Vina vs Vinardo: How to Choose a Scoring Function for Reproducible Molecular Docking

This guide is for computational chemists, molecular modelers, and research teams deciding whether a small-molecule docking workflow should use Vina, Vinardo, or a target-specific comparison of both. The practical question is not “Which scoring function wins everywhere?” It is “Which function is sufficiently reliable for the decision this project needs to make?”

That distinction matters because docking can be asked to recover a plausible pose, rank ligands for one target, enrich known actives in a screen, or approximate measured affinity. Those are related tasks, but they are not interchangeable. A scoring function can perform differently across them [1].

Vina is the default empirical scoring function associated with AutoDock Vina. Vinardo—short for Vina RaDii Optimized—was developed from the Vina framework and originally evaluated in scoring, docking, ranking, and virtual-screening tests by Quiroga and Villarreal [1]. Current AutoDock Vina exposes vina and vinardo as selectable scoring-function names in its official API [2].

Vinardo retains the broad families of steric, hydrophobic, and nondirectional hydrogen-bond interactions used by Vina, but it is not merely a renamed preset. Its published formulation removes Vina's second attractive Gaussian term and changes atomic radii, parameters, and term weights [1]. Those changes alter the landscape used to evaluate and search protein–ligand poses.

The original Vinardo study separates four capabilities that are often compressed into the word “accuracy” [1]. Keeping them separate makes a software and protocol decision much easier to defend.

A fair comparison changes one variable: the scoring function. If protonation, receptor coordinates, ligand geometry, search space, or random seed also changes, the result can no longer be attributed to Vina versus Vinardo alone. AutoDock Vina's official API exposes the scoring function and seed as explicit inputs, making both part of the recorded protocol [2].

Preparation deserves particular attention. Use the same prepared receptor and ligand files—not merely files with the same names. Retain checksums when possible. Then preserve the complete commands, software version, pose files, logs, and analysis method alongside the result.

The 2016 Vinardo paper reported better results than Vina across the datasets and tests used in that study [1]. That is meaningful published evidence for considering Vinardo. It is not proof that Vinardo must win for every protein family, receptor state, ligand series, preparation protocol, or screening objective.

For pose prediction, begin with multiple crystallographic complexes relevant to the target or close analogs—not one convenient success. Predefine the near-native criterion, commonly an RMSD threshold for the redocked ligand, and report both successes and failures. For screening, use target-specific actives and well-designed decoys or inactives, then compare the metric that matches the project: early enrichment, ranking, recall at a budget, or another prespecified endpoint.

To demonstrate the comparison operationally, we redocked imatinib (STI-571) into the c-Abl kinase structure PDB 1IEP. RCSB PDB identifies 1IEP as an X-ray structure of the c-Abl kinase domain in complex with STI-571 at 2.10 Å resolution [4]. The official AutoDock Vina basic docking tutorial also uses this complex and explicitly documents how to switch the same example to Vinardo [3].

We ran both functions through MolNexus 0.1.0 and its bundled AutoDock Vina 1.2.7 engine. The prepared receptor, prepared ligand, reference ligand, search box, exhaustiveness, requested modes, energy range, seed, and CPU setting were identical. Only --scoring vina versus --scoring vinardo changed.

Both runs recovered the crystallographic pose well below the prespecified 2 Å threshold. Vina produced the more negative numerical score; Vinardo produced the slightly lower crystallographic RMSD. If the choice had been made from score magnitude alone, the pose-recovery result would have been ignored. If it had been made from this one RMSD alone, the comparison would still be too narrow to support a universal claim.

The useful conclusion is narrower and stronger: both functions succeeded for 1IEP under this protocol, and they produced measurably different score, pose, and retention outcomes while every declared input was held fixed. The manifests, commands, reports, output poses, checksums, and source captures are retained with this article's evidence record.

Running both functions can be useful, but “two scores are better than one” is not a validation method. Decide in advance how disagreement will be handled. Will one function be selected from a validation set? Will each generate an independent shortlist? Will a compound advance only when its pose is chemically plausible under both? The rule should exist before the results are visible.

For large libraries, test the protocol on a representative validation subset before committing the entire compute budget. This is also where workflow friction becomes visible: repeated preparation, untracked parameter changes, ambiguous filenames, or results separated from their poses can cost more than the additional docking run.

MolNexus exposes Vina and Vinardo within the same local desktop workflow. The scoring selector sits beside exhaustiveness, requested poses, energy range, and seed, while the receptor, ligand list, search box, three-dimensional pose, and ranked results remain in context. That does not validate a scoring function automatically; it reduces the operational distance between a fair comparison and its evidence.

Completed work can be retained in local SQLite history and reopened with its context. Results can be exported to PDF and Excel, and selected complexes can be exported for downstream inspection. If your larger problem is organizing receptors, ligands, runs, and poses, see the guide to keeping molecular docking runs organized. For broader purchasing criteria, use the molecular docking software selection guide.

A scoring-function choice becomes defensible when another researcher can reconstruct why it was made. Start with the endpoint, control the protocol, retain the evidence, and let target-relevant performance—not familiarity or the most negative number—decide.

If predicted structures are part of the upstream workflow, the guide to AlphaFold structures and molecular docking explains where receptor confidence enters the decision before either scoring function is run.