---
title: Direct Binding Free Energy (DBFE)
url: https://www.emergentmind.com/topics/direct-binding-free-energy-dbfe
type: topic
---

# Direct Binding Free Energy (DBFE)

Searching arXiv for the cited works to ground the article in published sources.
Direct Binding Free Energy (DBFE) denotes the absolute thermodynamic free energy change associated with forming a bound complex from its unbound components. Across the literature, the underlying quantity is usually the standard binding free energy, written as $\Delta G^\circ_{\mathrm{bind}}$ or $\Delta G$, and related to equilibrium association and dissociation constants by standard statistical thermodynamics. The label “DBFE” is not used uniformly: in some works it refers to direct end-state absolute binding free energy estimation without alchemical intermediates, in others to rigorous statistical-mechanical formulations of absolute binding, to path-based physical unbinding free energies, or to machine-learning prediction of experimentally measured $\Delta G$ values [2603.12253] [1208.4885] [2508.19632]. This suggests that DBFE is best understood as a family resemblance term: methods differ substantially, but they aim at the same absolute binding thermodynamics rather than the relative mutation quantity $\Delta\Delta G$.

## 1. Thermodynamic definition and scope

The common thermodynamic object is the standard free energy for association. For a receptor and ligand, one formulation is
$$
\Delta G^\circ_{\mathrm{bind}} = G^\circ_{\mathrm{aq}}(RL) - G^\circ_{\mathrm{aq}}(R) - G^\circ_{\mathrm{aq}}(L),
$$
while for protein-protein binding one paper writes
$$
\Delta G_{\mathrm{bind}} = G_{AB}^{\mathrm{bound}} - \left(G_A^{\mathrm{unbound}} + G_B^{\mathrm{unbound}}\right).
$$
Equivalent equilibrium relations are reported as
$$
\Delta G = -RT \ln K_a = RT \ln K_d
$$
and
$$
\Delta G^\circ_{\mathrm{bind}} = -k_B T \ln K_a.
$$
These expressions appear in antibody-antigen prediction, inverse-folding-based protein-protein binding analysis, and solution-phase free-energy treatments [2508.19632] [2410.09543] [1501.04428].

A central distinction in the literature is between absolute binding free energy and mutational free-energy differences. Relative mutational effects are defined as
$$
\Delta\Delta G = \Delta G_{\mathrm{mut}} - \Delta G_{\mathrm{wt}},
$$
or, in the protein-protein setting, as the difference between mutant and wild-type binding free energies. Several papers emphasize that DBFE corresponds to the absolute $\Delta G$ of a bound complex, whereas $\Delta\Delta G$ is a derived relative quantity more suitable for mutation scans and affinity optimization [2508.19632] [2410.09543].

Standard-state treatment is integral to the definition. Multiple sources write the correction in the form
$$
\Delta G_{\mathrm{std}} = -k_B T \ln\!\left(C^\circ V_{\mathrm{site}}\right),
$$
or equivalent variants involving translational and rotational phase-space factors such as $8\pi^2$ or an effective binding-region volume $\Omega$ [1904.11058] [1208.4885] [2603.12253]. In path-based formulations, omission of such terms is explicitly identified as the reason a reported free energy should be interpreted as qualitative or path-comparative rather than an absolute standard-state DBFE [2311.17925].

## 2. Rigorous statistical-mechanical formulations

A rigorous formulation is provided by implicit ligand theory (ILT), which separates receptor sampling from ligand sampling while preserving exact expressions for the standard binding free energy. ILT defines an effective interaction energy $\Psi(r_{RL})$ and a receptor-configuration-dependent binding potential of mean force $B(r_R)$, then expresses the standard binding free energy as
$$
\Delta G^\circ = -\beta^{-1}\ln \left< e^{-\beta B} \right>_{R}^{r_R} + \Delta G_\xi,
$$
with
$$
\Delta G_\xi = -\beta^{-1}\ln\!\left( \frac{\Omega\, C^\circ}{8\pi^2} \right).
$$
This is presented as a direct route to standard binding free energies from docking-like rigid-receptor calculations averaged over a receptor ensemble [1208.4885].

The same ILT perspective underlies later FFT-accelerated work, where the binding potential of mean force is estimated efficiently by evaluating rigid-body interaction energies over many ligand poses. For a fixed receptor conformation, the pose-dependent BPMF is written as
$$
B(q_R, R) = -k_B T \ln \left\langle e^{-\beta U_{\mathrm{int}}(q_R, q_L, R)} \right\rangle_{\mathrm{apo}\,L},
$$
and receptor-ensemble averaging yields the final standard binding free energy with the standard-state correction expressed through $V_{\mathrm{site}}$ [1708.07045]. In that study, the resulting standard binding free energies for T4 lysozyme ligands are reported to agree well with previous alchemical calculations, with Pearson $R \approx 0.87$–$0.90$ against a flexible-complex benchmark for 24 systems and $R \approx 0.82$ against an alternative rigid-receptor method for 141 ligands at fine grid spacing [1708.07045].

A conceptually different but still rigorous direct treatment appears in Monte Carlo integration. There, the gas-phase DBFE is obtained from a Boltzmann-weighted interaction enthalpy plus entropic penalties inferred from insertion statistics and standard-state volume restriction:
$$
\Delta G_{\mathrm{bind}}^{\mathrm{gas},\circ}
=
\frac{\sum \Delta E_i e^{-\beta \Delta E_i}}{\sum e^{-\beta \Delta E_i}}
-
k_B T \ln(N_{\mathrm{acc}}/N_{\mathrm{att}})
-
k_B T \ln(V_{\mathrm{integration}}/V^\circ).
$$
The method is described as direct because it does not require alchemical transformations, restraints, or pose predefinition, although the reported implementation uses a rigid receptor and approximate solvation corrections [1701.05073].

These statistical-mechanical treatments share a common principle: DBFE can be written as a free-energy difference between physical bound and unbound states, but practical estimators differ in how they generate overlap, define the binding region, and treat solvent and configurational entropy.

## 3. Direct end-state versus alchemical DBFE

One major use of the term DBFE refers specifically to end-state absolute binding free energy methods that avoid alchemical intermediates. “Binding Free Energies without Alchemy” presents a method explicitly named Direct Binding Free Energy (DBFE), formulated in implicit solvent and based on only three simulations per ligand: apo receptor, free ligand, and holo complex [2603.12253]. The method constructs a thermodynamic cycle with a decoupled unrestrained state, a decoupled restrained state, a clash-filtered decoupled state, and the coupled bound state:
$$
\Delta G_{\mathrm{bind}}^\circ
=
\Delta G_{0\rightarrow r}
+
\Delta G_{r\rightarrow rc}
+
\Delta G_{rc\rightarrow 1}.
$$
Its restraining potential combines a Gaussian translational term and a Bingham rotational term,
$$
U_\zeta(\zeta) = k_B T \left[\frac{1}{2}(t - \mu)^T \Sigma^{-1}(t - \mu) - q^T M Z M^T q \right],
$$
with an analytic standard-state correction and an entropic clash-filtering term
$$
\Delta G_{r\rightarrow rc} = -k_B T \ln \alpha.
$$
The coupled-versus-clash-filtered leg is estimated by MBAR [2603.12253].

In the host-guest benchmark reported there, DBFE in OBC2 implicit solvent achieves RMSE $4.1$ kcal/mol, Pearson $r = 0.58$, and Spearman $\rho = 0.56$, outperforming OBC2 double decoupling in correlation and strongly outperforming OBC2 MM/GBSA, while TIP3P double decoupling remains best overall. On a protein-ligand benchmark, DBFE gives RMSE $5.3$ kcal/mol, $r = 0.65$, and $\rho = 0.64$, which is better in RMSE than OBC2 double decoupling but below TIP3P explicit-solvent performance [2603.12253].

A related non-alchemical framework based on the energy representation (ER) theory of solution also computes binding free energies directly from endpoint ensembles. There the binding free energy is written as
$$
\Delta G = \Delta\mu_G^B - \Delta\mu_G^D + \Delta G_{\mathrm{corr}}^\circ,
$$
where the bound and dissociated solvation free energies are obtained from energy-distribution functionals rather than alchemical interpolation [2410.04423]. For aspirin binding to $\beta$-cyclodextrin, the ER-OR method reports $\Delta G^\circ = -5.2 \pm 0.1$ kcal mol$^{-1}$ for the primary complex and $-5.03 \pm 0.09$ kcal mol$^{-1}$ for the secondary complex, close to BAR values of $-4.2 \pm 0.2$ and $-4.1 \pm 0.2$ kcal mol$^{-1}$, respectively [2410.04423].

By contrast, alchemical DBFE methods estimate the same absolute quantity through nonphysical intermediate Hamiltonians. The dopamine D3 receptor study uses a single-decoupling method with a $\lambda$-dependent Hamiltonian
$$
U_{\lambda}(\mathbf{r})=U_{0}(\mathbf{r})+\lambda u(\mathbf{r}),
$$
plus a standard-state correction
$$
\Delta G_{\mathrm{t}}^{\circ}=-k_{B}T\ln C^{\circ}V_{\mathrm{site}}.
$$
Its central finding is that enclosed hydration effects must be embedded into the solvation model: without HSA-informed hydration terms, RMSE is $7.9$ kcal/mol and correlation is $r=-0.014$, whereas inclusion of enclosed hydration yields RMSE $1.2$ kcal/mol and $r=0.64$ [1904.11058].

This contrast clarifies a recurring ambiguity. In some publications, DBFE means “absolute binding free energy” irrespective of whether the estimator is direct or alchemical; in others, especially [2603.12253], it means an explicitly non-alchemical end-state method. The distinction is methodological, not thermodynamic.

## 4. Pathway-based and restraint-based direct calculations

A separate class of direct methods computes DBFE from physical binding or unbinding pathways. In the iMD-VR-FE framework, interactive molecular dynamics in virtual reality is used to generate ligand unbinding paths, which are then projected into a six-dimensional collective-variable space and profiled by umbrella sampling with WHAM [2311.17925]. The path collective variables are
$$
s(\mathbf{r}) = \frac{\sum_{i} t_i \exp\!\left[-\lambda\, d\!\left(\boldsymbol{\theta}(\mathbf{r}), \mathbf{z}(t_i)\right)\right]}{\sum_{i} \exp\!\left[-\lambda\, d\!\left(\boldsymbol{\theta}(\mathbf{r}), \mathbf{z}(t_i)\right)\right]}
$$
and
$$
z(\mathbf{r}) = -\frac{1}{\lambda} \ln \left[\sum_{i} \exp\!\left[-\lambda\, d\!\left(\boldsymbol{\theta}(\mathbf{r}), \mathbf{z}(t_i)\right)\right]\right],
$$
with a metric-corrected distance in CV space. The unbinding and binding free energies are then estimated from PMF end-state differences,
$$
\Delta G_{\text{unbind}} = W(s_{\text{unbound}}) - W(s_{\text{bound}}), \qquad
\Delta G_{\text{bind}} = -\Delta G_{\text{unbind}}.
$$
For benzamidine–trypsin, the reported path-based estimate is $\Delta G_{\mathrm{bind}} \approx -22.5$ kcal mol$^{-1}$, substantially more favorable than the experimental value of approximately $-6.2$ kcal mol$^{-1}$ because the implementation used implicit solvent, protein backbone restraints, and no explicit standard-state or restraint-release corrections [2311.17925].

The paper is explicit that rigorous absolute DBFE from such PMF protocols requires standard-state and restraint corrections, including terms of the form
$$
\Delta G^\circ_{\mathrm{ss}} = -k_B T \ln\!\left(C^\circ V_{\mathrm{eff}}\right),
$$
and, when orientational restraints are used, analytic correction formulas akin to the Boresch treatment [2311.17925]. This makes clear that pathway methods can produce either qualitative mechanistic free-energy profiles or quantitative DBFEs, depending on whether the full thermodynamic bookkeeping is included.

Attach–pull–release, umbrella sampling, funnel metadynamics, and related physical-coordinate approaches are described in comparative context in the SAMPL9 cyclodextrin study. That work notes that direct absolute routes can become difficult when binding involves large host reorganization and multiple competing poses, circumstances under which the authors chose the Alchemical Transfer Method instead of a direct PMF-style absolute route [2302.08620]. This suggests that “directness” in DBFE is often limited by overlap and coordinate-design issues rather than by thermodynamic principle.

## 5. Machine-learning uses of DBFE

In machine-learning papers, DBFE is used as a prediction target rather than a simulated free energy. TopoBind models antibody-antigen binding free energy as a scalar thermodynamic quantity $\Delta G$ in kcal/mol and treats this as the direct, absolute binding free energy of the complex [2508.19632]. Its curated dataset begins with 1705 antibody-antigen complexes from 472 unique PDB IDs and, after modality filtering, retains 303 PDB IDs corresponding to 1398 usable instances, split into 978/209/211 train/validation/test instances with no PDB ID overlap [2508.19632]. The method fuses pooled ESM-2 embeddings with a 100-dimensional topological descriptor built from contact map metrics, interface contact statistics, distance map statistics, and persistent homology invariants up to dimension 2. Bidirectional cross-attention aligns sequence and topology, and the main model uses a Lasso regressor on the fused embedding [2508.19632].

The reported test performance for TopoBind with Lasso is MSE $3.8160$, RMSE $1.9535$, MAE $1.4255$, Pearson $r = 0.5905$, $R^2 = 0.3390$, and classification accuracy $0.7500$ for strong binders defined by $\Delta G < -10$ kcal/mol. The best topology hyperparameters are a contact threshold near $8.0$ Å and persistent-homology top-$k = 6$, with performance described as robust near those choices [2508.19632].

A different ML framing appears in the Boltzmann-Aligned inverse folding model for protein-protein interactions. There, DBFE is approximated from inverse-folding log-likelihood differences between the bound complex and the decoupled monomers:
$$
\widehat{\Delta G}_{\mathrm{bind}}
\approx
-k_B T \left[
\ln p_{\theta}(s_{AB}\mid S_{AB}^{\mathrm{bnd}})
-
\big(
\ln p_{\theta}(s_A\mid S_A)
+
\ln p_{\theta}(s_B\mid S_B)
\big)
\right].
$$
The explicit treatment of the unbound state is presented as the key physical inductive bias. On antibody-antigen binding energies in the SAbDab test, the resulting DBFE estimator achieves a Spearman correlation of $0.385$, outperforming DSMBind and other baselines according to the paper [2410.09543].

These ML uses broaden the meaning of DBFE. The target remains the absolute $\Delta G$ of binding, but the estimator may be a supervised regressor trained on experimental labels or a Boltzmann-aligned proxy derived from sequence-conditioned structural likelihoods. A plausible implication is that, in current ML usage, DBFE often functions as the experimentally defined endpoint quantity against which surrogate models are calibrated rather than as a directly simulated thermodynamic observable.

## 6. Practical difficulties, corrections, and recurring controversies

Several recurring error sources cut across DBFE methodologies. Standard-state handling is one of the most universal. ILT, SDM, end-state DBFE, ER theory, and pathway approaches all require a correction converting a confined or restrained binding region into the $1$ M reference state [1208.4885] [1904.11058] [2603.12253] [2410.04423]. When such terms are omitted, as in the qualitative iMD-VR-FE implementation, the resulting free energy is explicitly not an absolute standard-state DBFE [2311.17925].

Hydration and solvent representation are another major issue. The D3 receptor study shows that confined, energetically frustrated waters can dominate binding thermodynamics and that neglecting them can produce gross underestimation of affinity in implicit-solvent alchemical DBFE [1904.11058]. The 2015 solution-phase review similarly identifies explicit solvent and ion effects, wrong ionization states, errors in ionic solvation free energies, molecular symmetry, anharmonicity, and insufficient conformational sampling as factors that can each contribute roughly $1$–$3$ kcal/mol errors at $298$ K [1501.04428].

Multiple poses and chemical species complicate any absolute DBFE definition when the experimentally observed affinity is an aggregate over substates. The SAMPL9 cyclodextrin study treats pose-specific and species-specific binding constants through a combination formula,
$$
K_b = \sum_i P_0(i)\,K_b(i),
$$
equivalently,
$$
\Delta G^\circ_{\mathrm{bind,total}}
=
-k_B T
\ln
\left[
\sum_i P_0(i)\,e^{-\beta \Delta G^\circ_{\mathrm{bind},i}}
\right].
$$
For mixtures of species, the aggregation extends to
$$
\Delta G^\circ_{\mathrm{mix}}
=
-k_B T
\ln
\left[
\sum_s f_s \sum_{p\in\text{poses}} e^{-\beta \Delta G^\circ_{s,p}}
\right].
$$
The paper’s conclusion is that apparently simple host-guest systems may require explicit treatment of multiple poses, enantiomers, protonation states, and slow conformational modes for reliable free-energy prediction [2302.08620].

Interpretability also varies strongly by method. Classical statistical-mechanical and end-state approaches can identify physically meaningful corrections such as clash entropy, hydration displacement, confinement volume, or reorganization free energy [2603.12253] [1904.11058]. ML approaches can be predictive but are only indirectly thermodynamic: sparse Lasso weights improve interpretability in TopoBind, yet cross-attention and topology encoders remain black-box components according to the paper [2508.19632].

A common misconception is that “direct” automatically means more physically rigorous. The literature does not support that equivalence. Some direct methods are fully rigorous when all corrections are included, as in ILT-based and ER-based formulations [1208.4885] [2410.04423]. Others are explicitly approximate or qualitative because of solvent models, rigid receptors, missing corrections, or proxy-based inference [2311.17925] [1701.05073] [2410.09543].

## 7. Comparative perspective

DBFE now spans several methodological regimes with a shared target but different computational philosophies.

| Regime | Core idea | Representative paper |
|---|---|---|
| Rigorous direct statistical mechanics | Receptor/ligand sampling separated or insertion-based free-energy averaging | [1208.4885], [1708.07045], [1701.05073] |
| End-state non-alchemical ABFE | Bound and decoupled end states connected without $\lambda$ intermediates | [2603.12253], [2410.04423] |
| Alchemical absolute binding free energy | $\lambda$-dependent Hamiltonians with standard-state and restraint corrections | [1904.11058], [2302.08620] |
| Pathway/PMF direct methods | Physical unbinding pathways profiled along CVs | [2311.17925] |
| Machine-learning DBFE prediction | Experimental or proxy $\Delta G$ predicted from structure/sequence representations | [2508.19632], [2410.09543] |

Taken together, the literature shows that DBFE is less a single algorithm than a unifying thermodynamic objective. In the narrowest modern sense, especially in “Binding Free Energies without Alchemy,” DBFE denotes an end-state absolute binding free energy method that avoids alchemical intermediates [2603.12253]. In the broader sense used across statistical mechanics, simulation, and ML, it denotes the absolute binding free energy $\Delta G$ itself—distinguished from $\Delta\Delta G$ and from qualitative binding scores—regardless of whether it is obtained by rigorous ensemble averages, physical-pathway PMFs, alchemical cycles, or predictive surrogates [2508.19632] [2410.09543] [1208.4885].

Source: https://www.emergentmind.com/topics/direct-binding-free-energy-dbfe