Open Source DockMate-VS Framework Standardises Molecular Docking Protocol Development and Virtual Screening

Computational drug discovery workflows depend heavily on molecular docking algorithms to predict how small molecules bind to receptor targets, yet early technical decisions can dramatically alter experimental rankings before prospective biological testing even begins. Software developer Gerrit Koorsen has introduced DockMate-VS, an open-source Python application designed to formalise molecular docking protocol development, validation, and structure-based virtual screening within an inspectable framework. Documented in SoftwareX, the platform addresses a persistent methodological challenge in structure-based virtual screening: protocol choices such as protein protonation, crystallographic water retention, grid box dimensions, search exhaustiveness, and scoring functions frequently bias pose prediction and compound prioritisation. Rather than assuming a single arbitrary docking configuration, the software treats parameter selection prior to a screening campaign as an empirical experiment, systematically comparing multiple docking engines, binding-site definitions, and rescoring approaches across reproducible benchmarks.

Standard virtual screening pipelines often rely on custom command-line orchestration scripts to link standalone utilities like AutoDock Vina or Smina with external preparation tools. This ad hoc approach frequently leads to reproducibility hurdles, unrecorded parameters, lost score provenance, and obscured failure modes during execution. Moreover, pooling receiver operating characteristic area under the curve metrics across heterogeneous receptor structures risks masking severe underperformance at individual protein sites. DockMate-VS resolves these technical fragmentation issues by coordinating protein and ligand preparation, binding-site specification, docking execution, root-mean-square deviation calculations, and diagnostic enrichment analyses within a modular graphical and headless architecture. Built using Python, Tkinter, RDKit, Biopython, MDAnalysis, and Gemmi, the software structures computational campaigns into two distinct sequential stages: a protocol development phase that evaluates parameter permutations on co-crystal complexes, and a screening phase that applies a frozen protocol to labelled controls and uncharacterised chemical libraries.

During ligand preparation, the tool standardises input chemical structures, retains the principal molecular fragment, neutralises structures where permissible using RDKit or preserves supplied formal charges, and records resulting charges for downstream auditing. The pipeline enumerates conformers and tautomers within user-defined constraints and performs three-dimensional embedding and energy minimisation, with Open Babel generating rigid ligand and receptor PDBQT representations. The current release intentionally omits automated pKa prediction and pH-dependent ionisation state enumeration, docking either user-specified or neutralised states while caching variants for future workflows. Receptor preparation integrates structural repair through OpenMM and PDBFixer, optional PROPKA and Reduce processing, polar hydrogen addition, and configurable crystallographic water treatments ranging from total removal or complete retention to selective preservation based on crystallographic criteria. Cavity definitions can be derived directly from co-crystallised ligands, user-specified residues, fixed coordinate boxes, or automated fpocket cavity predictions.

The software establishes an explicit separation between pose generation, pose ranking, and active-inactive compound discrimination. A protocol capable of recovering a crystallographic pose does not inherently discriminate genuine active compounds from property-matched decoys, just as high enrichment factors do not confirm that top-ranked geometries are structurally sound. In protocol development mode, DockMate-VS sweeps across supported primary docking engines including Smina and AutoDock Vina, with optional rDock support, executing calculations through command-line adapters defined in a shared registry. For AutoDock Vina and Smina calculations, baseline pose rankings are recorded alongside optional rescoring functions like Vinardo across saved configurations. The platform computes heavy-atom root-mean-square deviations while accounting for symmetry-equivalent atom mappings, allowing investigators to isolate whether docking failures stem from conformational search limitations or scoring function deficiencies before fixing parameters for extensive screening libraries.

To illustrate the exploratory protocol development workflow, the study investigated the public LIT-PCBA ESR1 antagonist target utilising the 1.8-angstrom crystallographic structure of estrogen receptor alpha (PDB entry 1XP1, chain A) bound to the noncovalent ligand AIH. The sweep examined AutoDock Vina, Smina, and rDock engines across box margins of 4, 5, and 6 angstroms, three distinct water-handling policies, exhaustiveness values of 8, 16, and 32, and native versus Vinardo-rescored rankings, with random seeds set to 42. After collapsing redundant rDock exhaustiveness conditions, the analysis evaluated 126 distinct protocol configurations. Docking runs generated a sub-2-angstrom pose in 116 out of 126 conditions, recovering sub-2-angstrom poses within the top five and top ten ranks in 47 and 108 conditions, respectively. However, no protocol ranked a sub-2-angstrom pose in the top position, pinpointing scoring function limitations rather than insufficient spatial sampling. Furthermore, Vinardo rescoring failed to improve Top-5 root-mean-square deviations across 63 paired evaluations, leaving 26 unchanged and worsening 37 with a median deviation increase of 0.82 angstroms.

The screening module was demonstrated on a curated, deterministic derivative of the LIT-PCBA ESR1 antagonist benchmark comprising 80 verified active compounds and 800 inactive molecules selected without replacement. Using the frozen AutoDock Vina configuration with a 6-angstrom box margin, retained crystallographic waters, an exhaustiveness setting of 8, and seed 42, all 880 compounds completed execution successfully. Active compounds achieved an average docking score of minus 7.86 kilocalories per mole compared to minus 6.91 kilocalories per mole for inactive compounds, yielding a receiver operating characteristic area under the curve of 0.695 and an average precision of 0.174 against an active prevalence baseline of 0.091. Stratified bootstrapping across 10,000 replicates established 95 percent confidence intervals of 0.633 to 0.752 for the area under the curve and 0.142 to 0.235 for average precision, with the highest-ranked active compound emerging at rank five and expected enrichment factors reaching 1.83, 2.10, and 2.19 at the top 1, 5, and 10 percent of the screened library.

Diagnostic routines embedded in DockMate-VS revealed that molecular size heavily influenced the observed screening separation. Actives possessed higher average molecular weights than inactive decoys (328.2 daltons versus 267.1 daltons), and molecular weight correlated inversely with raw docking scores (Spearman rho of minus 0.6489, corresponding to a positive correlation of 0.65 with negative docking score). Ranking compounds strictly by molecular weight yielded an area under the curve of 0.712 with a 95 percent confidence interval of 0.660 to 0.760, outperforming the raw docking score itself. When removing rank dependence on molecular weight, discrimination declined to 0.594, while evaluating 80 molecular-weight-matched active and inactive pairs produced an area under the curve of 0.549 with a 95 percent confidence interval of 0.468 to 0.623, encompassing random classification. The partial Spearman correlation between score and biological activity fell from 0.194 to 0.076 after controlling for molecular weight, demonstrating that docking scores contributed minimal predictive power beyond ligand size for this dataset.

To verify execution fidelity against historical literature, the author reproduced an established Smina screening experiment on the ACHE-DEKOIS 2.0 dataset originally published by Tran-Nguyen and Ballester. Reusing archived 1EVE receptor coordinates and prepared MOL2 libraries containing 40 active compounds and 1,200 decoys under identical parameters (exhaustiveness 1, seed 0, and native Vina scoring), all 1,240 screening calculations completed without failure. DockMate-VS matched published results exactly, yielding identical trapezoidal precision-recall area under the curve values of 0.194743 and average precision of 0.195322, both rounding precisely to the historical reference value of 0.195. Recalculated receiver operating characteristic area under the curve values matched at 0.648833, all 1,240 native one-decimal affinity scores were identical with a maximum difference of 0.0 kilocalories per mole, and compound ranking was fully preserved at a Spearman rho of 1.000000.

Coordinate verification against the historical ACHE benchmark confirmed identical spatial configurations for all 1,240 compounds at stored precision. Detailed file audits identified that 101 decoys contained altered atom-type annotations and 172 exhibited differing partial-charge annotations compared to the primary benchmark archive, though coordinates, types, and charges matched the authors’ companion AVE-vs-DEKOIS reference archive completely. For operational durability, DockMate-VS records execution parameters, tool versions, and random seeds within JSON run manifests, while case outputs are written incrementally. This design allows interrupted campaigns to resume seamlessly by re-attempting failed or unwritten cases, while automatically invalidating partial results if conflicting docking or scoring settings are introduced.

DockMate-VS provides structured outputs designed for downstream biophysical workflows rather than fully automated simulation setups. Generated directory structures preserve prepared receptor PDB and PDBQT files, ligand variants, multi-model docked PDBQT poses, individual rescored geometries, and crystallographic reference coordinates alongside CSV, JSON, and Markdown summaries. Dedicated export commands extract selected poses as structure-data files preserving input ligand topology or construct receptor-ligand complex PDB assemblies for molecular dynamics engines including GROMACS, AMBER, and NAMD, or free-energy frameworks like BFEE2 and BAT.py. The software currently focuses on non-covalent docking, requires external parameterisation for solvation and force-field generation, operates officially on Linux and macOS with community Docker support for Windows, and maintains continuous integration verification through 309 automated tests under Python versions 3.9 through 3.12.

Subject of Research: Open-source molecular docking protocol development and virtual screening software

Article Title: DockMate-VS: Molecular docking protocol development and structure-based virtual screening

Article References: Koorsen, G. (2026). DockMate-VS: Molecular docking protocol development and structure-based virtual screening. SoftwareX, 36, Article 103226. https://doi.org/10.1016/j.softx.2026.103226

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103226

Keywords: molecular docking, virtual screening, DockMate-VS, structure-based drug design, AutoDock Vina, Smina, rDock, Vinardo, computational chemistry, pose recovery, enrichment analysis, receptor preparation, ligand preparation, bioinformatics software, reproducible research, ESR1, open-source software