Accurate protein-ligand binding free energy calculations are a valuable tool in computational drug discovery as they enable the identification and prioritization of highly active compounds, thus reducing experimental costs. Free Energy Perturbation (FEP) is a gold-standard method for computing binding free energies, yet its accuracy depends significantly on the quality of the input protein-ligand complex structures, and it incurs a high computational and potentially high financial cost.

AI-driven protein structure prediction models offer ways to produce homology models using available protein structure data, these can be used to compute ligand binding free energies for systems where no crystal structure is available. However, this approach is still heavily reliant on the accuracy of the homology model. While models such as AlphaFold2 have shown some success in reproducing binding free energies comparable to crystal structures, its inability to predict holo protein structures limits its direct application in an efficient FEP workflow.¹

Baidu’s PaddleHelix team developed HelixFold3 (HF3),² a protein structure prediction model, at a time when AlphaFold3 was only available as a webservice.³ Designed to emulate AlphaFold3, HF3 provided a more accessible solution to protein-ligand complex prediction while maintaining accuracy. Furui and Ohue have used Flare FEP to validate the practicality of the HF3 model.⁴

In this article we will showcase Flare’s FEP solution and how it can be efficiently applied to accurately predict ligand binding free energies where the protein models have been generated with HF3 which validates their protein-ligand structure prediction approach.

Calculating binding free energies using Flare FEP

FEP is a computational technique used to predict the binding free energies of molecules. Accurate and efficient calculation of these values enables researchers to prioritize the most promising candidates early in the drug discovery process, reducing the time and cost associated with experimental testing. Cresset is at the forefront of digitized chemistry solutions, offering advanced relative and absolute FEP methods via Flare FEP.

Figure 1. Image showing the JNK1 derivative set in the binding site with the protein electrostatic potential surface and the FEP map for those ligands, generated automatically using Flare FEP. The ligands were obtained from the supporting material of the Furui et.al. paper

Relative Binding Free Energy (RBFE) calculations is a flavor of FEP that calculates the binding free energies between chemically related ligands relative to each other. It works by first generating an FEP map (Figure 1) which links chemically similar ligands together and then calculating the free energy change between the linked ligands by transforming one into another via an alchemical pathway. This process is non-trivial, but Flare FEP has been designed to make performing these steps as accurate, efficient and accessible as possible. Therefore, map generation has been automated using Cresset’s extensively modified version of LOMAP ⁴, then if there are any transformations in the map which are too complex to occur in a single step, Flare is able to intelligently identify and insert intermediates between the molecules.

The alchemical pathway used for a transformation between two end states is composed of a series of lambda windows, since each lambda window consists of a short molecular dynamics simulation, the number of lambda windows used can greatly impact the efficiency of the calculation. Flare FEP uses the ‘adaptive lambda window algorithm’⁵ to automatically assess and apply the optimal number of lambda windows for each transformation necessary to achieve convergence of the FEP calculations, typically reducing the number of windows simulated by 30%. These features ensure Flare FEP remains accurate, reliable, efficient and user-friendly, making it well-suited for seamless adoption and exploratory use.

Validating HelixFold3 predicted protein-ligand complex structures using Flare FEP

The most successful FEP projects rely on a good understanding of the target system and an accurate representation of the protein-ligand structure. In an ideal world there would be a well-resolved protein-ligand complex to ensure the free energy calculations are performed on a biologically relevant conformation. However, since these are not always available, AI models can be used to predict them.

Furui and Ohue used HF3 to predict the structures of 8 targets (BACE, CDK2, JNK1, MCL1, P38, PTP1B, Thrombin, TYK2) from the Wang et. al. FEP benchmark dataset and compared them with the experimentally derived crystal structures.⁶ First, they predicted five apo and holo structures for each target, selected one of the five and then assessed three RMSDs: the ligand RMSD, the global RMSD and the binding site RMSD. They found that across all structures the ligand RMSD was below 2 Å, the global RMSD was lower for the holo predicted structures than the apo, with the exceptions of BACE and JNK1. Comparing binding site RMSDs between the apo and holo structures, showed improvements for CDK2, MCL1, P38, PTP1B holo structures over the apo, whereas the RMSDs remained approximately the same for other targets. Despite the improvements in the binding site RMSDs for JNK1 and P38, they remained above 2 Å, indicating that HF3 struggles to accurately predict these binding sites with respect to RMSD.

Calculating binding free energies offers a more practical assessment on the accuracy of a predicted protein-ligand complex. Therefore, Furui and Ohue calculated the binding free energies of the dataset ligands with crystal, HF3 apo and HF3 holo protein structures, comparing them to experimental values. Their analysis of the Wang et al. dataset revealed some variability in model performance. Thrombin consistently showed the strongest predictive accuracy, with R² ranging from 0.856 to 0.882 and MUE between 0.152 and 0.381 kcal/mol. In contrast, BACE consistently performed the worst, with lower R² (0.123 to 0.325) and higher MUE (0.986 to 1.007 kcal/mol) ranges, while the crystal PTP1B had the lowest individual R² at just 0.002. Despite these fluctuations, if they exclude the crystal PTP1B and HF3 apo MCL1 complex the overlapping 90% confidence intervals, calculated from 10,000 bootstrap samplings, suggest that HF3 and Flare FEP can predict free energies with the same accuracy as crystal structures.

HF3 is trained on the structures from the PDB and therefore the training data will contain the structures in the Wang et. al. dataset. To challenge their model Furui and Ohue predicted derivative structures of the dataset, which were modeled using ligands absent from the training data. Their results showed that HF3 effectively predicted ligand positioning for most derivatives, with RMSD values typically staying below 2 Å. However, JNK1 derivatives showed notable deviations as 8 out of the 21 derivatives had RMSD values exceeding this threshold. Figure 2 illustrates the overlay of the JNK1 crystal structure with the most misaligned ligand, highlighting where HF3’s prediction diverged. A closer look suggests the difference stems from the crystal structure ligand lacking a chlorophenyl group, which, in our model recreated using the project files from the supporting information, tucks deeper into the protein pocket (Figure 2). Despite this variability, Flare FEP can reasonably predict the binding free energy of the ligands across the derivative set with MUE values of less than 1 kcal/mol for all targets except MCL1 which had an MUE of 1.5 kcal/mol.

Figure 2. Superimposed structures of JNK1. In purple is the HelixFold3 predicted structure with the ligand 18627-1 and in cyan is the crystal structure with the ligand 17124-1. This image was created using the files supplied by Furui and Ohue in the supplementary information

Conclusion

This article has showcased Flare FEP as a reliable and user-friendly method to predict ligand binding free energies. Furui and Ohue have shown the effectiveness of their HF3 model in predicting both apo protein structures and holo protein-ligand complexes. By integrating HF3 with Flare FEP, they were able to seamlessly predict binding free energies with some success, highlighting the potential to use Flare FEP in method validation and seamlessly integrate it into a computational drug design workflow.

References

1. Beuming, T.; Martín, H.; Díaz-Rovira, A. M.; Díaz, L.; Guallar, V.; Ray, S. S. Are deep learning structural models sufficiently accurate for free-energy calculations? Application of FEP+ to AlphaFold2-predicted structures. J. Chem. Inf. Model. 2022, 62, 4351–4360. https://doi.org/10.1021/acs.jcim.2c00796
2. Liu, L. et al. (2024) “Technical Report of HelixFold3 for Biomolecular Structure Prediction.” Available at: http://arxiv.org/abs/2408.16975
3. Abramson, J., Adler, J., Dunger, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 2024, 630, 493–500. https://doi.org/10.1038/s41586-024-07487-w
4. Furui, K., Ohue M., Benchmarking HelixFold3-Predicted Holo Structures for Relative Free Energy Perturbation Calculations. ACS Omega 2025. https://doi.org/10.1021/acsomega.4c11413
5. Liu, S., Wu, Y., Lin, T. et al. Lead optimization mapper: automating free energy calculations for lead optimization. J. Comput. Aided Mol. Des., 2013, 27, 9, 755-770. https://doi.org/10.1007/s10822-013-9678-y
6. Scott D. Midgley, Sofia Bariami, Matthew Habgood, and Mark Mackey, Adaptive Lambda Scheduling: A Method for Computational Efficiency in Free Energy Perturbation Simulations. J. Chem. Inf. Model. 2025 65 (2), 512-516. https://doi.org/10.1021/acs.jcim.4c01668
7. Wang, L. et al. Accurate and Reliable Prediction of Relative Ligand Binding Potency in Prospective Drug Discovery by Way of a Modern Free-Energy Calculation Protocol and Force Field. J. Am. Chem. Soc. 2015, 137, 2695–2703. https://doi.org/10.1021/ja512751q