Boltz-ABFE: Integrating Prediction & ABFE

• Boltz-ABFE is a workflow that combines Boltz-2 co-folding with ABFE, enabling accurate binding free energy computations without experimental protein–ligand structures.
• It employs a staged process including pose prediction, refinement, and rigorous MD preparation to generate simulation-ready models for free energy perturbation.
• Benchmark tests on targets like TYK2 and JNK1 show that Boltz-ABFE approaches crystal-based methods, broadening early-stage drug discovery applications.

Boltz-ABFE is a workflow for absolute binding free energy (ABFE) calculation in the absence of an experimental protein–ligand crystal structure. Its central design is to use the Boltz-2 co-folding model to predict an initial protein–ligand complex, automatically refine that prediction into a simulation-ready pose, and then apply an absolute free energy perturbation (FEP) protocol. The method is motivated by a specific bottleneck in structure-based drug discovery: classical ABFE is most reliable when initialized from a high-quality experimental complex structure, yet in early-stage discovery many targets lack a co-crystal structure with the ligand of interest, and available structures may instead be apo, derived from distant homologues, or predicted models. Boltz-ABFE is intended to remove that practical dependency without changing ABFE itself (Thaler et al., 26 Aug 2025).

1. Motivation and problem setting

Classical ABFE calculations are most reliable when the starting structure already resolves the ligand pose, binding-site waters, and local protein geometry. In practical discovery campaigns, however, many targets have no crystal structure with the ligand of interest, the binding pose may be uncertain, and the available structural information may be incomplete. Boltz-ABFE was developed to address this bottleneck by combining three components: a structure prediction model, automatic structure preparation or refinement, and an ABFE pipeline that can operate from a predicted complex rather than a crystal complex (Thaler et al., 26 Aug 2025).

A key conceptual point is that the method does not redefine the thermodynamic basis of ABFE. Instead, it shifts the entry point of the workflow. The objective is to make ABFE usable in early-stage drug discovery, where structural information is often absent or only weakly constrained. This suggests that Boltz-ABFE should be understood primarily as an integration of structure prediction with physics-based affinity estimation, rather than as a new free energy formalism.

The paper frames this as an expansion of the domain of applicability of FEP. In that framing, the limiting factor is not the underlying alchemical protocol but the availability of an accurate initial protein–ligand complex for molecular dynamics.

2. Boltz-2 as the complex initializer

The workflow begins with Boltz-2 co-folding of the protein and ligand. In the benchmark comparison, Boltz-2 generates an initial predicted complex structure directly from the target sequence and ligand identity. The paper distinguishes several starting-structure strategies: Crystal, Native, Boltz-2, Boltz-2+P, and Boltz-2-A. Here, Boltz-2 denotes the raw co-folded structure, Boltz-2+P denotes a Boltz-2 structure followed by POSIT redocking or refinement, and Boltz-2-A denotes affinity prediction from the Boltz-2 affinity module (Thaler et al., 26 Aug 2025).

The supporting information indicates that Boltz-2 is markedly better than Boltz-1 at generating clash-free structures suitable for downstream preparation. It also reports that, in many co-folding scenarios, generating 10 models per protein–ligand pair was sufficient to obtain at least one structure compatible with the OpenEye/Spruce preparation pipeline. In operational terms, Boltz-2 therefore functions as the pose generator and complex initializer for downstream ABFE.

This distinction between raw and refined Boltz-2 outputs is important. The predicted complex can be used directly, but it can also serve as an intermediate representation to be filtered and improved before molecular simulation. A plausible implication is that the quality of the initial co-folded ensemble, rather than a single best model, becomes a major determinant of downstream ABFE usability.

3. Preparation and refinement of predicted complexes

Predicted complexes are often not immediately simulation-ready. The workflow explicitly addresses steric clashes, incorrect bond orders or chemistry, strained ligand geometries, incompatible protonation or tautomer states, and poses that fail standard force-field preparation. The supporting information evaluates whether co-folded structures are “compatible with the OpenEye Spruce pipeline,” indicating that automated structure cleaning and preparation are a formal part of the method rather than an ad hoc preprocessing step (Thaler et al., 26 Aug 2025).

For some systems, predicted structures were refined using POSIT redocking. The supporting information compares raw Boltz predictions with Boltz-1 + POSIT and Boltz-2 + POSIT, indicating a pose-refinement stage after structure prediction and before ABFE. The practical significance is straightforward: Boltz-ABFE is not simply a protocol that runs FEP on an arbitrary predicted pose. It is a staged pipeline in which the complex is predicted, filtered or refined, prepared for molecular dynamics, and only then passed into the ABFE calculation.

The paper’s discussion of clash-free, processable structures is especially consequential. Boltz-2 produces far more such structures than Boltz-1, improving the likelihood that ligand chemistry is valid, that the pose can be parameterized, that the system can pass solvation and minimization, and that the resulting structure is suitable for molecular dynamics and FEP. This suggests that structure preparation is a first-order component of the method’s success, not a peripheral implementation detail.

4. Alchemical ABFE protocol and evaluation procedure

The full workflow can be summarized in four operational stages: prediction of one or more protein–ligand complexes with Boltz-2, selection or refinement of a usable pose, preparation of an MD-ready system, and execution of an ABFE calculation via an alchemical thermodynamic cycle (Thaler et al., 26 Aug 2025).

In the conceptual ABFE cycle, the ligand is decoupled from the protein in the bound state and from solvent in the unbound state, with additional restraint and standard-state corrections. The relation given is

$$\Delta G_{\text{bind}} = \Delta G_{\text{decouple}}^{\text{complex}} - \Delta G_{\text{decouple}}^{\text{solvent}} + \Delta G_{\text{restraints}} + \Delta G_{\text{standard-state}}.$$

The paper also notes a standard FEP-based estimator across alchemical windows, for example

$$\Delta G = -k_B T \ln \left\langle e^{-\beta \Delta U} \right\rangle,$$

or, more generally, estimation by BAR or MBAR across windows.

The evaluation protocol includes a specific post-processing step. RMSE and MUE were computed after centering predicted $\Delta G$ to the experimental mean, because the experiments report IC50-derived affinities whereas ABFE estimates $K_d$. As a result, the reported comparisons remove an overall offset and assess within-target performance rather than direct equality of observables. The tables report the mean of three replicates and bootstrap standard deviation as uncertainty, indicating that robustness was assessed through repeated ABFE calculations and resampling-based error estimation.

5. Benchmark targets and quantitative performance

The benchmark comprises four FEP+ targets: TYK2, CDK2, JNK1, and P38. Across these systems, the reported overall pattern is that crystal-structure ABFE remains the strongest or near-strongest baseline, Boltz-2-derived structures can support useful ABFE calculations without a crystal structure, POSIT-refined Boltz-2 sometimes improves performance but not uniformly, and the Boltz-2 affinity module is sometimes competitive, especially on TYK2 and JNK1 (Thaler et al., 26 Aug 2025).

Strategy	Overall error	Overall correlation
Crystal	RMSE 0.80 kcal/mol, MUE 0.64	$r = 0.75$, $\rho = 0.73$, $\tau = 0.53$
Boltz-1+P	RMSE 1.01, MUE 0.74	not reported in the data block
Boltz-2	RMSE 1.26, MUE 0.96	lower than Crystal and Boltz-2-A overall
Boltz-2+P	RMSE 1.23, MUE 0.93	not reported in the data block
Boltz-2-A	RMSE 0.77, MUE 0.64	$r = 0.61$, $\rho = 0.67$, $\tau = 0.50$

For TYK2, the reported values are: Crystal RMSE 0.66 kcal/mol, MUE 0.55, $$\Delta G = -k_B T \ln \left\langle e^{-\beta \Delta U} \right\rangle,$$0; Boltz-1+P RMSE 0.56, MUE 0.42, $$\Delta G = -k_B T \ln \left\langle e^{-\beta \Delta U} \right\rangle,$$1; Boltz-2 RMSE 1.71, MUE 1.44, $$\Delta G = -k_B T \ln \left\langle e^{-\beta \Delta U} \right\rangle,$$2; Boltz-2+P RMSE 1.63, MUE 1.26, $$\Delta G = -k_B T \ln \left\langle e^{-\beta \Delta U} \right\rangle,$$3; and Boltz-2-A RMSE 0.72, MUE 0.63, $$\Delta G = -k_B T \ln \left\langle e^{-\beta \Delta U} \right\rangle,$$4. The supplementary comparison of Boltz-1 refinement strategies on TYK2 reports: Crystal RMSE $$\Delta G = -k_B T \ln \left\langle e^{-\beta \Delta U} \right\rangle,$$5, Native Boltz-1 $$\Delta G = -k_B T \ln \left\langle e^{-\beta \Delta U} \right\rangle,$$6, POSIT $$\Delta G = -k_B T \ln \left\langle e^{-\beta \Delta U} \right\rangle,$$7, and HYBRID $$\Delta G = -k_B T \ln \left\langle e^{-\beta \Delta U} \right\rangle,$$8. This indicates strong dependence on refinement strategy.

For CDK2, the values are: Crystal RMSE 0.68, $$\Delta G = -k_B T \ln \left\langle e^{-\beta \Delta U} \right\rangle,$$9; Boltz-1+P RMSE 1.00; Boltz-2 RMSE 0.77, $\Delta G$0; Boltz-2+P RMSE 0.88; and Boltz-2-A RMSE 0.81. In this target, Boltz-2 is close to the crystal baseline and better than Boltz-1+P.

For JNK1, the values are: Crystal RMSE 0.89, $\Delta G$1; Boltz-1+P RMSE 1.12; Boltz-2 RMSE 1.31; Boltz-2+P RMSE 1.10; and Boltz-2-A RMSE 0.53, $\Delta G$2. The data emphasize that the affinity module substantially outperforms the raw co-folded structure on this target.

For P38, the values are: Crystal RMSE 0.86, $\Delta G$3; Boltz-1+P RMSE 1.16; Boltz-2 RMSE 1.25; Boltz-2+P RMSE 1.28; and Boltz-2-A RMSE 0.93. P38 is described as the most challenging of the four in terms of correlation, and the predicted-structure approaches do not clearly improve over crystal-based ABFE.

6. Robustness, failure modes, and practical interpretation

The supporting information emphasizes a practical metric beyond RMSE and correlation: the success rate of producing clash-free, pipeline-compatible structures. The main conclusions are that 10 Boltz-2 models per complex are often sufficient to obtain at least one usable structure, that Boltz-2 consistently improves over Boltz-1, and that some targets remain difficult, including CDK8, SHP2, and one internal target set identified as “target 7” (Thaler et al., 26 Aug 2025).

Several failure modes are identified. First, incorrect chemistry can occur: four ligands in the TYK2 set were generated with incorrect chemistry by Boltz-1. Second, many predicted structures fail preparation unless multiple models are sampled. Third, success is target-dependent rather than uniform across proteins. Fourth, a raw predicted pose may underperform relative to a refined or affinity-based output, as seen especially in TYK2 and JNK1.

These points address a common misconception about predicted-structure ABFE workflows. The limiting factor is not simply whether a co-folding model can place a ligand in a pocket; it is whether the resulting complex survives chemistry validation, force-field preparation, solvation, minimization, and subsequent alchemical simulation. In that sense, Boltz-ABFE is a robustness-oriented workflow rather than a single prediction model.

The reported use of multiple model generations, POSIT refinement, repeated ABFE replicates, and bootstrap uncertainty indicates an attempt to make the pipeline robust enough for screening use. A plausible implication is that ensemble generation and automated triage are structurally necessary components when ABFE is deployed without experimental co-structures.

7. Scope, limitations, and significance for early-stage discovery

The method’s principal significance lies in enabling ABFE in settings where no experimental co-structure exists. The paper’s bottom-line conclusion is that Boltz-ABFE extends ABFE/FEP beyond the narrow setting of crystal structures by using Boltz-2 to propose a bound pose, automatically screening or refining that pose into an MD-compatible structure, and then applying a standard alchemical ABFE workflow with repeated replicates and bootstrap error estimates (Thaler et al., 26 Aug 2025).

The limitations are equally explicit. Structure prediction remains the bottleneck: if Boltz-2 gives the wrong pose or wrong chemistry, ABFE cannot rescue it reliably. Not all targets are equally tractable. POSIT refinement can help but does not uniformly solve all issues. Absolute accuracy remains target-dependent, with P38 illustrating weaker ranking performance. Experimental comparison also requires centering because the experimental observables are often IC50-derived whereas ABFE estimates $\Delta G$4.

Within those constraints, the method is positioned for cases in which no crystal structure exists but a plausible binding mode is needed quickly and a physics-based estimate is preferred to a docking score alone. The data also indicate that, for some projects, the Boltz-2 affinity module may already provide a useful approximate ranking, while ABFE supplies a more rigorous thermodynamic estimate when the pose is reliable. This suggests that Boltz-ABFE is best interpreted as a bridge between modern co-folding models and established alchemical free energy methods, with its practical value determined primarily by the quality and consistency of the predicted complex.