Docking score, binding energy and binding affinity: what the numbers mean

Three terms that share a unit and mean different things

17 min

Line drawing of biotin in the eight-stranded beta barrel of streptavidin, seen from above with the ureido ring shaded blue, beside a to-scale step ladder of its binding enthalpy, entropy cost and net free energy.

Docking score vs binding energy vs affinity

Binding affinity is how tightly a ligand binds a target, measured as a dissociation constant (Kd). Binding free energy (ΔG) is that same fact as an energy, ΔG = RT ln Kd. A docking score is a software estimate of ΔG for one pose. "Binding energy" is looser: it can mean the free energy, a force-field interaction energy or, in chemistry and physics, a bond or nuclear binding energy. Only affinity and ΔG are measured.

A shared unit hides the gap. AutoDock Vina prints −12.51 kcal/mol for imatinib on c-Abl, while its measured apparent Kd of 106 nM is −9.52 kcal/mol.[11] MM/PBSA stretches it further: for three charged biotin analogues in avidin it gives −35 to −53 kcal/mol, while the tightest of seven measured avidin ligands reaches −20.4.[6] The numbers look alike and are not.

Which number is which

Table 1 lists the quantities that get called affinity or energy. Only the first two are the same thing in different units, and only the first is measured directly.

QuantityWhat it isUnitWhere it comes from
Binding affinity (Kd, Ki)How tightly a ligand binds one target at equilibrium; lower Kd means tighterMolar concentration, or pKdCalorimetry, surface plasmon resonance, binding assays
Binding free energy (ΔG)The affinity as an energy, ΔG = RT ln Kd; more negative means tighterkcal/mol or kJ/molConverted from Kd, or read from calorimetry
Binding energy, interaction energyForce-field energy of the complex minus its parts, with or without solvationkcal/molMinimization, MM/PBSA, MM/GBSA
Docking scoreValue of a scoring function for one posekcal/mol for some programs, arbitrary units for othersDocking programs
Predicted affinityA model's estimate of pK or log IC50 for a predicted complexpK or log unitsMachine-learning scoring such as GNINA and Boltz-2
Table 1. Quantities that are called binding affinity, binding energy or docking score. The first two describe the same equilibrium. The others estimate it, with different inputs and different omissions.

Binding affinity and the dissociation constant

Affinity is a statement about equilibrium. For a protein P and ligand L that form a complex PL, the dissociation constant is Kd = [P][L]/[PL], the free ligand concentration at which half the protein is occupied.[16] A smaller Kd means tighter binding, so a 1 nM ligand is a thousand times stronger than a 1 µM one. Kd also belongs to the complex, not to the experiment, unlike an IC50, which shifts with substrate and enzyme concentration.

Affinity is often quoted as a negative logarithm, pKd = −log10 Kd in molar units, so that bigger numbers mean tighter binding. A Kd of 10 nM is a pKd of 8, and pK is the scale that machine-learning tools such as GNINA predict on.

Affinity is also conditional. A Kd belongs to one protein construct, one buffer and one temperature, and it describes one ligand binding one site. Two labs measuring the same complex under different conditions commonly differ by a few-fold, a spread the error section below returns to.[4]

Converting ΔG to Kd and back

Kd and ΔG are the same information on different scales. The standard binding free energy is

ΔG∘=RTln⁡Kdc∘⟺Kd=c∘ eΔG∘/RT\Delta G^\circ = RT \ln \frac{K_d}{c^\circ} \qquad\Longleftrightarrow\qquad K_d = c^\circ \, e^{\Delta G^\circ / RT}

where R is the gas constant, T the absolute temperature and c° the standard concentration of 1 M.[1] At 25 °C, RT is 0.592 kcal/mol (2.48 kJ/mol), so every tenfold change in Kd moves ΔG by 1.364 kcal/mol (5.71 kJ/mol).

Table 2 shows the conversion at 25 °C. The scale is steep: a docking score that is off by 2.4 kcal/mol is off by a factor of about 57 in Kd.

ΔG (kcal/mol)ΔG (kJ/mol)Kd
−4−16.71.2 mM
−6−25.140 µM
−8−33.51.4 µM
−10−41.847 nM
−12−50.21.6 nM
−14−58.655 pM
−16−66.91.9 pM
−18−75.364 fM
Table 2. Binding free energy and the dissociation constant at 25 °C. We calculated each Kd with Kd = exp(ΔG/RT), RT = 0.5925 kcal/mol, and a 1 M standard state.

Imatinib shows the conversion in practice. Its apparent Kd of 106 nM gives ΔG = 0.5925 × ln(106 × 10−9) = −9.52 kcal/mol.[11] Run it backwards and the Vina score of −12.51 for the same compound would mean a Kd of 0.68 nM, about 150 times tighter than the measurement. That gap is why a docking score should never be converted and reported as an affinity.

Most wrong conversions come from units and inputs. Putting kJ/mol into a formula built for kcal/mol is a factor of 4.184, and the temperature matters little, since 37 °C changes RT by only 4%. The bigger trap is the input: an IC50 is not a Kd, and the IC50 to Ki calculator applies the assay correction first. Femtomolar values like the last row of Table 2 are usually kinetic or displacement estimates with large uncertainty.

What binding energy means in each field

"Binding energy" has three common meanings, and which one applies depends on who is writing.

In drug design and simulation

In papers about docking and molecular dynamics, binding energy usually means an energy difference between the bound complex and the separated partners, in kcal/mol. When the authors have included solvation and entropy it is a binding free energy, and when they have not it is an interaction energy. Most MM/PBSA and MM/GBSA studies leave the entropy term out, because it is costly and noisy, so their values are not free energies.[6]

The difference is large. In the avidin calculations above, the MM/PBSA values of −35 to −53 kcal/mol for charged biotin analogues lie far past the −20.4 kcal/mol that the best of seven measured avidin ligands reaches.[6] Such numbers can rank related ligands on one target, but they are not affinities. Much of the excess is the entropy term that MM/PBSA and MM/GBSA studies usually leave out.

In chemistry

For a chemist, binding energy is the energy needed to break something apart, usually written as a bond energy or bond dissociation energy. Splitting the C-C bond of ethane takes about 90 kcal/mol, and splitting H2 takes about 104 kcal/mol.[9] Both are enthalpies of breaking a covalent bond in the gas phase, a few electronvolts per bond.

A drug sticking to a protein is a far smaller event. A ΔG of −10 kcal/mol is a ninth of the C-C bond energy, and it is only the net of many weak contacts (hydrogen bonds, van der Waals contacts and buried hydrophobic surface) after paying to strip water from both partners.

In physics

A physicist means nuclear binding energy: the energy released when separate protons and neutrons form a nucleus, equal to the mass defect times c². Helium-4 is bound by about 28.3 MeV in total, and the curve of binding energy per nucleon peaks near iron-56 at 8.79 MeV and nickel-62 at 8.7945 MeV, the highest of any known nuclide.[7] The same word also describes the energy that holds an electron in an atom: 13.6 eV for hydrogen.[8]

Table 3 puts the scales side by side, using 1 eV per particle = 96.485 kJ/mol. A helium nucleus binds about 65 million times more energy than a drug does per complex, which is why binding energy in nuclear physics and in drug design never meet in the same equation.

What is boundEnergy per particleEnergy per mole
Drug in a protein pocket (ΔG)0.43 eV41.8 kJ/mol
C-C bond in ethane3.9 eV377 kJ/mol
H-H bond in H24.5 eV436 kJ/mol
Electron in a hydrogen atom13.6 eV1,312 kJ/mol
Helium-4 nucleus28.3 MeV2.7 billion kJ/mol
Table 3. Binding energy at five scales. Values per particle and per mole use 1 eV = 96.485 kJ/mol. The drug value is a ΔG of −10 kcal/mol, a Kd of 47 nM at 25 °C. Sources: Blanksby and Ellison (2003), NIST CODATA 2022, Wang et al. (2021).[9][8][7]

Why a binding energy is not a binding free energy

Affinity responds to the free energy, which is the energy change minus a temperature-weighted entropy term, ΔG = ΔH − TΔS. Binding between a protein and a ligand gives off heat and, at the same time, costs the molecules freedom of motion. The two terms are large and nearly cancel, so the small difference between them is what sets Kd.

Streptavidin and biotin show it at the extreme. Calorimetry on wild-type streptavidin at 300 K gives an enthalpy of −29.4 kcal/mol, an entropy cost of +11.4 kcal/mol and a net ΔG of −18.0 kcal/mol, which corresponds to a Kd of about 8 × 10−14 M by our calculation. Kinetic measurements in the same study put the Kd of biotin at 6.5 × 10−14 M.[5]

TermValue (kcal/mol)What it represents
ΔH−29.4Heat released: contacts formed minus contacts broken
−TΔS+11.4Cost of freezing motion and ordering the complex
ΔG−18.0Net free energy; sets Kd
Table 4. Thermodynamics of biotin binding to wild-type streptavidin. Isothermal titration calorimetry at 300 K. The printed table headers show the signs reversed; we give the signed values. ΔG = ΔH − TΔS. Source: Magalhães et al. (2011), Table II.[5]

A calculation that captured only the enthalpy would overshoot the free energy by more than half. That is the pattern in MM/PBSA output, and it is why interaction energies run well past measured ΔG values. The drawing at the top of this page shows the three steps to scale.

What a docking score is

A docking program searches for ligand poses in a receptor and ranks them with a scoring function. Scoring functions are fast approximations, summing simple terms for contacts, hydrogen bonds, clashes and sometimes flexibility. They look at one static structure, so they leave out most of what separates ΔH from ΔG: all but a rotatable-bond penalty for entropy, the cost of moving water and the energy of relaxing the protein around the ligand.

Some programs fit their function to measured affinities so the output lands on a kcal/mol scale. AutoDock Vina does this, and the Vina manual calls its score a predicted binding affinity.[12] Others do not. GOLD's fitness functions, including ChemPLP, are dimensionless, with higher meaning better.[13] The unit on the page tells you little about whether the number is an energy.

Figure 1. A docking score as it appears in a ProteinIQ job. The public AutoDock Vina c-Abl batch example (PDB 1IEP, imatinib, nilotinib and ponatinib) lists each pose's Vina score as Affinity in kcal/mol. The panel under the table splits the selected pose's value into inter energy and intra energy, so the headline score is built from force-field-like terms, not read from an assay.

In Figure 1 the Affinity column is the score, and the energy terms underneath it show what the score is made of. Other tools report on other scales, and some count better in the opposite direction (Table 5).

ToolValue reportedUnitBetter is
AutoDock VinaVina score per posekcal/molLower
GOLD (ChemPLP)Fitness scoreDimensionlessHigher
GNINACNNaffinity, a predicted pKpK (−log10 of molar Kd)Higher
Boltz-2affinity_pred_value, log10 of an IC50 in µMLog unitsLower
PRODIGYΔG and Kd for protein-protein complexes from interface contactskcal/mol and MLower ΔG, lower Kd
gmx_MMPBSAEnd-point binding energy from simulation snapshotskcal/molLower
Table 5. What common tools report and which direction is better. Units and meaning differ, so scores from different tools cannot be compared. Sources: AutoDock Vina manual, CCDC GOLD documentation, Boltz documentation, Vangone and Bonvin (2015).[12][13][14][15]

Converting a predicted pK is simple arithmetic: a CNNaffinity of 7 is a nominal Kd of 100 nM and a ΔG of −9.5 kcal/mol. The arithmetic adds no accuracy, though. A CNNaffinity typically misses by about 1.4 pK units, and Boltz-2 predicts a log IC50, which needs an assay correction before it means anything like a Kd.

How close a docking score gets to the measured affinity

Binding free energies have a noise floor even when measured. Ross and colleagues compared compounds that were tested in more than one assay and found a pairwise RMSE of 0.91 kcal/mol between assays of the same molecules, against 0.23 kcal/mol for repeats within one assay.[4] In the same study, FEP+ free energy perturbation calculations on 1,237 ligands reached a pairwise RMSE of 1.25 kcal/mol, not far above the disagreement between two assays.[4]

Docking is in a different league. Trott and Olson reported a standard error of 2.85 kcal/mol between Vina scores and experimental free energies for 190 complexes.[2] On the CASF-2016 benchmark of 285 complexes, Vina correlated with measured affinity at R = 0.604, with a standard deviation of 1.73 pK units, equal to about 2.4 kcal/mol.[3] A 2.4 kcal/mol error is a factor of 54 to 57 in Kd.

Figure 2. Typical error in binding free energy, from experiment to docking. Bars show kcal/mol. The label gives the same error as a factor in Kd, 10 raised to the error divided by 1.364. The Vina CASF-2016 value converts the 1.73 pK standard deviation. Sources: Ross et al. (2023), Su et al. (2019), Trott and Olson (2010). Reuse under CC BY 4.0.

The consequence for any conversion is that a Vina score of −10, or 47 nM, is consistent with a Kd from about 1 nM to 2.5 µM. Docking is good at telling a plausible binder from an implausible one and at proposing poses. It is a poor instrument for ordering potent binders: for the three c-Abl inhibitors in Figure 1, Vina's ranking is the reverse of the measured one. Rigorous methods such as free energy perturbation cost far more compute and still reach only about 1 kcal/mol.

What is a good docking score

No score is good in itself. The same −8 kcal/mol is strong for a fragment and weak for a drug-sized molecule, because every extra heavy atom in contact adds favorable terms. A common way to remove that size effect is ligand efficiency, the free energy divided by the number of heavy atoms. About 0.3 kcal/mol per heavy atom is the floor often proposed for drug candidates.[10]

Heavy atomsΔG at 0.3 per atomEquivalent Kd
15−4.5 kcal/mol500 µM
25−7.5 kcal/mol3.2 µM
35−10.5 kcal/mol20 nM
40−12.0 kcal/mol1.6 nM
Table 6. What a ligand efficiency of 0.3 kcal/mol per heavy atom implies. We calculated ΔG = −0.3 × heavy atoms and converted to Kd at 25 °C. Heavy-atom threshold from Hopkins et al. (2014).[10]

Read the table as a way to judge a docking score against the size of the ligand. A 25-atom compound scoring −7.5 is as efficient as a 40-atom compound scoring −12, so the second is not the better lead. Experimentally, the strongest binders gain about 1.5 kcal/mol per heavy atom up to roughly 15 atoms and then level off.[17] Docking scores keep rising with size, which is why large ligands score better whether or not they bind better.

For a threshold you can defend, use your own target. Redock the co-crystallized ligand and use its score as the bar, dock known actives against decoys and choose the cutoff that separates them, or take the top fraction of a screen. Each cutoff holds only for the docking program and settings that produced it.

How to name each number

Most overstated results come from a label, not a calculation. Table 7 gives the name that matches what was actually computed.

What you haveCall itAvoid
Vina score of −9.1Vina docking score, −9.1 kcal/molBinding energy, affinity or ΔG
MM/GBSA value without entropyMM/GBSA interaction energy, entropy not includedBinding free energy
GNINA CNNaffinity of 7.2Predicted pK, nominal Kd of 63 nMMeasured or experimental Kd
Calorimetry ΔG and KdMeasured binding free energy and affinity, with temperature and bufferAnything that omits the conditions
IC50 from an activity assayPotency in that assay, or Ki after correctionKd
Table 7. What to call each kind of number.

Docking and binding tools on ProteinIQ

Dock, check the pose, then rescore or simulate, in that order, before treating any number as an affinity. That is the usual path from a molecular docking run to a binding free energy estimate.

TaskTool
Dock a ligand and get a score in kcal/molAutoDock Vina, smina
Rescore poses with a neural networkGNINA
Check that a pose is physically plausiblePoseBusters, PLIP
Measure pose distance from a crystal ligandRMSD Calculator
Estimate ΔG and Kd for a protein-protein complexPRODIGY
Score simulation snapshots with end-point energiesgmx_MMPBSA
Predict a complex and its affinityBoltz-2
Convert IC50 to Ki and pIC50IC50 to Ki calculator
Table 8. Tools for scoring, checking and converting binding results.

Frequently asked questions

What is a good docking score?

There is no universal threshold. A good score is one that beats the score of a known ligand docked into the same receptor with the same settings, after allowing for ligand size. Ligand efficiency of about 0.3 kcal/mol per heavy atom is a useful floor for drug-sized molecules.[10]

Is binding energy the same as binding affinity?

No. Binding affinity is how tightly a ligand binds, measured as Kd. Binding free energy is that same quantity as an energy. Binding energy is a looser phrase that often means a force-field interaction energy with no entropy term, and it does not convert to a Kd.[6]

How do I convert ΔG to Kd?

Use Kd = exp(ΔG/RT) with a 1 M standard state. At 25 °C, RT is 0.592 kcal/mol, so a ΔG of −10 kcal/mol is a Kd of 47 nM.[1] Rearranged, Kd = 10^(ΔG/1.364) M at 25 °C with ΔG in kcal/mol.

How do I convert Kd to ΔG?

Use ΔG = RT ln Kd with Kd in molar units. A Kd of 10 nM gives 0.592 × ln(10−8) = −10.9 kcal/mol at 25 °C. Each tenfold change in Kd shifts ΔG by 1.364 kcal/mol.

Why is my docking score more negative than the measured ΔG?

Docking scores add favorable contacts for every heavy atom and model little of the entropy and desolvation cost, so large ligands tend to score better than they bind. Imatinib scores −12.51 in Vina against a measured −9.52 kcal/mol.[11]

Does a more negative score always mean stronger binding?

Only within one method, one receptor and one protocol, and only when the difference is larger than the method's error. Differences below about 2 kcal/mol between compounds are within the typical error of docking.[3]