;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;2.0 \u0026Aring;\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCentroid distance\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExcellent localization\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCorrect localization\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWrong site\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIn-place RMSD\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNear-native conformation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAcceptable conformation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDistorted pose\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAligned RMSD\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHigh-accuracy pose\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCorrect pose\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eIncorrect pose\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e2.5.4 RMSD script\u003c/h2\u003e \u003cp\u003eRMSD-based metrics were calculated using custom Python script implementing RDkit which performs three measures: (i) centroid distance between geometric centers of docked and crystallographic ligands; (ii) in-place RMSD computed in the fixed receptor coordinate frame; and (iii) aligned RMSD obtained after optimal heavy-atom superposition of ligand coordinates, with RMSD\u0026thinsp;\u0026lt;\u0026thinsp;2.0 \u0026Aring; considered as threshold. Script is symmetry-aware, incorporate atom-mapping procedures to ensure consistency across differing atom orders, and robust to ligand rotations and conformational variability. The script and the usage instructions have been provided in data repository accessible at - \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://data.mendeley.com/datasets/sv7jcdc4xj/1\u003c/span\u003e\u003cspan address=\"https://data.mendeley.com/datasets/sv7jcdc4xj/1\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Reproducibility and Robustness Metrics (Aim 2)\u003c/h2\u003e \u003cdiv id=\"Sec17\" class=\"Section3\"\u003e \u003ch2\u003e2.6.1. Pose Reproducibility and algorithmic stability (Primary Robustness Metric)\u003c/h2\u003e \u003cp\u003ePose reproducibility quantifies the consistency of a docking server across independent executions. For each protein\u0026ndash;ligand pair, docking was performed in three independent runs using identical inputs and default parameters, retaining only the top-ranked pose from each run. Run-to-run consistency was assessed by computing pairwise heavy-atom RMSDs between top-ranked poses from all run pairs using In-place RMSD (no superposition) between triplicate docking runs (Run1\u0026ndash;Run2, Run1\u0026ndash;Run3, Run2\u0026ndash;Run3). For each protein\u0026ndash;ligand system, the median inter-run RMSD was calculated and aggregated across proteins to obtain server-level reproducibility statistics. Low median inter-run RMSD indicates a deterministically stable docking process, whereas high variance indicates stochastic sampling behavior or convergence failure\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section3\"\u003e \u003ch2\u003e2.6.2. Robustness of correct pose prediction\u003c/h2\u003e \u003cp\u003e \u003cb\u003eRobustness of correct pose prediction\u003c/b\u003e evaluates whether a docking server can reproducibly generate a correct ligand binding pose within the crystallographic binding site across repeated docking runs. Correct predictions are defined using In-place heavy-atom RMSD\u0026thinsp;\u0026le;\u0026thinsp;2.0 \u0026Aring; relative to the experimental ligand structure, ensuring simultaneous correctness of binding-site localization, orientation, and conformation. For each protein\u0026ndash;server pair, triplicate outcomes were collapsed into protein-level categories reflecting consistent (3/3), partial (2/3), sporadic (1/3), or absent (0/3) native pose recovery. Results were summarized as counts per server, enabling evaluation of reproducibility of correct docking outcomes rather than average error magnitude.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section3\"\u003e \u003ch2\u003e2.6.3. Relative Robustness Against a Baseline Server\u003c/h2\u003e \u003cp\u003eTo provide a reference context for reproducibility behavior, robustness metrics were interpreted relative to MolModa, which was included as a representative Vina-class implementation. This comparison illustrates differences between platforms that occasionally yield accurate poses but exhibit variable stability and those that display more consistent behavior, highlighting practical trade-offs in docking reproducibility rather than establishing performance rankings.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Active-Site Localization and Pose Accuracy\u003c/h2\u003e \u003cp\u003eResults are interpreted as indicative trends within this curated dataset and are not intended as universal performance rankings.\u003c/p\u003e \u003cdiv id=\"Sec22\" class=\"Section3\"\u003e \u003ch2\u003e3.1.1. Accuracy of Binding Site Localization\u003c/h2\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA and Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, differences in binding-site localization behavior were observed across the evaluated servers. MolModa localized all targets within the centroid-distance success threshold (10/10), indicating consistent binding-site confinement within this dataset. SwissDock exhibited similar localization behavior, with a small number of isolated deviations resulting in a 90% localization success rate. In contrast, GalaxyDock displayed greater dispersion in centroid distances, including occasional large displacements. PatchDock showed the broadest centroid-distance distribution and the lowest localization success rate. ParDOCK exhibited intermediate localization behavior but failed to generate a pose for one target (1Z95), which was treated as a localization failure. Complete per-run centroid distance values are provided in Supplementary \u003cb\u003eTable S1\u003c/b\u003e.\u003c/p\u003e \u003cp\u003eCollectively, these observations illustrate server-dependent differences in binding-site localization behavior under identical input conditions, highlighting the utility of centroid-based analysis within the proposed evaluation framework\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003e3.1.2. In-place RMSD Overall pose accuracy across servers\u003c/h2\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB and Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, MolModa produced consistently low In-place RMSD values across targets. Notably, these values closely tracked centroid distances, indicating minimal additional orientation or conformational error once the binding site was localized. GalaxyDock exhibited increased variability in In-place RMSD, with several targets exceeding the 2.0 \u0026Aring; threshold despite successful pocket localization. SwissDock showed a comparable pattern, suggesting reduced consistency in final pose placement relative to localization performance. Rigid-body\u0026ndash;dominated approaches showed larger deviations: PatchDock and ParDOCK exhibited substantially higher In-place RMSD values, reflecting limited capacity to recover native ligand orientations and internal conformations under the evaluated conditions. Complete per-run data are provided in Supplementary \u003cb\u003eTable S2\u003c/b\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section3\"\u003e \u003ch2\u003e3.1.3. Aligned Pose RMSD\u003c/h2\u003e \u003cp\u003eMolModa maintained low aligned RMSD values across targets, consistent with its In-place RMSD behavior. GalaxyDock also demonstrated low aligned RMSD values for many targets despite increased variability in pose placement, indicating partial decoupling between conformational similarity and spatial accuracy for this server. SwissDock showed reduced conformational similarity relative to its localization behavior, reflecting an inverse trend between pocket identification and internal geometry reproduction. PatchDock produced near-zero aligned RMSD values with negligible dispersion, a direct consequence of its rigid-body protocol that preserves the input ligand conformation without flexible sampling. In contrast, ParDOCK exhibited higher aligned RMSD values, indicating reduced consistency in reproducing bioactive ligand geometries (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC; Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Complete per-run aligned RMSD values are provided in Supplementary \u003cb\u003eTable S3\u003c/b\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eEvaluation of metrics for binding site localization and pose accuracy across five docking servers.\u003c/b\u003e Values represent the median with the interquartile range (IQR) in parentheses. Success rate denotes the percentage of targets where the docked ligand centroid was within 2.0 \u0026Aring; of the crystallographic reference.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eServer\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCentroid distance Median (\u0026Aring;) (IQR)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIn-place rmsd (\u0026Aring;) Median (IQR)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAligned Pose rmsd (\u0026Aring;) Median (IQR)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSuccess rate (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMolModa\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.1925 (0.19)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.4029 (0.23)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.2304 (0.11)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eGalaxyDock\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.366 (1.15)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7876 (1.94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.3649 (0.78)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSwissDock\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.292 (0.35)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.6912 (1.59)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.6258 (0.81)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePatchDock\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.225 (2.77)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.4175 (5.07)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.0005 (0.000075)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e60\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eParDock\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.899 (0.77)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.7944 (2.07)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.7407 (1.07)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e70\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e \u003ch2\u003e3.1.4. Protein-class\u0026ndash;dependent trends in binding-site localization and pose accuracy\u003c/h2\u003e \u003cp\u003eWhen docking behavior was stratified by protein class, systematic differences in binding-site localization and fixed-frame pose placement were observed across enzymes, ATPases, and nuclear receptors.\u003c/p\u003e \u003cp\u003eFor enzyme targets (1IA1, 1S3V, 1T40), centroid distances and In-place RMSD values were generally low for MolModa and SwissDock, with relatively narrow dispersion compared to other protein classes. In contrast, GalaxyDock and PatchDock exhibited increased variability for individual enzyme complexes, reflected by elevated centroid displacement and In-place RMSD values for specific targets, while ParDOCK showed intermediate deviations across the same set.\u003c/p\u003e \u003cp\u003eFor ATPase targets (1OPK, 1T46, 1YWR), broader dispersion in both centroid distance and In-place RMSD was observed across several servers. GalaxyDock and PatchDock displayed higher variability in fixed-frame pose placement for multiple ATPase complexes, whereas MolModa and SwissDock showed comparatively lower centroid displacement with reduced spread across targets. ParDOCK exhibited elevated In-place RMSD values for multiple ATPase structures.\u003c/p\u003e \u003cp\u003eFor nuclear receptor targets (1S19, 1SJ0, 1SQN, 1Z95), centroid distances were generally low for MolModa, GalaxyDock, and SwissDock, indicating consistent binding-site localization across these servers. However, In-place RMSD values showed wider distributions for GalaxyDock, PatchDock, and ParDOCK, reflecting increased variability in pose placement despite preserved localization to the binding site.\u003c/p\u003e \u003cp\u003eOverall, stratification by protein class illustrates that docking behavior varies with target family and server type, reinforcing the utility of class-aware analysis within the proposed evaluation framework.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Algorithmic Stability and Pose Reproducibility\u003c/h2\u003e \u003cdiv id=\"Sec27\" class=\"Section3\"\u003e \u003ch2\u003e3.2.1. Inter-run pose consistency\u003c/h2\u003e \u003cp\u003eMolModa, PatchDock, and ParDOCK returned identical top-ranked poses across repeated executions for all evaluated systems, indicating deterministic pose output under identical input conditions. In contrast, GalaxyDock and SwissDock exhibited measurable inter-run variability, with several targets showing substantial differences between replicate docking runs, including cases with large deviations in fixed-frame pose placement (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e; Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eSummary of inter-run pose reproducibility across docking servers.\u003c/b\u003e Median and interquartile range (IQR) of pairwise inter-run \u003cem\u003ein-place\u003c/em\u003e RMSD values (\u0026Aring;) calculated from triplicate docking runs. Values close to zero indicate high pose reproducibility across independent runs.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eServer\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMedian (IQR)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMolModa\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eGalaxyDock\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.4568 (0.61)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSwissDock\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.4342 (1.05)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePatchDock\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eParDock\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section3\"\u003e \u003ch2\u003e3.2.2. Robustness of correct pose prediction\u003c/h2\u003e \u003cp\u003eMolModa exhibited perfect robustness, achieving consistent native pose recovery for all evaluated targets. In contrast, the stochastic docking servers GalaxyDock and SwissDock displayed reduced robustness, with consistent recovery observed for only a subset of targets and complete failure for others. PatchDock showed moderate robustness, recovering the native pose consistently for approximately half of the dataset, whereas ParDOCK demonstrated the poorest performance, with consistent native pose recovery observed for only a single target. Detailed categorical outcomes are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eConsistency of native pose recovery across repeated docking runs.\u003c/b\u003e Number of protein\u0026ndash;server pairs achieving successful native pose recovery (In-Place RMSD\u0026thinsp;\u0026le;\u0026thinsp;2.0 \u0026Aring;) in all three runs (3/3), two runs (2/3), one run (1/3), or none of the three runs (0/3) across triplicate docking experiments.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eServer\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3/3 Success\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2/3 Success\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1/3 Success\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0/3 Success\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMolModa\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eGalaxyDock\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSwissDock\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePatchDock\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eParDock\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec29\" class=\"Section3\"\u003e \u003ch2\u003e3.2.3. Relative Robustness Against a Baseline Server\u003c/h2\u003e \u003cp\u003eMolModa, which exhibited deterministic inter-run behavior and consistent native pose recovery across all evaluated targets, was used as a representative Vina-class reference implementation for contextualizing reproducibility behavior. Relative to this reference, GalaxyDock and SwissDock displayed reduced robustness, characterized by non-zero inter-run deviations and consistent native pose recovery for only a subset of targets. PatchDock showed near-deterministic inter-run behavior similar to the reference, but achieved consistent native pose recovery for only a portion of the dataset. ParDOCK exhibited minimal inter-run RMSD dispersion but failed to reproducibly recover the native pose for the majority of targets.\u003c/p\u003e \u003cp\u003eCollectively, these reference-contextualized results illustrate server-dependent differences in operational stability and provide a reproducibility-focused perspective that is independent of pose accuracy.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThe objective of this work is to illustrate the utility of reproducibility-aware evaluation metrics rather than to derive definitive global performance hierarchies across protein classes. From a biological perspective, unreliable binding-site localization or unstable pose recovery can lead to incorrect interpretation of protein\u0026ndash;ligand interactions, residue contributions, and structure\u0026ndash;function relationships.\u003c/p\u003e \u003cp\u003eAlthough Cross et al. applied correlation and paired statistical tests to compare mean RMSDs across docking programs, those analyses were designed to assess inter-method agreement across distinct protein\u0026ndash;ligand complexes. In the present study, RMSD metrics arise from repeated stochastic executions on the same complexes and therefore do not constitute independent samples suitable for inferential hypothesis testing.\u003c/p\u003e \u003cp\u003eIn this context, the term \u003cem\u003estochastic\u003c/em\u003e refers to docking platforms whose search and/or ranking procedures explicitly incorporate random sampling steps, resulting in non-identical pose rankings across independent executions, rather than simply to randomized initial coordinate placement.\u003c/p\u003e \u003cdiv id=\"Sec31\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Decoupling Pose Accuracy from Reproducibility Reveals Distinct Failure Modes\u003c/h2\u003e \u003cp\u003eThe primary contribution of this study is the demonstration that pose accuracy and run-to-run reproducibility represent complementary and non-interchangeable dimensions of docking performance. Conventional single-run evaluations implicitly assume that a server producing an accurate pose once will do so reliably. The reproducibility-aware framework presented here shows that this assumption does not necessarily hold for web-based docking platforms.\u003c/p\u003e \u003cp\u003eWhen localization, fixed-frame pose placement, and inter-run consistency metrics are considered together, multiple failure modes become apparent. Some docking platforms generate geometrically accurate ligand conformations, as reflected by low aligned RMSD values, yet exhibit substantial variability in pose ranking across independent executions. This behavior indicates that accurate conformations can be sampled but are not consistently selected as the top-ranked solution. Conversely, other platforms produce identical outputs across repeated runs but nevertheless fail to recover native-like poses for a subset of complexes, demonstrating that deterministic behavior alone does not guarantee correctness.\u003c/p\u003e \u003cp\u003eThese observations highlight that pose accuracy and reproducibility interrogate fundamentally different aspects of docking algorithms. Pose accuracy reflects the capacity of a search strategy and scoring function to identify a near-native pose within the energy landscape, whereas reproducibility reflects the stability of the ranking process with respect to stochastic sampling. A docking platform may therefore perform well by one criterion while performing poorly by the other. By explicitly separating these dimensions, the present framework exposes algorithm-dependent trade-offs between sampling breadth and ranking stability. Platforms incorporating extensive stochastic exploration may intermittently access correct poses but display variable convergence behavior, while more deterministic or geometry-driven approaches tend to yield stable outputs that may lack sufficient adaptability for diverse ligand\u0026ndash;receptor geometries.\u003c/p\u003e \u003cp\u003eImportantly, these findings emphasize that pose accuracy alone is insufficient for qualifying web-based docking results under black-box usage conditions. Stability-aware evaluation is required to determine whether an observed pose represents a reproducible prediction or a single-run artifact. The framework introduced here provides a practical means to make this distinction.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec32\" class=\"Section2\"\u003e \u003ch2\u003e4.2. Limitations of Rigid and Geometric Docking Approaches\u003c/h2\u003e \u003cp\u003eThe behavior of PatchDock illustrates methodological constraints associated with rigid-body, geometry-driven docking strategies under re-docking conditions. PatchDock consistently produced near-zero aligned heavy-atom RMSD values across the dataset. This behavior reflects the rigid-body nature of the algorithm, which preserves the internal geometry of the input ligand and does not perform conformational sampling. Because the docking ligand was derived directly from the crystallographic pose and displaced only by rigid-body translation, close geometric agreement following optimal superposition is expected and does not, by itself, indicate successful binding-site localization. Accordingly, centroid distance and In-place RMSD metrics were essential for distinguishing cases of true pocket localization from geometrically aligned poses located outside the binding site.\u003c/p\u003e \u003cp\u003eIn the absence of internal energy minimization or induced-fit modeling, rigid geometric approaches may be limited in their ability to accommodate subtle steric conflicts or conformational adjustments that contribute to native-like pose placement, even under re-docking conditions.\u003c/p\u003e \u003cp\u003eParDOCK exhibited a distinct failure mode characterized by limited reproducible native pose recovery across the evaluated systems. Although both PatchDock and ParDOCK represent ligands using rigid geometries, their underlying search strategies differ substantially. PatchDock relies on purely geometric matching, whereas ParDOCK applies energy-based Monte Carlo sampling with explicit ligand reorientation. Despite employing an all-atom, energy-based protocol, ParDOCK showed limited consistency in native pose recovery under the tested conditions. This discrepancy suggests sensitivity to factors such as sampling efficiency, force-field representation, or scoring discrimination rather than to ligand rigidity alone.\u003c/p\u003e \u003cp\u003eAn additional practical limitation was encountered for ParDOCK during docking of the nuclear receptor target PDB ID: 1Z95 [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Despite repeated submissions using standardized ligand preparation and geometry optimization protocols, ParDOCK consistently rejected the ligand input for this system, preventing pose generation. Consequently, this target was excluded from ParDOCK-specific analyses. While the precise cause of this behavior could not be determined from the web interface, it highlights a general constraint of web-based docking platforms, where input handling and error reporting are not always transparent to the user.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec33\" class=\"Section2\"\u003e \u003ch2\u003e4.3. Implications for \u0026ldquo;Black-Box\u0026rdquo; Users of Web-Based Docking Servers\u003c/h2\u003e \u003cp\u003eA central practical implication of this study concerns how results from web-based docking platforms should be interpreted under routine \u0026ldquo;black-box\u0026rdquo; usage conditions. The increasing accessibility of such tools has contributed to an implicit assumption that single-submission docking outputs are deterministic and directly actionable. The reproducibilty-aware evaluation framework presented here demonstrates that this assumption is not universally valid.\u003c/p\u003e \u003cp\u003eFor docking platforms employing stochastic search or ranking procedures, single-run outputs cannot be assumed to represent stable predictions. Although such platforms may sample native-like binding modes, run-to-run variability indicates that individual submissions may yield alternative high-ranking poses. Within this context, absence of convergence across repeated runs should not be interpreted as lack of algorithmic capability, but rather as a manifestation of stochastic sampling behavior. Replicate docking therefore becomes a methodological requirement rather than an optional refinement step.\u003c/p\u003e \u003cp\u003eBy contrast, platforms exhibiting deterministic or near-deterministic behavior provide consistent pose outputs under identical input conditions. Such behavior supports their use in rapid screening or hypothesis-generation workflows, particularly when computational resources, queue limits, or user expertise restrict the feasibility of repeated submissions. However, deterministic behavior alone does not guarantee correct binding-site localization or accurate pose placement, underscoring the need for complementary geometric validation metrics.\u003c/p\u003e \u003cp\u003eRigid and geometry-driven docking strategies introduce an additional interpretive layer. Deterministic geometric matching can produce reproducible outputs, yet geometric similarity under optimal superposition does not necessarily imply successful localization within the intended binding pocket. Consequently, reliance on aligned RMSD alone is insufficient for evaluating docking success when using rigid or geometry-driven approaches, and localization-aware metrics are essential.\u003c/p\u003e \u003cp\u003eBased on these considerations, the present framework supports a tiered interpretation strategy for web-based docking results. Deterministic platforms are well suited for rapid exploratory analyses, whereas stochastic platforms require replicate sampling and convergence assessment to establish confidence in predicted poses. Rigid or geometry-driven platforms may be useful for coarse shape-based exploration but should be complemented with localization and fixed-frame pose metrics when used in re-docking scenarios.\u003c/p\u003e \u003cp\u003eCollectively, these observations emphasize that effective use of web-based docking servers depends not only on server selection but also on adoption of stability-aware evaluation practices. The framework introduced here provides a practical basis for such interpretation without assuming universal performance hierarchies across docking algorithms\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec34\" class=\"Section2\"\u003e \u003ch2\u003e4.4. Limitations and Future Directions\u003c/h2\u003e \u003cp\u003eThe primary scope of this study is the introduction and demonstration of a reproducibility-aware geometric evaluation framework rather than exhaustive characterization of docking performance across large chemical or target spaces. Accordingly, the selected set of protein\u0026ndash;ligand complexes was chosen to provide controlled test cases for illustrating distinct failure modes in localization, pose placement, and reproducibility, rather than to support statistical generalization across protein families.\u003c/p\u003e \u003cp\u003eWhile the present work demonstrates the utility of decoupling localization, fixed-frame pose accuracy, and inter-run stability, additional validation on larger and more diverse datasets will be necessary to further refine recommended threshold values and to explore how these metrics behave across broader classes of targets and ligand chemotypes. Extension of the framework to systems exhibiting substantial backbone rearrangements, shallow or highly solvent-exposed binding sites, and allosteric pockets represents a particularly important direction.\u003c/p\u003e \u003cp\u003eThe framework currently focuses on pose geometry and reproducibility and does not incorporate binding affinity prediction or enrichment-based virtual screening metrics. Integration of reproducibility and stability-aware geometric evaluation with scoring-function benchmarking represents a natural future extension.\u003c/p\u003e \u003cp\u003eFinally, computational efficiency and wall-clock runtime were not analyzed in this study. For web-based platforms, reported runtimes are strongly influenced by server load, queue policies, and backend hardware in addition to algorithmic complexity. As such, runtime benchmarking is better addressed through controlled local implementations or standardized cloud environments.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusions","content":"\u003cp\u003eThis work introduces a reproducibilty-aware geometric evaluation framework for assessing web-based molecular docking outputs under typical black-box usage conditions. By explicitly decoupling binding-site localization, fixed-frame pose accuracy, and run-to-run reproducibility, the framework demonstrates that conventional single-run docking assessments can mask distinct and practically important failure modes.\u003c/p\u003e \u003cp\u003eApplication of this framework illustrates that accurate pose generation, correct binding-site localization, and reproducible ranking behavior are separable algorithmic properties. Consequently, pose accuracy alone is insufficient to characterize the practical reliability of docking predictions. Incorporation of reproducibility and localization-aware metrics provides essential contextual information for interpreting docking results generated by heterogeneous web-based platforms.\u003c/p\u003e \u003cp\u003eRather than establishing universal performance hierarchies, this study provides a generalizable methodology for diagnosing docking behavior and for identifying when replicate sampling and complementary geometric metrics are necessary. Adoption of such reproducibility-aware evaluation practices can improve confidence in docking-derived hypotheses and support more reliable use of web-based docking tools in structure-based drug design.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eRMSD, Root Mean Square Deviation\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eConflicts of interest statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eVenkata ramana singamaneni is an employee of Cambrex, Charles city, Iowa, USA. This work was conducted outside the scope of his employment, and Cambrex had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The authors declare no other competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll input protein structures, reference ligands, docked pose files, per-run result files and per-run RMSD tables required to reproduce the analyses are provided in data repository, Analysis script used for centroid distance, in-place RMSD, and aligned RMSD calculations along with usage instructions is also included as plain-text files. All data are available without restriction in the Mendeley repository- https://data.mendeley.com/datasets/sv7jcdc4xj/1 \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding-\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors did not receive support from any organization for the submitted work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCRediT authorship contribution statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eManiratnam Puli\u003c/strong\u003e \u0026ndash; Supervision (lead); Conceptualisation (lead); Data curation (lead); writing \u0026ndash; original draft (lead); formal analysis (lead); software (lead); visualisation (lead). \u003cstrong\u003eVenkata Ramana Singamaneni\u003c/strong\u003e- software (supporting), formal analysis (supporting).\u003cstrong\u003e\u0026nbsp;Nikitha Bennuri \u0026ndash;\u0026nbsp;\u003c/strong\u003eData curation (equal); formal analysis (equal); software (equal). \u003cstrong\u003eSravya Peethani\u003c/strong\u003e - Data curation (equal); formal analysis (equal); software (equal). \u003cstrong\u003eHawanika Durgam\u003c/strong\u003e - Data curation (equal); formal analysis (equal); software (equal). \u003cstrong\u003eSowjanya Varipelli\u003c/strong\u003e - Data curation (equal); formal analysis (equal); software (equal).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgment-\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNone\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eKitchen DB, Decornez H, Furr JR, Bajorath J (2004) Docking and scoring in virtual screening for drug discovery: methods and applications. nat rev drug discov 3:935\u0026thinsp;\u0026ndash;\u0026thinsp;49. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/nrd1549\u003c/span\u003e\u003cspan address=\"10.1038/nrd1549\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePinzi L, Rastelli G (2019) Molecular docking: shifting paradigms in drug discovery. Int j of mol sci 20:4331.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKellenberger E, Rodrigo J, Muller P, Rognan D (2004) Comparative evaluation of eight docking tools for docking and virtual screening accuracy. proteins: struct, funct, and Bioinform 57:225\u0026thinsp;\u0026ndash;\u0026thinsp;42.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMartis EA, T\u0026eacute;letch\u0026eacute;a S (2025) Ten quick tips to perform meaningful and reproducible molecular docking calculations. PLOS comput biol 21: e1013030.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCross JB, Thompson DC, Rai BK, Baber JC, Fan KY, Hu Y, Humblet C (2009) Comparison of several molecular docking programs: pose prediction and virtual screening accuracy. J chem inf model. 49:1455\u0026ndash;74.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHartshorn MJ, Verdonk ML, Chessari G, Brewerton SC, Mooij WT, Mortenson PN, Murray CW (2007) Diverse, high-quality test set for the validation of protein\u0026ndash; ligand docking performance. J med chem 50:726\u0026ndash;41.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSahu MK, Nayak AK, Hailemeskel B, Eyupoglu OE (2024) Exploring recent updates on molecular docking: Types, method, application, limitation \u0026amp; future prospects. Int J Pharm Res Allied Sci 13:24\u0026ndash;40.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFeinstein WP, Brylinski M (2015) Calculating an optimal box size for ligand docking and virtual screening against experimental and predicted binding pockets. J cheminf 7:18.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePlewczynski D, Łaźniewski M, Augustyniak R, Ginalski K (2011) Can we trust docking results? Evaluation of seven commonly used programs on PDBbind database. J comput chem 32:742\u0026ndash;55.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWhitlow M, Howard AJ, Stewart D, Hardman KD, Chan JH, Baccanari DP, Tansik RL, Hong JS, Kuyper LF (2001) X-ray Crystal Structures of Candida albicans Dihydrofolate Reductase: High Resolution Ternary Complexes in Which the Dihydronicotinamide Moiety of NADPH Is Displaced by an Inhibitor. J med chem 44:2928\u0026ndash;32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCody V, Luft JR, Pangborn W, Gangjee A, Queener SF. Structure determination of tetrahydroquinazoline antifolates in complex with human and Pneumocystis carinii dihydrofolate reductase: correlations between enzyme selectivity and stereochemistry. Biol Crystallogr 60:646\u0026ndash;55.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRuiz F, Hazemann I, Mitschler A, Joachimiak A, Schneider T, Karplus M, Podjarny A (2004) The crystallographic structure of the aldose reductase\u0026ndash;IDD552 complex shows direct proton donation from tyrosine 48. Biol Crystallogr 60:1347\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNagar B, Hantschel O, Young MA, Scheffzek K, Veach D, Bornmann W, Clarkson B, Superti-Furga G, Kuriyan J (2003) Structural basis for the autoinhibition of c-Abl tyrosine kinase. Cell 112:859\u0026ndash;71.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMol CD, Dougan DR, Schneider TR, Skene RJ, Kraus ML, Scheibe DN, Snell GP, Zou H, Sang BC, Wilson KP (2004) Structural basis for the autoinhibition and STI-571 inhibition of c-Kit tyrosine kinase. J Biol Chem 279:31655\u0026ndash;63.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGolebiowski A, Townes JA, Laufersweiler MJ, Brugel TA, Clark MP, Clark CM, Djung JF, Laughlin SK, Sabat MP, Bookland RG, VanRens JC, De B, Hsieh LC, Janusz MJ, Walter RL, Webster ME, Mekel MJ (2005) The development of monocyclic pyrazolone based cytokine synthesis inhibitors. Bioorg Med Chem Lett 15:2285\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTocchini-Valentini G, Rochel N, Wurtz JM, Moras D (2004) Crystal structures of the vitamin D nuclear receptor liganded with the vitamin D side chain analogues calcipotriol and seocalcitol, receptor agonists of clinical importance. Insights into a structural basis for the switching of calcipotriol to a receptor antagonist by further side chain modification. J Med Chem 47:1956\u0026ndash;61.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim S, Wu JY, Birzin ET, Frisch K, Chan W, Pai LY, Yang YT, Mosley RT, Fitzgerald PM, Sharma N, Dahllund J, Thorsell AG, DiNinno F, Rohrer SP, Schaeffer JM, Hammond ML (2004) Estrogen receptor ligands. II. Discovery of benzoxathiins as potent, selective estrogen receptor alpha modulators. J Med Chem 47:2171\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMadauss KP, Deng SJ, Austin RJ, Lambert MH, McLay I, Pritchard J, Short SA, Stewart EL, Uings IJ, Williams SP (2004) Progesterone receptor ligand binding pocket flexibility: crystal structures of the norethindrone and mometasone furoate complexes. J Med Chem. 47:3381\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBohl CE, Gao W, Miller DD, Bell CE, Dalton JT (2005) Structural basis for antagonism and resistance of bicalutamide in prostate cancer. Proc Natl Acad Sci U S A 102:6201\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePettersen EF, Goddard TD, Huang CC, Meng EC, Couch GS, Croll TI, et al (2021) UCSF ChimeraX: Structure visualization for researchers, educators, and developers. Protein Sci 30:70\u0026ndash;82.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKochnev Y, Ahmed M, Maldonado AM, Durrant JD (2024) MolModa: accessible and secure molecular docking in a web browser. Nucleic Acids Res 52: W498\u0026ndash;W506.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eO'Boyle NM, Banck M, James CA, Morley C, Vandermeersch T, Hutchison GR (2011) Open Babel: An open chemical toolbox. J cheminform 3:33.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWord JM, Lovell SC, Richardson JS, Richardson DC (1999) Asparagine and glutamine: using hydrogen atom contacts in the choice of side-chain amide orientation. J Mol Biol 285:1735\u0026ndash;47.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEberhardt J, Santos-Martins D, Tillack AF, Forli S (2021) AutoDock Vina 1.2. 0: new docking methods, expanded force field, and python bindings. J chem Inform Model 61:3891\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBugnon M, R\u0026ouml;hrig UF, Goullieux M, Perez MA, Daina A, Michielin O, Zoete V (2024) SwissDock 2024: major enhancements for small-molecule docking with Attracting Cavities and AutoDock Vina. Nucleic Acids Res 52: W324-32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGrosdidier A, Zoete V, Michielin O (2011) SwissDock, a protein-small molecule docking web service based on EADock DSS. Nucleic Acids Res 39: W270-7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZoete V, Schuepbach T, Bovigny C, Chaskar P, Daina A, R\u0026ouml;hrig UF, Michielin O (2016) Attracting cavities for docking. Replacing the rough energy landscape of the protein by a smooth attracting landscape. J Comput Chem 37:437\u0026ndash;47.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRohrig UF, Goullieux M, Bugnon M, Zoete V (2023) Attracting cavities 2.0: improving the flexibility and robustness for small-molecule docking. J Chem Inform Model 63:3925\u0026ndash;40.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGupta A, Gandhimathi A, Sharma P, Jayaram B (2007) ParDOCK: an all atom energy based Monte Carlo docking protocol for protein-ligand complexes. Protein and peptide letters. 14:632\u0026ndash;46.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShin WH, Kim JK, Kim DS, Seok C (2013) GalaxyDock2: Protein\u0026ndash;ligand docking using beta-complex and global optimization. J Comput Chem 34:2647\u0026ndash;56.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaek M, Shin WH, Chung HW, Seok C (2017) GalaxyDock BP2 score: a hybrid scoring function for accurate protein\u0026ndash;ligand docking. J Comput Aided Mol Des 31:653\u0026ndash;66.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuhovny D, Nussinov R, Wolfson HJ (2002) Efficient unbound docking of rigid molecules. In International workshop on algorithms in bioinformatics Sep 17 (pp. 185\u0026ndash;200). Berlin, Heidelberg: Springer Berlin Heidelberg.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchneidman-Duhovny D, Inbar Y, Nussinov R, Wolfson HJ (2005) PatchDock and SymmDock: servers for rigid and symmetric docking. Nucleic Acids Res 33(suppl_2): W363-7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVolkamer A, Griewel A, Grombacher T, Rarey M (2010) Analyzing the topology of active sites: on the prediction of pockets and subpockets. J Chem Inform Model 50:2041\u0026ndash;52.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTripathi A, Suri K, K S, Murugan NA (2025) Assessing the accuracy of binding pose prediction for kinase proteins and 7-azaindole inhibitors: a study with AutoDock4, Vina, DOCK 6, and GNINA 1.0. RSC Adv 15:47051\u0026ndash;47065.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"journal-of-molecular-modeling","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jmmo","sideBox":"Learn more about [Journal of Molecular Modeling](https://www.springer.com/journal/894)","snPcode":"894","submissionUrl":"https://submission.nature.com/new-submission/894/3","title":"Journal of Molecular Modeling","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Web-based docking servers, Molecular docking, Active site localization Pose accuracy, Reproducibility, Algorithmic stability","lastPublishedDoi":"10.21203/rs.3.rs-8908837/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8908837/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eContext\u003c/h2\u003e \u003cp\u003eWeb-based molecular docking servers are widely used in structure-based modeling due to their accessibility, yet their outputs are commonly evaluated using single-run pose accuracy metrics that implicitly assume deterministic behavior. Such assumptions can obscure variability arising from stochastic sampling and ranking procedures, particularly in black-box usage scenarios where replicate validation is uncommon. In this study, we introduce a \u003cb\u003ereproducibility-aware geometric evaluation framework\u003c/b\u003e designed to disentangle binding-site localization, fixed-frame pose accuracy, and run-to-run stability. Application of this framework to representative protein\u0026ndash;ligand systems demonstrates that accurate pose generation, correct site localization, and reproducible pose ranking are separable properties. The results illustrate that pose accuracy alone does not fully characterize the practical reliability of web-based docking outputs and motivate inclusion of stability-aware metrics in docking validation workflows.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eFive widely used web-based docking servers representing distinct algorithmic paradigms were evaluated, including a Vina-class reference implementation (MolModa), stochastic sampling platforms (GalaxyDock, SwissDock, PaRDOCK), and a geometry-driven rigid-body approach (PatchDock). Docking outputs were analyzed using complementary geometric metrics: centroid-based localization to assess binding-site placement, in-place root-mean-square deviation (RMSD) to evaluate fixed-frame pose accuracy, and aligned RMSD to quantify conformational agreement independent of spatial placement. Operational stability was assessed using triplicate docking runs under unbiased re-docking conditions. All geometric analyses were performed using custom Python script designed to ensure coordinate-frame consistency and symmetry-aware RMSD evaluation.\u003c/p\u003e","manuscriptTitle":"Beyond Pose Accuracy: Reproducibility-Oriented Framework For Interpreting Web-Based Molecular Docking Outputs Using Decoupled Localization and Pose Fidelity Metrics","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-22 18:10:05","doi":"10.21203/rs.3.rs-8908837/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-05-09T07:06:54+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"333402513789435896408630280401612964100","date":"2026-05-02T04:31:00+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"60673429037237128563477161309458195531","date":"2026-04-26T21:33:11+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-14T15:13:43+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-23T08:14:43+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-23T08:13:46+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of Molecular Modeling","date":"2026-02-18T11:17:19+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-molecular-modeling","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jmmo","sideBox":"Learn more about [Journal of Molecular Modeling](https://www.springer.com/journal/894)","snPcode":"894","submissionUrl":"https://submission.nature.com/new-submission/894/3","title":"Journal of Molecular Modeling","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"9e131f81-22ba-4deb-91c5-efdc6551dc69","owner":[],"postedDate":"April 22nd, 2026","published":true,"recentEditorialEvents":[{"type":"editorInvitedReview","content":"","date":"2026-05-09T07:06:54+00:00","index":32,"fulltext":""},{"type":"reviewerAgreed","content":"333402513789435896408630280401612964100","date":"2026-05-02T04:31:00+00:00","index":31,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-22T18:10:05+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-22 18:10:05","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8908837","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8908837","identity":"rs-8908837","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}