---
title: How MM-GBSA results compare to the bench
description: 'What an MM-GBSA binding energy has been shown to agree with experimentally, and why this implementation returns no binding free energy at all.'
tags:
  - mmgbsa
  - validation
  - translatability
---

**You have an MM-GBSA result. Here is what it supports.**

- **This implementation returns no binding free energy.** The job performs a single-point OpenMM evaluation of the minimized complex. It does not compute ΔG_bind = E_complex − E_receptor − E_ligand, so `delta_g_bind` is always `null` and `binding_affinity_available` is `false`. What you get are the complex's own energies (`complex_potential_energy_kcal`, `nonbonded_energy_kcal`, the Generalized Born term) plus a `caveat` string. See [interpreting MM-GBSA outputs](/docs/mmgbsa/outputs).
- **So there is no affinity number on this page to grade.** Everything below describes what MM-GBSA as a *method* has been shown to achieve in the literature. None of those figures describes what this tool currently emits, and you should not transfer them to it.
- **Where the method does produce a ΔG, it ranks better than it measures.** Published correlations against measured affinity run from r ≈ 0.9 to *negative* depending on the target, and absolute errors of 4–21 kcal/mol are routine. It is a re-ranking tool, not an affinity assay.
- **Do not convert an MM-GBSA ΔG to a Kd.** The arithmetic is exact, the input is not. See [the Kd question](#the-kd-conversion-is-not-defensible).
- **Confirm affinity with** a binding measurement — ITC, SPR, or an enzymatic IC50/Ki. Nothing computational substitutes for one.

Jump to [agreement with wet lab](#agreement-with-wet-lab) or [against other in silico methods](#against-other-in-silico-methods). The general framing for every tool is in [reading an in silico result against the bench](/docs/translatability).

## How in silico is used currently

Affinity prediction is used to **re-order a shortlist**, not to decide potency. In a live program that means three things:

- **Post-docking rescoring.** You have tens or hundreds of poses from a docking run and an empirical score that correlates poorly with potency. An end-point free-energy method is applied to the survivors to reshuffle them before anyone buys or synthesizes. This is MM-GBSA's dominant published use, and it is the use the platform wires it into ([DiffDock](/docs/diffdock) → rescore).
- **Congeneric-series ranking.** Within one target and one chemical series, where systematic errors partly cancel, the method is used to order analogs. Outside a series — comparing unrelated scaffolds, or comparing across targets — the cancellation does not happen and the published correlations fall apart.
- **Mechanistic reasoning.** Decomposing an interaction energy into van der Waals, electrostatic and solvation parts to argue about *why* a ligand binds, and which residues to mutate. Note that this platform does not produce a per-residue decomposition: `include_per_residue` is rejected rather than silently under-delivered, because there is no ΔG_bind to decompose.

What MM-GBSA is not used for, by anyone reporting their validation honestly, is producing an absolute binding affinity.

## Against other in silico methods

The affinity accuracy ladder, with each rung's inputs stated:

1. **Empirical docking score** — a fitted function evaluated on one pose. Fast, and a poor affinity ranker; MM/GBSA rescoring improved AutoDock's screening and ranking power on ABL, ALK and BRAF, which is the standard argument for running it at all.
2. **MM-GBSA / MM-PBSA** — a force-field interaction energy plus an implicit-solvent term, from one minimized structure or an MD ensemble. Correlations as above: usable for intra-series ranking, unusable for absolute affinity.
3. **Alchemical free energy (FEP, TI)** — a physically rigorous path between two ligands, requiring explicit solvent, extensive MD sampling, and hours to days per perturbation. Pairwise RMSE **1.25 kcal/mol** against experiment, approaching the **0.91 kcal/mol** reproducibility of the experiments themselves. Independently, non-equilibrium TI and a commercial FEP implementation both reached an AUE of ~**0.87 kcal/mol** over 482 perturbations across 13 protein-ligand datasets, with 66–68% of estimates within 1 kcal/mol.
4. **Experiment** — 0.91 kcal/mol between independent assays, 0.23 kcal/mol between repeats of one assay. Even this is not a hard floor: one surveyed comparison, SPR against mass spectrometry on carbonic anhydrase I ligands, disagreed by more than 2.5 kcal/mol, which is worse than any individual FEP result in that benchmark.

⚠︎ **This platform does not run alchemical FEP, and `fep_estimate` is not it.** The `fep_estimate` tool computes a relative ΔΔG heuristic from molecular descriptor deltas (`heuristic_descriptor_delta_v1`). It takes SMILES rather than poses and runs no molecular dynamics at all. It must never be described as, compared to, or allowed to inherit the accuracy figures of rung 3. The honest ordering for what is actually available here is: `fep_estimate` (descriptor heuristic) and the docking score at the bottom, MM-GBSA above them *when it produces a ΔG*, and nothing on rung 3.

Within MM-GBSA itself, two comparisons are worth knowing. **GB model choice is not cosmetic:** GB_OBC1 beat GB_OBC2 on four of six systems in the study above, and GB_OBC2 produced negative correlations on two of them. **Ensemble averaging beats a single minimized structure** where it has been tested head to head (R² 0.36 → 0.69 on six antithrombin ligands) — though the large PDBbind evaluation found MD largely unnecessary for MM/GBSA specifically, so the two findings genuinely disagree and the resolution is probably system-dependent.

> Ross GA, Lu C, Scarabelli G, Albanese SK, Houang E, Abel R, Harder ED, Wang L. The maximal and current accuracy of rigorous protein-ligand binding free energy calculations. *Commun Chem*. 2023;6(1):222. doi: [10.1038/s42004-023-01019-9](https://doi.org/10.1038/s42004-023-01019-9)
>
> Gapsys V, Pérez-Benito L, Aldeghi M, Seeliger D, van Vlijmen H, Tresadern G, de Groot BL. Large scale relative protein ligand binding affinities using non-equilibrium alchemy. *Chem Sci*. 2019;11(4):1140-1152. doi: [10.1039/c9sc03754c](https://doi.org/10.1039/c9sc03754c)
>
> Sun H, Li Y, Shen M, Tian S, Xu L, Pan P, Guan Y, Hou T. Assessing the performance of MM/PBSA and MM/GBSA methods. 5. Improved docking performance using high solute dielectric constant MM/GBSA and MM/PBSA rescoring. *Phys Chem Chem Phys*. 2014;16(40):22035-22045. doi: [10.1039/c4cp03179b](https://doi.org/10.1039/c4cp03179b)

## In silico vs. wet lab

The standard wet-lab approach for the question MM-GBSA addresses is a **direct binding measurement** — isothermal titration calorimetry (ITC) or surface plasmon resonance (SPR) for a dissociation constant, or an enzymatic IC50/Ki where the target has a catalytic readout. These give a number in molar units that converts to a free energy; they are the thing every affinity prediction is graded against.

Unlike the pose question, agreement here has no single settled threshold. The field reports two metrics: **correlation** (Pearson r or Spearman ρ against measured affinity, for ranking) and **error in kcal/mol** (RMSE, MUE, or unsigned mean error, for magnitude). They answer different questions and a method can be good at one and useless at the other — which is precisely MM-GBSA's situation.

| | Time to answer | Cost basis | Accuracy, and its metric |
|---|---|---|---|
| **Standard wet lab** — ITC, SPR, or enzymatic IC50/Ki | Days to weeks per series, once protein and assay exist; months to get there | SPR $357.50/day (about 17,800 runs here per day) at a core facility's industry tier; ITC and MST have no public price of any kind. [cost basis](/docs/translatability#what-the-wet-lab-cost-figures-mean) | Ground truth. Two independent assays of the same compound series agree to a pairwise RMSE of **0.91 kcal/mol** (95% CI 0.83–1.11) over 1,867 compounds; repeats of a single assay, **0.23 kcal/mol** (350 compounds, 15 assays) |
| **Conventional in silico** — empirical docking score | Seconds to minutes per pose, CPU | No per-run price is published for this class. [cost basis](/docs/translatability#what-the-wet-lab-cost-figures-mean) | Poor as an affinity ranker; see the confidence-score figures on [the DiffDock page](/docs/diffdock/translatability#against-other-in-silico-methods). MM/GBSA rescoring measurably improved AutoDock's screening and ranking power on ABL, ALK and BRAF |
| **SOTA in silico — not run here** — alchemical free energy (FEP, TI) | Hours to days per perturbation, GPU | No per-run price is published for this class. [cost basis](/docs/translatability#what-the-wet-lab-cost-figures-mean) | Pairwise RMSE **1.25 kcal/mol** (95% CI 1.17–1.33) against experiment; AUE **3.64 ± 0.14 kJ/mol** (0.87 kcal/mol) over 482 perturbations across 13 datasets, with 66–68% of estimates within 1 kcal/mol |
| **Combined tooling here** — [DiffDock-L](/docs/diffdock) → MM-GBSA | 3–35 min for a pose, plus 1–5 min per rescored complex | 5 credits | **No affinity figure exists.** DiffDock-L reaches 50% of poses ≤ 2 Å on PoseBusters; the MM-GBSA step then returns no ΔG_bind, so the chain produces a geometry with no graded affinity attached to it |

**Read the accuracy column by cell, never down the column.** The wet-lab cell is a reproducibility figure — the floor below which no prediction can be graded. The alchemical cell is an error against that ground truth. The conventional cell is a ranking claim. The last cell is an admission.

On that reading, the honest position for this chain is: **minutes instead of weeks, at an affinity accuracy that is currently unmeasurable, because the affinity step does not emit the quantity it names.** The pose half of the chain is real and graded; the affinity half is not.

Two asymmetries matter beyond the table. A binding measurement needs purified, behaving protein and a compound in hand — which is most of why computational affinity prediction exists. And an assay reports what *is*, at a stated buffer and temperature, while MM-GBSA reports a force-field energy under an implicit-solvent approximation. The two are not interchangeable at any correlation coefficient.

## Agreement with wet lab

**This section grades the method, not this tool.** Every figure below comes from an implementation that computed a real ΔG_bind by partitioning the complex into receptor and ligand and parameterizing the small molecule. The platform's implementation does neither. The figures are here because they are the correct prior for what MM-GBSA can do, and because they bound what a future implementation here could claim.

Read every one of them with its **n and its target**, because target dependence is the dominant effect in this literature.

**On 59 ligands across six proteins** (α-thrombin n=7, avidin n=7, cytochrome C peroxidase n=18, neuraminidase n=8, P450cam n=12, penicillopepsin n=7), with a single-trajectory protocol and normal-mode entropy:

- **Absolute binding free energy is not usable.** MM/GBSA unsigned mean errors against experiment were 21.31 kcal/mol (α-thrombin), 10.58 (penicillopepsin), 9.30 (avidin), 6.07 (P450cam), 5.62 (cytochrome C peroxidase) and 4.42 (neuraminidase). Not one system came under 4 kcal/mol.
- **Ranking correlation depends on the GB model and the target.** With GB_OBC2 — the model this platform defaults to — at an interior dielectric of 1, Pearson r against measured affinity was 0.92 (avidin), 0.73 (α-thrombin), 0.35 (cytochrome C peroxidase), 0.23 (penicillopepsin), **−0.16** (neuraminidase) and **−0.42** (P450cam). Two of six targets ranked worse than chance.
- **GB_OBC1 did better than GB_OBC2** on four of the six systems and was the best of the three GB models overall in that study. The platform's default is the one that lost.
- **A single outlier can carry the whole correlation.** The study's own worked example: MM/PBSA on α-thrombin gave r = 0.81, which fell to **r = 0.01** when one ligand (benzamidine) was removed. With n = 7, a correlation coefficient is close to meaningless.

**On 1,864 PDBbind crystal structures**, MM/GBSA on minimized structures reached an overall Pearson rp of **0.579 ± 0.002**. That headline is biased by over-represented protein families: on the unbiased one-protein-family/one-binding-ligand sampling it falls to **rp = 0.408 ± 0.006** at an interior dielectric of 4, and averaged over fold families the correlation score was **0.516 ± 0.047** with a ranking score of **0.463 ± 0.047**. Per-family rp spanned **0 to 0.9**. Ligand charge mattered strongly: rp fell from **0.621 ± 0.003** for neutral ligands to **0.125 ± 0.142** for ligands carrying a formal charge of 5. The same study found MD sampling largely unnecessary for MM/GBSA specifically (rp 0.579 minimized vs 0.564 after 1 ns).

**On 855 diverse protein-ligand complexes** with an automated MM/GBSA implementation, R² between calculated and experimental affinities was **0.63** — but only "after carefully removing flawed structures," and the paper attributes its outliers to exactly the approximations the method makes: neglect of explicit solvent, ligand strain, and entropy. A separate study on *Plasmodium falciparum* DHFR inhibitors reported that correlations from a **single energy-minimized structure** were similar to those from averaging over MD snapshots with explicit water — relevant here, because single-structure minimization is what this platform does.

**On six antithrombin ligands**, single-structure MM/GBSA rescoring gave **R² = 0.36** against affinities derived from measured dissociation constants, rising to **0.69** with ensemble averaging over 20 MD simulations, and to **0.85** after removing one outlier. At n = 6 all three numbers are fragile; the useful content is the direction — ensemble averaging roughly doubled the explained variance over the single-structure protocol.

> Hou T, Wang J, Li Y, Wang W. Assessing the performance of the MM/PBSA and MM/GBSA methods. 1. The accuracy of binding free energy calculations based on molecular dynamics simulations. *J Chem Inf Model*. 2011;51(1):69-82. doi: [10.1021/ci100275a](https://doi.org/10.1021/ci100275a)
>
> Sun H, Li Y, Tian S, Xu L, Hou T. Assessing the performance of MM/PBSA and MM/GBSA methods. 4. Accuracies of MM/PBSA and MM/GBSA methodologies evaluated by various simulation protocols using PDBbind data set. *Phys Chem Chem Phys*. 2014;16(31):16719-16729. doi: [10.1039/c4cp01388c](https://doi.org/10.1039/c4cp01388c)
>
> Greenidge PA, Kramer C, Mozziconacci JC, Wolf RM. MM/GBSA binding energy prediction on the PDBbind data set: successes, failures, and directions for further improvement. *J Chem Inf Model*. 2013;53(1):201-209. doi: [10.1021/ci300425v](https://doi.org/10.1021/ci300425v)
>
> Rastelli G, Del Rio A, Degliesposti G, Sgobba M. Fast and accurate predictions of binding free energies using MM-PBSA and MM-GBSA. *J Comput Chem*. 2010;31(4):797-810. doi: [10.1002/jcc.21372](https://doi.org/10.1002/jcc.21372)
>
> Zhang X, Perez-Sanchez H, Lightstone FC. A comprehensive docking and MM/GBSA rescoring study of ligand recognition upon binding antithrombin. *Curr Top Med Chem*. 2017;17(14):1631-1639. doi: [10.2174/1568026616666161117112604](https://doi.org/10.2174/1568026616666161117112604)

### The Kd conversion is not defensible

The existing [outputs page](/docs/mmgbsa/outputs) has historically described a canvas output that maps a binding energy to "very strong / strong / moderate / weak" with an estimated Kd in nanomolar. That mapping is the Boltzmann relation Kd = exp(ΔG/RT), and at 298 K (RT = 0.5925 kcal/mol) its arithmetic checks out: −12 kcal/mol gives 1.6 nM, −9 gives 250 nM, −6 gives 40 µM, −3 gives 6.3 mM. The arithmetic is not the problem.

The problem is that exponentiation **amplifies the input error geometrically**. A 1 kcal/mol error in ΔG is a 5.4-fold error in Kd; 2 kcal/mol is 29-fold; 3 kcal/mol is 158-fold. Against the verified absolute errors above, the *best* of six systems — neuraminidase at 4.42 kcal/mol unsigned mean error — implies a Kd wrong by a factor of roughly 1.7 × 10³. The worst implies a factor of about 10¹⁵. A "very strong, Kd < 1 nM" verdict built on that input carries no information about whether the compound binds at 1 nM or 1 mM.

And on this platform the question is moot in the worse direction: there is no ΔG_bind to exponentiate. **Nothing in the running code emits an `estimated_kd` or a binding-strength label** — the classifier was removed from the backend precisely because it had been fed the complex's raw potential energy and consequently returned "very strong / drug-like" for everything. The documented Kd column described output the tool does not produce. It has been removed from the docs rather than left standing, and the verdict here is: **do not reintroduce a ΔG → Kd conversion, even if a real ΔG_bind decomposition is implemented later.** Report the energy with its units and let the reader keep the error bars.

## What the figure doesn't cover

- **The largest gap is the implementation, not the literature.** Every correlation on this page was produced by code that partitions the complex and parameterizes the ligand. This app evaluates the whole complex as one system with a protein force field. Transferring any figure above to a result from this tool would be the central error this page exists to prevent.
- **Sample sizes are small and target dependence is large.** Per-target n in the best-documented study ranges from 7 to 18. Per-family Pearson r across 1,864 PDBbind structures spans 0 to 0.9. A single correlation coefficient for "MM-GBSA accuracy" does not exist, and quoting one without its target and n is the most common error in reading this literature.
- **The dielectric constant is tuned after the fact.** In the published evaluations the interior dielectric (1, 2 or 4) that gave the best correlation differed per system, and no universal value worked. A retrospective best-case correlation chosen over three dielectrics is not a prospective accuracy.
- **Charged ligands fail.** Correlation collapsed from 0.621 to 0.125 as formal charge rose from 0 to +5. If your series carries formal charge, the headline figures do not describe it.
- **Entropy is omitted, and adding it back is not obviously an improvement.** Standard MM-GBSA drops the configurational entropy term. In a study of more than 1,500 protein-ligand systems, adding a truncated normal-mode entropy to **minimized** structures *degraded* overall accuracy relative to the enthalpy alone; it helped only on MD trajectories. So "entropy is neglected" is a real approximation, but it is not a defect with a cheap fix.
- **Metals and water are outside the model.** The two systems where GB_OBC2 gave negative correlations above — cytochrome C peroxidase and P450cam — both have iron in the binding site, and the authors attribute the failure to missing GB parameters for Fe. Implicit solvent also cannot represent a structurally bound water that mediates the interaction.
- **Garbage in, confident number out.** MM-GBSA scores the geometry it is handed. The upstream pose is roughly half right: DiffDock-L reaches 50% ≤ 2 Å on PoseBusters, and [that page](/docs/diffdock/translatability) documents how much lower that gets on novel targets. A physics-based energy computed on a wrong pose is a precise answer to the wrong question, and nothing in the output flags it. Inspect the pose before you read the energy.
- **No prospective validation exists for this chain.** No published study has run a docking-plus-MM-GBSA pipeline on this platform's configuration, nominated compounds, and reported the measured affinities. Campaign hit rates from multi-tool pipelines belong to those pipelines.
- **Reviews are not evidence here.** MM-GBSA has a very large review and protocol literature, and much of it restates a "typical r = 0.5–0.7" range whose provenance dissolves on inspection. Every figure on this page is from a primary comparative study, and the ones taken from an abstract rather than a table are marked as such above.

## Go / no-go

**Go — act on the result when:**

- You are reading the energy decomposition as a *mechanistic* description of a complex you already believe in — which contacts dominate, whether burial or electrostatics drives the interaction.
- You are confirming that a minimized complex is physically sane before taking it into a longer simulation. The minimized PDB is a real, useful artifact of this job.
- You have inspected the upstream pose and are using the rescoring step as one input among several, in the consensus pattern published campaigns actually used.
- (For a future implementation that emits a real ΔG_bind) you are ranking a **congeneric series against one target**, on more than a handful of compounds, with no formal charges and no metal in the site — the conditions under which the published correlations are strongest.

**No-go — do not act on the result when:**

- You want a binding affinity, a Kd, or a "strong binder" verdict. This tool does not produce one, and the method would not support converting one if it did.
- You are comparing across targets, or across unrelated scaffolds. Systematic error cancellation is the whole reason intra-series ranking works, and it does not apply.
- Your target has a metal in the binding site, or your ligands carry formal charges, or a structural water mediates the key contact.
- The upstream pose has not been checked. Compounding a 50%-accurate pose with a method that cannot tell you the geometry is wrong gets you a confident number on the wrong structure.
- You need a go/no-go on potency. Run ITC, SPR, or the enzymatic assay. That is the measurement; everything on this page is a way of deciding what to measure first.

**Caveats on reading this page:** every figure above is quoted from the cited source with its original metric and sample. Figures from different studies are not directly comparable — test sets, protocols, dielectric constants, entropy treatments and GB models all differ between them. Figures marked with a review comment were read from the published abstract rather than the paper's own table, and have not been verified against primary text.

## Read next