> ## Documentation Index > Fetch the complete documentation index at: https://docs.revilico.bio/llms.txt > Use this file to discover all available pages before exploring further. # Protein Family Benchmarking across 6 Engines > Six docking and co-folding engines compared on pose accuracy, affinity scoring and physical validity across 14 protein families. Six docking and co-folding engines, each run on co-crystal structures from 14 protein families and scored against the crystal pose and against measured binding affinity. The benchmark is described in our white paper, [*Structural Context Determines Docking Engine Performance: A Family-Stratified Benchmark of Six Engines*](https://www.biorxiv.org/content/10.64898/2026.08.11.744016v1) (Alejo et al., bioRxiv, 2026). This page reproduces its results and adds interactive views of the per-family data.> **[Read the white paper](https://www.biorxiv.org/content/10.64898/2026.08.11.744016v1)** > > Structural Context Determines Docking Engine Performance: A Family-Stratified Benchmark of Six Engines. bioRxiv preprint, doi:10.64898/2026.08.11.744016.> [!NOTE] > > **RevDock** is Revilico's fork of GNINA 1.3. All six engines were run on the same set of structures, so the results are directly comparable. **Overall** numbers are the ones published in the paper. **Per-family** numbers come from the benchmark data release that accompanies it, with the sample size of each cell shown or one hover away. The release covers five of the six engines; PandaDock appears in the overall results only.## Where each engine standsFour ways to measure a docking engine: **docking power** (does it put the ligand where the crystal has it), **scoring power** (does its score track measured affinity), **ranking power** (does it order ligands correctly), and **physical validity** (are its poses chemically sensible). \ \
Engine
\**Bold:** a RevDock value that beats at least one other engine in the column. *Italic, underlined:* the best value in the column. **-**: not reported (DiffDock and rDock scores do not map onto an affinity scale, so they have no scoring power; the paper does not give a mean ρ for PandaDock). Docking power, scoring power and physical validity are overall results from the paper. † Mean of the 14 per-family values in the data release. Lower median RMSD is better. RevDock has the highest pose accuracy, 73.3% of poses within 2 Å of the crystal pose, and ties Boltz2 on affinity correlation (mean r of 0.605 against 0.604). PandaDock matches it on pose accuracy (72.1%) and has the lowest median RMSD. DiffDock and rDock have no usable ranking power overall. RevDock has the lowest physical-validity score of the six, 73.2% of its poses passing the PoseBusters checks against 93% to 98% for the others, so its poses are the most accurate and the least often chemically clean.## Overall results, one metric at a time## Results by protein familyThe ranking is not the same in every family, which is the paper's main point: which engine is best depends on the structural character of the target, not on the overall ranking. Switch the metric to see where each engine leads, and hover a cell for its sample size.
Family
Families led
**Bold:** a RevDock value that beats at least one other engine in the family. *Italic, underlined:* the best value in the family. Cell shading runs from weak to strong. **-**: not reported (DiffDock and rDock have no scoring power). † Fewer than 10 structures: shown faded and not counted in "Families led". Ties count for every engine that shares the best value. Hover a cell for its sample size. PandaDock is not shown here because its per-family results are not in the data release; it appears in the overall results above. Docking-power rows use the co-crystal structures retrieved from RCSB (the PDBbind structures cannot be redistributed), so their sample sizes are smaller than the overall figures.PandaDock is not in this view because its per-family results are not part of the released tables; its overall results are in the scoreboard above.For example, RevDock is the most accurate engine on Phosphatases (94% of poses within 2 Å) while DiffDock is the best on Protein kinases (90%), and on affinity ranking Boltz2 and RevDock both peak on PPI targets (ρ of 0.84 and 0.87). Glycosidases have only two structures with a crystal pose in the released set, so their docking-power cells are shown faded and are not counted.## Predicted against experimental affinityEvery structure with a measured affinity, for the three engines whose scores map onto an affinity scale and that have per-family results in the data release. Pick a family to see how the relationship changes.Pearson rstructuresPredictions and experimental values are z-scored within each family, so engines with different score units share one axis. The dashed line is perfect agreement. With all families selected, r is pooled over the z-scores, so it can differ from the mean of the per-family values above. Only engines whose scores map onto an affinity scale and that have per-family results in the data release are shown.## How the benchmark was run> **The six engines** > > | Engine | Type | Native score used | Units | > | - | - | - | - | > | **RevDock** | Deep-learning docking with affinity scoring (Revilico fork of GNINA 1.3) | CNNaffinity | pKd | > | **PandaDock** (v4.1.1, expanded sampling) | Physics-based docking | top-ranked score | ΔG, kcal/mol | > | **Boltz2** | Co-folding model | predicted affinity | ln(IC50) | > | **AutoDock-GPU** | GPU genetic-algorithm docking | top score of 20 runs | ΔG, kcal/mol | > | **DiffDock** | Diffusion-based blind docking | confidence | dimensionless | > | **rDock** | CPU cavity docking | score (negated before ranking) | dimensionless | > > Boltz2, RevDock, PandaDock and AutoDock-GPU produce scores that map onto an experimental affinity or free-energy scale, so they get scoring power (Pearson r, R²). DiffDock and rDock scores do not, so for them only the rank correlation is reported.> **Structures and filters** > > Candidate targets were identified through ChEMBL and gathered from the RCSB Protein Data Bank, PDBbind v2020, the CASF-2016 core set and the PoseBusters set. Only X-ray structures were kept, with resolution of 2.5 Å or better, R-free of 0.30 or lower, a ligand of 100 to 800 Da with at least 5 heavy atoms, and at least one measured IC50, Ki, Kd or EC50. Ions, buffer components, crystallographic sugars and modified residues were excluded. Each structure belongs to exactly one family. In total, 12,420 docking runs were attempted and 11,036 completed.> **How the metrics are computed** > > * **Docking power:** the share of poses within 2.0 Å heavy-atom RMSD of the crystal ligand, and the median RMSD, using symmetry-corrected RMSD without realignment (Kabsch protein alignment first for Boltz2, which predicts its own frame). > * **Scoring power:** Pearson r and R² between the predicted and experimental affinity, within each family, in each engine's native units. > * **Ranking power:** Spearman ρ over the same pairs. > * **Physical validity:** the share of poses passing the PoseBusters molecule checks (bond geometry, steric clashes and similar).> **Where each number comes from** > > The overall engine results are the values reported in the paper. The per-family tables and the scatter plots come from the benchmark data release, which holds five of the six engines, so PandaDock appears only in the overall results. The release does not include the PDBbind index files, which are licensed, so the affinity analyses cannot be reproduced from the release alone.## Reading these results> **Overall and per-family docking power use different sets** > > The overall docking-power figures use every structure with a crystal pose. The per-family tables use the co-crystal structures retrieved from RCSB, because the PDBbind structures cannot be redistributed, so their sample sizes are smaller and some families have only a handful of structures. That is why, for example, the per-family numbers for rDock and DiffDock average lower than the overall figures.> **Small families are noisy** > > Several families have few structures with a crystal pose or a measured affinity. Cells with fewer than 10 are shown faded and are left out of the "families led" count, and the Glycosidases docking-power cells rest on two structures each.> **No decoys, so no screening power** > > The benchmark contains binders and their co-crystal structures only. It does not measure how well an engine separates binders from non-binders in a screen.> **Scores are not on one scale** > > Each engine reports a different kind of number, so comparing r or ρ across engines compares how well each engine's own score orders the same ligands, not the size of its predicted affinities. The scatter plot z-scores each family so the engines can share an axis.## Sources- **Paper:** Alejo, K., Fisher, S., Kalluri, T., More, B., Rajgure, H., Panda, P. K., Korban, C., Chung, C. *Structural Context Determines Docking Engine Performance: A Family-Stratified Benchmark of Six Engines.* bioRxiv (2026). [doi:10.64898/2026.08.11.744016](https://www.biorxiv.org/content/10.64898/2026.08.11.744016v1) - **Overall results:** the paper's Results and Discussion sections. - **Per-family results and scatter plots:** the Protein Family Docking Benchmark data release that accompanies the paper (raw docking outputs, processed correlation and RMSD tables). - **Structures and affinities:** RCSB Protein Data Bank, PDBbind v2020, CASF-2016 and the PoseBusters benchmark set. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.