Docking Score Interpretation: What -10 kcal/mol Means

Someone hands you a docking result: -10.35 kcal/mol, labelled "very strong binding". The number has energy units, so it looks like a binding free energy. It is not one. If you treat it as one, you will over-promise a compound to your project team and be wrong by orders of magnitude in Kd.

This is what that number actually is, what it can carry, and where it breaks.

The worked example: sotorasib into KRAS G12C

We ran sotorasib (AMG 510) against KRAS G12C through covadock_dock on GPU. The receptor was the deposited crystal structure, PDB 6OIM. Ligand structure came from covabasic_pubchem (CID 137278711, C30H30F2N6O3, MW 560.6) and was checked with covabasic_validate (InChIKey NXQKSXLFSAEQCZ-SFHVURJKSA-N).

Verbatim outputs:

That is a good result. Sotorasib is a real approved KRAS G12C drug, and the pose lands in the switch-II pocket the drug class is designed for. The contacts are chemically the right ones. So this is close to the best case for docking: known drug, high-quality co-crystal receptor, well-defined pocket.

And even in that best case, the number is not a binding free energy.

Where the "kcal/mol" comes from

The default engine behind covadock_dock is AutoDock Vina (default exhaustiveness=16, 10 poses per engine, PoseBusters validity check on the top poses). Vina's scoring function is an empirical sum of weighted terms - steric, hydrophobic, hydrogen bonding, plus a rotatable-bond penalty - with weights fitted to reproduce measured affinities across a training set.

Two consequences follow directly:

1. The output has energy units because the fit target had energy units. There is no partition function, no explicit solvent, no entropy calculation. It is a regression output wearing a thermodynamic costume. 2. Its accuracy is bounded by the fit, not by physics. Off the training distribution, it degrades without warning.

The label CovaDock prints is a fixed threshold scale, published in the tool docstring so you can see the cut points: weak (> -6), moderate (-6 to -8), strong (-8 to -10), very strong (< -10). Sotorasib at -10.35 clears the last bin by 0.35 kcal/mol. That margin is far inside the method's error bar.

How big is the error bar? The benchmark numbers

The standard reference is CASF-2016 (Su et al.), a core set of 285 protein-ligand complexes across 57 targets with 5 ligands each. Published AutoDock Vina results on that set:

• Docking power (does the top-ranked pose reproduce the crystal pose within 2.0 A RMSD): 90.2% top-1 success. Increases to 94.4% when structural waters are included.
• Scoring power (correlation of score with measured affinity): Pearson R = 0.604, standard deviation 1.73 log-affinity units.
• Ranking power (ordering ligands correctly within one target): Spearman rho = 0.528.

Read those three lines together, because they say three different things.

Vina finds the right pose about nine times in ten on well-behaved crystal structures. That is the part that works, and it is why the 61 switch-II contacts in the sotorasib run are worth looking at.

Vina's affinity number carries R = 0.604 with an SD of 1.73 log units. At 298 K, one log unit of Kd is RT ln(10) = 1.36 kcal/mol, so 1.73 log units is roughly 2.4 kcal/mol of residual scatter. Sotorasib's -10.35 therefore sits inside a distribution that comfortably spans -8 to -13 in equivalent terms. A 0.35 kcal/mol difference between two compounds is noise. A 3 kcal/mol difference is a weak signal.

Spearman 0.528 is the number most people should care about most, because within-target ranking is what a hit list actually is. Roughly: the ordering is better than random and clearly worse than reliable.

Enrichment versus affinity: the distinction that matters

These are two different jobs, and docking is only good at one of them.

Practical translation:

• Use scores to cut a large list. Good.
• Use scores to order the top 20 for synthesis. Weak; combine with interaction patterns, pose sanity and chemical judgement.
• Use scores to predict Kd or to claim compound A beats compound B by 0.5 kcal/mol. Do not.

The covalent problem, illustrated by this exact compound

Sotorasib works by forming a covalent bond between its acrylamide warhead and Cys12. That bond is the entire mechanism.

The CovaDock run scored it non-covalently. The scoring function has no term for bond formation. So -10.35 kcal/mol describes only the reversible recognition step that positions the warhead - it understates the real thermodynamics of the drug by a large and unquantified amount.

This cuts both ways and is worth internalising:

• For a covalent series, the docking score is a proxy for warhead placement, not potency. Check whether the reactive atom sits near the target cysteine in the pose. That is the readout with information in it.
• Comparing a covalent compound against a non-covalent one on docking score alone is meaningless. You are comparing two different physical quantities.

What the interaction count does and does not add

61 interactions sounds like strong evidence. It is useful, but not as a magnitude.

Interaction counts depend on the geometric cutoffs the analyser uses, and they scale with ligand size. A 560 Da molecule will make more contacts than a 300 Da fragment, whether or not it binds better. What the list is genuinely good for is identity: the contacts landed at Gly24, Ala25, Gly74 and Gln75, which is the switch-II region. That tells you the pose is in the right place. A high score with contacts scattered outside the known pocket is a red flag no number will surface for you.

Same logic applies to ligand efficiency. Sotorasib has 41 heavy atoms (from the C30H30F2N6O3 formula returned by covabasic_pubchem). LE = 10.35 / 41 = 0.25 kcal/mol per heavy atom. Respectable for a large late-stage molecule, unimpressive for a fragment. The raw score alone hides that entirely.

Honest limits

What a CovaDock score does not tell you:

• Not a Kd, Ki or IC50. No conversion factor exists that makes it one.
• No covalent chemistry. Warhead reactivity, k_inact and residence time are outside the model.
• Receptor-dependent. The 90.2% docking power figure comes from curated crystal structures. Dock into a predicted model or a poorly resolved pocket and both pose and score degrade. Our run used crystal 6OIM deliberately.
• Rigid-receptor by default. Induced fit, cryptic pockets and large loop rearrangements are not sampled.
• Waters and ions dropped. CASF shows a 90.2% to 94.4% swing on that choice alone.
• No selectivity, no ADMET, no developability. A -10 score says nothing about whether the compound is safe or gets anywhere near the target in vivo. That is a separate screen (covatox_structural_alerts, covatox_ich_m7).

CovaDock is triage. It narrows what you make and test. It does not replace a binding assay, and no docking output belongs in a regulatory filing as evidence of affinity.

Frequently asked questions

Related reading

• pLDDT explained: when a predicted structure is good enough to dock into

Run your own target and ligand through covadock_dock on the free tier and read the score the way this article suggests: pose first, contacts second, number last.