What Docking Scores Represent
Molecular docking produces two outputs: a binding pose (the predicted orientation of the ligand in the protein binding site) and a docking score (an estimate of how tightly that pose binds). The score is the output of a scoring function — a mathematical equation that converts atomic interactions into a single number.
Most scoring functions estimate the binding free energy (ΔG) in kcal/mol. The general principle is straightforward: more negative values indicate stronger predicted binding. A ligand scoring −10 kcal/mol is predicted to bind more tightly than one scoring −6 kcal/mol, all else being equal.
However, docking scores are approximations. They ignore solvent entropy, protein conformational change, and many other factors that contribute to real binding affinity. The correlation between docking scores and experimentally measured IC₅₀ or Kd values is typically moderate (R² ≈ 0.3–0.5). This means docking scores are useful for ranking compounds against the same target, but not for predicting absolute binding affinity.
AutoDock Vina Scores
AutoDock Vina uses a scoring function derived from machine learning on protein-ligand complexes. It combines knowledge-based terms (distance-dependent pair potentials) with empirical energy terms (hydrogen bonds, hydrophobic contacts, rotatable bond penalties).
Score Ranges
| Score Range | Interpretation | Action |
|---|---|---|
| Below −8.0 kcal/mol | Strong predicted binding | Worth experimental validation |
| −6.0 to −8.0 kcal/mol | Moderate binding | Possible hit; verify with consensus scoring |
| Above −5.0 kcal/mol | Weak or no binding | Unlikely to be a meaningful hit |
Imatinib (Gleevec) scores −10.2 kcal/mol against ABL1 kinase (PDB: 2HYY) in AutoDock Vina. The experimental IC₅₀ is approximately 25 nM, which corresponds to a ΔG of roughly −10.4 kcal/mol at 25°C. The Vina score is remarkably close in this case, but this level of agreement is not typical — expect 1–3 kcal/mol error for most systems.
What Changes the Score
- Exhaustiveness: Increasing from 8 to 32 may find better poses with lower (more negative) scores. The default exhaustiveness of 8 is adequate for most targets, but harder targets benefit from 32–64.
- Grid box size: Too small a box may exclude the correct binding site. Too large wastes computation. A 22 Å box centered on the binding site is a good starting point.
- Protein preparation: Missing hydrogens, wrong protonation states, or unsatisfied charges can shift scores by 2–4 kcal/mol.
- Ligand conformation: The starting conformation of the ligand matters less with high exhaustiveness, but a reasonable starting geometry helps.
GNINA CNN Scores
GNINA extends AutoDock Vina with a convolutional neural network (CNN) scoring function trained on receptor-ligand affinity data from PDBbind. The CNN score ranges from 0 to 1, where values closer to 1.0 indicate stronger predicted binding.
GNINA vs Vina
| Feature | AutoDock Vina | GNINA |
|---|---|---|
| Score type | kcal/mol (ΔG estimate) | CNN affinity (0–1) |
| Better binding | More negative | Higher (closer to 1.0) |
| Speed | Fast (5–10 sec) | Moderate (15–30 sec with GPU) |
| Training data | Knowledge-based potentials | PDBbind affinity data |
| Best for | Virtual screening | Pose scoring and rescoring |
GNINA's CNN scoring can rescore Vina poses to improve ranking accuracy. A common workflow is: dock with Vina, rescore the top poses with GNINA, then rank by CNN affinity.
RMSD: Structural Agreement
RMSD (Root Mean Square Deviation) measures how different two molecular structures are. In docking, it quantifies the distance between a predicted pose and a reference (usually the crystal structure).
Interpreting RMSD Values
| RMSD | Interpretation |
|---|---|
| Below 2.0 Å | Good agreement — pose reproduces experimental binding mode |
| 2.0–4.0 Å | Moderate agreement — key interactions may be preserved |
| Above 4.0 Å | Poor agreement — different binding mode from reference |
The industry benchmark is 2 Å RMSD: if a docking program reproduces the crystallographic pose within 2 Å, it is considered successful. AutoDock Vina achieves this in approximately 70–80% of cases on well-characterized binding sites.
When RMSD Matters More Than Score
For drug discovery, pose quality often matters more than the score. A pose with a score of −8.5 kcal/mol that shows clear hydrogen bonds to catalytic residues and good shape complementarity is more trustworthy than a pose scoring −10.0 kcal/mol with no clear interactions. Always visualize the best poses before making decisions.
Common Pitfalls
Understanding these pitfalls prevents costly mistakes in docking studies:
- Comparing scores across tools: Vina scores of −9.0 and SwissDock scores of −9.0 are not equivalent. Different scoring functions use different energy terms and scales. Always compare compounds using the same tool.
- Ignoring protein flexibility: Most scoring functions assume a rigid protein. Induced-fit effects (where the protein changes conformation upon ligand binding) can shift the true binding mode far from the docked pose.
- Over-trusting the top pose: Vina returns up to 9 poses. The top-ranked pose is not always the correct one. Check that poses 1–3 are structurally similar (low RMSD between them) for higher confidence.
- Ignoring solvent: Most scoring functions treat water implicitly. Binding sites with structured water molecules (like HIV protease) can have significant errors in docking scores.
- Selectivity blindness: A compound scoring −10 kcal/mol against one target may also bind related proteins with similar affinity. Always dock against close homologs if selectivity matters.
Put Score Interpretation into Practice
Dock imatinib against ABL1 kinase and examine the binding poses, scores, and interaction maps.
Run Docking Free →Frequently Asked Questions
What is a good docking score?
A good docking score depends on the scoring function used. For AutoDock Vina, scores below −8.0 kcal/mol suggest good binding affinity, −6.0 to −8.0 indicates moderate binding, and above −5.0 suggests weak or no meaningful binding. For GNINA CNN scores, values closer to 1.0 indicate stronger predicted binding. Always compare scores within the same scoring function and target, and validate predictions experimentally.
What is binding affinity?
Binding affinity is the strength of the interaction between a protein and a ligand, typically measured as the dissociation constant (Kd) or inhibition constant (Ki). Lower Kd values indicate tighter binding. Docking scores estimate binding free energy (ΔG) in kcal/mol, where more negative values indicate stronger predicted binding. The relationship between ΔG and Kd follows: ΔG = RT·ln(Kd), where R is the gas constant and T is temperature.
What is RMSD in docking?
RMSD (Root Mean Square Deviation) measures the structural difference between two protein-ligand poses. In docking, it quantifies how far a predicted pose deviates from a reference, typically the crystal structure. An RMSD below 2 Å is considered good agreement with the experimental structure. The industry benchmark is that docking programs should reproduce crystallographic poses within 2 Å in 70–80% of cases.
How accurate are docking scores?
Docking scores are approximations with moderate correlation to experimental binding affinities. Scoring functions typically achieve R² values of 0.3 to 0.5 against experimental data. They are most reliable for relative ranking of compounds against the same target, not for predicting absolute binding affinity. Combining docking with molecular dynamics, free energy calculations, or experimental validation significantly improves reliability.
What is a scoring function?
A scoring function is a mathematical equation that estimates the binding free energy between a protein and ligand based on their atomic interactions. Common types include force-field based (van der Waals, electrostatics), empirical (weighted energy terms trained on experimental data), and knowledge-based (statistical potentials derived from protein-ligand complexes). Each docking tool uses a different scoring function, which is why scores cannot be directly compared across tools.
References
- Trott O., Olson A.J. (2010). AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of Computational Chemistry, 31(2), 455-461.
- McNutt A.T., et al. (2021). GNINA 1.0: Molecular docking with deep learning. Journal of Cheminformatics, 13(1), 43.
- Warren G.L., et al. (2006). A critical assessment of docking programs and scoring functions. Journal of Medicinal Chemistry, 49(20), 5992-6016.
- Kontoyiannis M., et al. (2021). Scoring functions for molecular docking. WIREs Computational Molecular Science, 11(6), e1549.