Dual-LAO for calculating fast and robust relative binding free energies of simple and complex transformations Authors: Narjes Ansari, Félix Aviat, Jérôme Hénin, Jean-Philip Piquemal, Louis Lagardère Published: 2026 ABSTRACT AbstractRelative binding free energy (RBFE) calculations are a cornerstone of rational hit-to-lead and lead optimization in modern drug discovery. However, the high computational cost and limited reliability in tackling large or complex molecular transformations often prevent their routine, high-throughput use. Here we introduce Dual-LAO, a novel, highly efficient method for calculating RBFE. Building on the Lambda-ABF-OPES framework, this method combines a dual-topology setup and suitable restraints to dramatically accelerate free energy convergence. We demonstrate that Dual-LAO, in combination with the AMOEBA polarizable force field, achieves an unprecedented acceleration factor of 15 to 30 times compared to current state-of-the-art methods on standard drug targets. Crucially, the approach maintains high accuracy and successfully tackles previously prohibitive molecular changes, including scaffold-hopping, buried water displacement, charge changes, ring-opening, and binding pose perturbations. This significant leap in efficiency allows for the widespread, routine integration of predictive molecular simulations into the rapid optimization cycles of drug discovery, enabling chemists to confidently model historically challenging systems in timescales compatible with real-world project deadlines. Subject terms: Thermodynamics, Computational chemistry, Computational chemistry Introduction Accelerating therapeutic development critically hinges on accurately predicting the binding affinity (ΔG) between small molecules and macromolecular targets1. Computational methods play a pivotal role, offering the potential to streamline the costly and time-consuming process of drug discovery by prioritizing promising candidates for experimental investigation2–4. Among the most rigorous computational tools are physics-based alchemical free energy (AFE) methods, which leverage statistical mechanics and molecular simulations5. Alchemical methods rely on the exploration of non-physical intermediate states to perform a transformation between two physical end-states. One class of such methods, absolute binding free Energy (ABFE) calculations, aims to compute the total free energy change upon ligand binding6. While conceptually appealing, ABFE prediction remains computationally demanding, often struggling with slow convergence due to the extensive conformational sampling required for both the binding process and the large alchemical transformations involved (e.g., full ligand decoupling)7. Achieving reliable ABFE results necessitates both accurate force fields capable of capturing subtle molecular interactions (including polarization)8 and highly efficient sampling techniques to overcome significant free energy barriers9–18. Consequently, RBFE calculations have become the predominant approach in hit-to-lead and lead optimization campaigns19. RBFE methods calculate the difference in binding affinity (ΔΔG) between two structurally related ligands (L1 and L2) via a thermodynamic cycle20,21 by alchemically transforming L1 into L2 in both the protein-bound state (ΔGcomplex) and the unbound, solvated state (ΔGsolv): These methods benefit from significant error cancellation and typically involve smaller, more manageable perturbations than ABFEs19. This makes RBFE less computationally intensive than ABFE and often capable of higher precision for comparing congeneric series19. Furthermore, the computational cost for all these simulations has been dramatically reduced by the use of Graphics Processing Units (GPUs), enabling RBFE calculations to be performed with sufficient throughput to impact active drug discovery projects19. RBFE are typically performed using molecular dynamics (MD). The free energy difference is commonly estimated using methods like Free Energy Perturbation (FEP)22 or Thermodynamic Integration (TI)23, often involving multiple simulations at discrete intermediate states along the transformation path. To improve efficiency, alternative approaches have been developed. These include nonequilibrium methods24, which are highly parallelizable but may require significant computational effort to reach the same level of precision as equilibrium methods, a trade-off that remains a subject of ongoing discussion in the community25. Another class of methods involves expanded ensemble techniques, where the alchemical coupling parameter λ is treated as an extended variable that evolves during a single simulation. Prominent examples include λ-dynamics (λD)26 and its extension, multisite λ-dynamics (MSλD)27. MSλD is particularly scalable, allowing simultaneous perturbations at multiple molecular sites and enabling the exploration of large combinatorial chemical spaces with significantly fewer simulations than traditional FEP or TI. To practically model this alchemical transformation, several topology schemes can be employed. The single-topology approach requires mapping the atoms of L1 onto L2 and introducing “dummy” atoms for non-matching regions, which can become complex for dissimilar ligands20,28. The dual-topology approach simulates both L1 and L2 simultaneously, with the interactions of one ligand being “turned off” while the other is “turned on”20,29. This avoids dummy atoms but can require careful spatial restraints to manage the non-interacting components30. An alternative is the hybrid or modified dual-topology scheme, which defines a maximal common substructure (MCS) as a single, shared entity while treating the non-shared “appearing” and “disappearing” regions separately, thus avoiding both dummy atoms and the need for complex restraints31,32. Modern RBFE protocols, combining advanced force fields (e.g., OPLS3e33), enhanced sampling techniques (e.g., REST234), and automated workflows19, have demonstrated impressive performance in numerous retrospective and prospective studies, frequently reporting root-mean-square errors (RMSE) relative to experiment approaching 1 kcal.mol-1 35–37. This level of accuracy, often designated as “chemical accuracy”, allows RBFE predictions to effectively rank compounds and guide synthesis prioritization35. Large-scale prospective studies have confirmed the value of RBFE in enriching hit rates for potent compounds compared to other methods38. Despite these successes, RBFE methods face significant challenges that can impact their reliability and accuracy5,21. Accuracy is intrinsically limited by the underlying force field and, critically, by the extent of conformational sampling39. Insufficient sampling is a frequent pitfall, exacerbated by the chaotic nature of molecular dynamics, which mandates the use of ensemble simulations for statistically robust and reproducible results40. Furthermore, the resulting free energy distributions are often non-Gaussian, complicating error analysis41. Several specific scenarios are notoriously difficult: transformations involving large structural rearrangements (macrocyclization or scaffold hopping42,43), changes in ligand net charge44, covalent inhibitors19, uncertainties in the initial binding pose43, and perturbations involving the displacement or reorganization of buried water molecules5. Ambiguities in protein structure (e.g., missing loops) also pose challenges, often necessitating retrospective validation before prospective use. Furthermore, uncertainties in the precise binding site configuration and the ligand’s initial binding pose represent the primary bottlenecks for the prospective application of BFE calculations45. It is also crucial to contextualize computational accuracy targets against experimental reproducibility limits; independent experimental measurements of the same affinity can differ by 0.8–1.0 kcal.mol−1 (RMSE ≈ 0.6 to 0.7 pKi units), setting a practical limit on achievable computational accuracy46. Recently, the Lambda-ABF-OPES method was introduced as a highly efficient hybrid approach for calculating ABFE47. By coupling Lambda-adaptive biasing force (Lambda-ABF)48 with on-the-fly probability enhanced sampling (OPES)16,17, this technique achieves significantly faster convergence — up to 9-fold improvement compared to the original Lambda-ABF — while maintaining high accuracy, demonstrated with a mean absolute error of approximately 0.9 kcal.mol-1 for challenging bromodomain inhibitors47. Lambda-ABF-OPES sampling acceleration also relies on the use of several simulations running simultaneously, called “walkers”, exchanging at fixed intervals the information gathered by their respective ABF-estimators. Each such walker thus benefits from the exploration of the others, which effectively increases the rate of exploration for all. A key aspect contributing to the accuracy reached in this work is the use of the advanced AMOEBA polarizable force field47. Given the established difficulties in sampling for both ABFE and RBFE, applying such a demonstrably rapid and accurate methodology developed for ABFE within the RBFE context holds significant promise. In this work, we introduce Dual-LAO (Dual-topology, Dual-DBC Lambda-ABF-OPES), a novel RBFE strategy designed to significantly accelerate convergence and improve accuracy. This method integrates three core components: a dual-topology representation of the transformed ligands, enhanced alchemical sampling using the Lambda-ABF-OPES47 framework, and a dual-DBC scheme utilizing specific distance-to-bound configuration (DBC) restraints49. We implement this approach within the Tinker-HP simulation package, leveraging the advanced AMOEBA polarizable force field50. This combined approach benefits from the sampling power of Lambda-ABF-OPES, the flexibility of the dual topology method, and the narrowing down of phase space volume due to the DBC restraints. In the following, we rigorously evaluate the performance of this implementation across a spectrum of challenging scenarios frequently encountered in drug discovery. These transformations cover four distinct types, each posing unique challenges to sampling and convergence: (i) fragment-based transformations, including difficult charge-changing perturbations, (ii) perturbations involving problematic buried water displacement, (iii) fragment-like chemical perturbations, and (iv) complex scaffold-hopping perturbations, featuring challenging ring-opening transformations. Our findings demonstrate the high accuracy and robustness of the Dual-LAO method across this diverse and challenging set. Across the entire dataset comprising 30 compounds and 56 alchemical edges, the approach yields an overall edgewise RMSE of 0.52 kcal.mol−1 and a pairwise RMSE of 0.56 kcal.mol−1, with a coefficient of determination (R2) of 0.90. Crucially, this high level of accuracy is achieved with a simulation time of only 4 ns/walker for the complex and 2 ns/walker for the solvent phase per transformation (using 4 walkers). These results demonstrate that the Dual-LAO approach consistently achieves rapid convergence and high accuracy, with RMSE values below 1 kcal.mol−1 across all these RBFE test systems, in significantly smaller simulation time than conventional RBFE methods. To further accelerate calculations, particularly when assessing multiple related analogues, we implemented Multiple topology Dual-LAO within the Tinker-HP package. This extension allows for the simultaneous sampling of multiple alchemical transformations within a single simulation framework, leveraging the efficiency benefits observed in other multiple topology dynamic lambda approaches51. In this framework, all free energy differences are obtained concurrently from just one simulation run. As validation of this implementation, we calculated the relative hydration free energies for 15 distinct molecule pairs within the benzene family. Results and discussion This section details the RBFE results obtained from various alchemical transformations using Dual-LAO method. These transformations cover: fragment-based transformations, buried water displacement, fragment-like chemical perturbations, and scaffold-hopping perturbations. Representative examples are provided for each type. RBFEs of fragments for the PWWP1 domain of NSD3 Our relative binding free energy RBFE simulations were performed as part of a fragment elaboration study targeting the PWWP1 domain of NSD352. A set of 12 ligands (2D structures of the ligands are detailed in Supplementary Fig. S1) was strategically selected53, covering a significant range of binding affinities, specifically from an initial 160 μM fragment hit (ligand 8) up to the highly potent 170 nM optimized lead compound, 9321. Detailed affinity data are provided in the Supplementary data 1. The alchemical free energy network for these transformations — primarily focused on R-group modifications — was initially constructed using the LoMap algorithm54 (see Supplementary Fig. S2). This yielded a total of 15 alchemical transformation pairs that connected the 12 ligands. These pairs were strategically chosen to represent different types of chemical changes and were broadly categorized into two types, as illustrated in Fig. 1a1, a2. RBFE Alchemical Transformations and Systems. a PWWP1-ligand complex: a1 shows a side chain transformation where the change involves only the peripheral substituent, keeping the core binding pose conserved. a2 illustrates a binding pose transformation, where Ligand 1 (green) adopts a different pose compared to Ligand 2 (red), leading to new protein-ligand interactions. b BRD4-ligand complex: b1 and b2 display buried water displacement. The comparison shows the displacement of ordered buried water molecules (red and white spheres) as the ligand size increases and extends its binding pose. c P38-ligand complex: c1 and c2 represent fragment-based perturbations, where the transformation causes a change in the ligand’s interaction with the protein. d CHK1-ligand complex: d1, d2 show scaffold-hopping transformations, where the R1 and R2 side chains change from an open form to a closed ring structure. Side Chain Transformations (Fig. 1a1): These pairs feature two ligands that share the same central binding core (shown in yellow). The example of the transformation of ligand 10 to ligand 11 is shown in this figure, where the transformation occurs solely on a peripheral substituent. This type of transformation is typically less challenging to sample in RBFE calculations because the binding mode remains conserved. Side Chain Transformations (Fig. 1a1): These pairs feature two ligands that share the same central binding core (shown in yellow). The example of the transformation of ligand 10 to ligand 11 is shown in this figure, where the transformation occurs solely on a peripheral substituent. This type of transformation is typically less challenging to sample in RBFE calculations because the binding mode remains conserved. Binding Pose Transformations (Fig. 1a2): These pairs represent a more challenging scenario in which the two ligands adopt different binding poses (for example ligand 17 to ligand 9466), leading to changes in protein-ligand interactions, including hydrogen bonding (HB) and van der Waals forces. This results in a binding free energy difference of around 3–4 kcal.mol−1. Such a substantial change in the protein-ligand interaction pattern and binding pose requires robust conformational sampling and presents a stringent test for the RBFE methodology. Binding Pose Transformations (Fig. 1a2): These pairs represent a more challenging scenario in which the two ligands adopt different binding poses (for example ligand 17 to ligand 9466), leading to changes in protein-ligand interactions, including hydrogen bonding (HB) and van der Waals forces. This results in a binding free energy difference of around 3–4 kcal.mol−1. Such a substantial change in the protein-ligand interaction pattern and binding pose requires robust conformational sampling and presents a stringent test for the RBFE methodology. In addition to the pose changes, the network included a charge transfer transformation (pair: ligand 10 → ligand 12, see Supplementary Fig. S1). Such transformations, in which the charge state of a ligand changes between the initial and final states, are difficult to model accurately in the computational literature36. Our calculated result for this specific, challenging pair (ΔΔGCAL = − 1.22 ± 0.11 kcal.mol−1) shows excellent agreement with the experimental value (ΔΔGEXP=−1.04kcal⋅mol−1). It is noteworthy that our results achieved convergence (see Supplementary Figs. S3–S6) in less than 4 ns/walker for the complex phase and less than 2 ns/walker for the solvent phase. This contrasts sharply with state-of-the-art methods utilizing alchemical charge-changing perturbations, which typically require 24 λ windows, with each window simulated for 20 ns36. In total, for such challenging systems, our method demonstrates a convergence speed approximately 30 times faster. To further validate these results, we performed additional simulations incorporating co-alchemical ion corrections to maintain system neutrality; the results (detailed in Section Assessment of charge-change corrections of the Supplementary information) confirm that the standard Dual-LAO treatment effectively captures the binding physics of this system without the need for explicit correction factors. To ensure statistical rigor of the calculations, RBFE was calculated in both directions (L1 → L2 and L2 → L1) for each pair, resulting in a total of 30 calculated relative free energy values (ΔΔGCAL) throughout the network (see Supplementary data 1 and Supplementary Fig. S2). The calculations were performed using our dual topology framework and tested with the dual-DBC restraint scheme. The raw data, including the calculated ΔΔGCAL and its associated uncertainty for every leg, are compiled in Supplementary data 1. The performance of the RBFE calculations was assessed by comparing the simulated values (noted ΔΔGCAL) with experimentally obtained ones (noted ΔΔGEXP). Supplementary Fig. S7a illustrates the correlation plot for the ΔΔGCAL values, and Fig. 2a shows the correlation plot for the Absolute Binding Free Energy (ΔGCAL) values derived from the relative network, following the method described in Supplementary information. Experimental vs. calculated absolute binding free energies derived from RBFE. a PWWP1, b BRD4, c P38, and d CHK1 complexes. The calculated ΔGCAL results and associated errors represent the mean ± standard error derived from the weighted least-squares (WLS) fitting of the RBFE network (see SI for more details). The dark and light shaded regions indicate ± 1 kcal.mol−1 and ± 2 kcal.mol−1 deviations from the experimental values, respectively. The color bar displays the absolute difference between the experimental and computed values, ∣ΔGEXP−ΔGCAL∣. For each system, key performance statistics including Pearson’s r, RMSE, and MAE are reported. The dual-DBC restraint scheme yielded excellent agreement with the experimental data, with the total RMSE for ΔΔGCAL approximately 0.46 kcal.mol−1 and RMSE of 0.68 kcal.mol−1 for ΔGCAL. These values compare favorably to the work of Alibay et al.53, who reported an RMSE of 1.14 kcal.mol−1. To further assess the robustness and convergence of our protocol, we calculated the cycle-closure for all possible closed cycles within the PWWP1 dataset (see Supplementary Fig. S1 and Supplementary Table S1). The calculated closure values are notably low, with a maximum absolute deviation of 0.61 kcal.mol−1 and standard deviations consistently remaining below 1kcal.mol−1. This high degree of internal consistency suggests that no inherent bias arises from the dual-DBC protocol. This demonstrates the robustness of the methodology in different types of chemical transformations, including the pose changes and charge transfers. RBFEs of BRD4 with buried water displacement One of the significant challenge in both ABFE and RBFE simulations arises when perturbations involve displacing buried water molecules within the protein active site. To provide another challenging test for our current Dual-LAO method, we thus selected the Bromodomain 4 (BRD4) system, as shown in Fig. 1 panels b1 and b2. This target is a well-known benchmark for computational methods, recognized to be problematic for achieving accurate binding free energies. In our prior studies47,55 utilizing the Lambda-ABF-OPES methodology, we successfully calculated the ABFE for seven targets with different variety of ligands featuring such buried water displacement events, and all results demonstrated excellent predictive accuracy for ΔGCAL in these challenging systems. In the work by Ross et al.36 using FEP + , the BRD4 ligand set presented considerable difficulty, resulting in an RMSE of 2.05 ± 0.24 kcal.mol−1 and an R2 value near zero. We applied our Dual-LAO approach to this same challenging set of ligands, which consists of eight ligands and eleven alchemical transformation pairs (see Supplementary Fig. S8 for the two-dimensional (2D) structures of the ligands and Supplementary Fig. S9 for the LoMap network illustrating the various alchemical connections.) The correlation plots of ΔΔGCAL and ΔGCAL with experimental values are presented in Supplementary Fig. S7b and Fig. 2b, respectively. Additionally, as an example, the convergence of PMF and the λ exploration over time for the transformation of ligand 2 (with six buried water molecules, Fig. 1b1) to ligand 8 (with only four buried water molecules, Fig. 1b2) are shown in Supplementary Figs. S11-S13. Here again, the convergence was achieved in less than 4 ns/walker for the complex phase and 2 ns/walker for the solvent phase. Dual-LAO results show a marked improvement over the FEP + benchmark, demonstrating the method’s ability to handle systems involving complex buried water perturbations. Specifically, we achieved an RMSE of 0.64 kcal.mol−1 and an R2 of 0.74 for ΔΔGCAL, and an RMSE of 0.42 kcal.mol−1 and an R2 of 0.76 for ΔGCAL. In order to check the proper displacement of buried water molecules during the Dual-LAO simulations, we analyzed as an example the coordination number (CN) of water around the O1 atom (see Supplementary Fig. S10) for the alchemical transformation pair connecting ligand 2 (λ = 0) to ligand 5 (λ = 1). We focused specifically on this atom to ensure clear visualization and analysis of the water perturbation. Supplementary Fig. S10 displays the density plot of water molecules around the O1 atom as a function of the alchemical coupling parameter, λ. As shown in the figure, the coordination number of water molecules around the O1 atom is higher when the system is closer to the ligand 2 state (λ ≈ 0) compared to the ligand 5 state (λ ≈ 1). This difference confirms that the Dual-LAO methodology successfully drives the necessary water perturbation and displacement as the ligand transforms. The success of the Dual-LAO approach, particularly in resolving these challenging water rearrangements, can be further elucidated through a comparison with stochastic methods such as non-equilibrium grand canonical nonte carlo (NE-GCMC)56. While NE-GCMC facilitates water insertion/deletion through discrete non-equilibrium switching moves, Dual-LAO utilizes a continuous dynamical coupling enhanced by the OPESe and ABF algorithms. As demonstrated in our previous work47,55, the integration of OPESe effectively treats the solvent environment as an orthogonal degree of freedom. By adaptively biasing the probability distribution, OPESe actively drives the exploration of the configurational space, facilitating the transition of water molecules in and out of buried cavities. This provides a specific advantage over GCMC-based sampling in cases where water movement is strongly coupled to protein conformational plasticity. The joint alchemical-configurational landscape allows for a mutual relaxation of the binding pocket and the solvent environment, lowering the effective free energy barriers to water reorganization without the need for predefined GCMC trial volumes or discrete switching states. RBFEs of P38 with fragment-like chemical perturbations We applied the Dual-LAO method to the P38 fragment set from Ross et al.36, which comprises 5 ligands (See Supplementary Fig. S14 for 2D structures of the ligands) and 7 transformations (see Supplementary Fig. S15 for the LoMap network illustrating the various alchemical connections). Panels c1 and c2 in Fig. 1 show a representative transformation pair. Ligand 4 forms 3 stable hydrogen bonds (HBs) with residues THR103, HIE104, and MET106 (see Fig. 1 c1), while Ligand 7 interacts only with residue MET106 via 2 HBs (see Fig. 1c2). In this specific perturbation, the left side of the two ligands is identical, while the right side undergoes a complex change. The convergence of PMF and the λ exploration over time for the transformation of ligand 4 to ligand 7 are shown in Supplementary Figs. S16–S19. For these transformations, our results yielded an RMSE of 0.42 kcal.mol−1 for ABFE and 0.82 kcal.mol−1 for RBFE. These values represent an improvement over the benchmarks established by Steinbrecher et al.57, who reported an RMSE of 1.25 kcal.mol−1 for fragment-based systems. This excellent agreement further demonstrates the effectiveness of the Dual-LAO method for accurately modeling such fragment-like transformations. RBFEs of CHK1 with scaffold-hopping perturbations As mentioned earlier, scaffold-hopping perturbations represent the most challenging and aggressive class of alchemical transformations studied using RBFE methods. Scaffold hopping involves replacing the core structural motif (the ’scaffold’) of a ligand with a structurally distinct motif while aiming to retain similar potency and the overall binding mode. To further challenge our method, we utilized the CHK1 scaffold-hopping set from Ross et al.36. This set contains 5 ligands involved in 8 different transformations (the 2D structure of the ligands is available in Supplementary Fig. S20 and the LoMap network illustrating the various alchemical connections are shown in Supplementary Fig. S21). Fig. 1d1, d2 present one such transformation, where the CH2CH3 moiety in the R1 and R2 positions of Ligand 19 transforms into a closed 7-member ring in Ligand 21. This exemplifies a crucial ring opening/closing type of perturbation. The convergence of PMF and the λ exploration over time for the transformation of ligand 19 to ligand 21 are shown in Supplementary Figs. S22–S25. For this difficult transformation class, the Dual-LAO method yielded an RMSE of 0.31 kcal.mol−1 for ABFE and 0.58 kcal.mol−1 for RBFE, indicating an excellent correlation between experimental and simulation results (see Fig. 1). These results show an improvement over the work of Ries et al.58, who used Replica-Exchange Enveloping Distribution Sampling procedure. These results highlight the ability of the Dual-LAO approach to circumvent the numerical instabilities traditionally associated with ring-breaking transformations in single-topology frameworks. As noted by Liu et al.59, ring opening via dummy atoms often leads to convergence failures due to poorly defined bonded potentials at the alchemical endpoints. In contrast, our method utilizes a dual-topology scheme where the initial and final ligand states are defined as independent topologies. This architecture eliminates the need to break or close bond constraints within a single molecule, and Lambda-ABF-OPES mitigates the challenge of ensuring configurational overlap between independent topologies. This algorithm provides a robust biasing force that samples the joint alchemical-configurational space, ensuring that even during drastic structural transitions, the system remains numerically stable and the free energy converges rapidly through consistent sampling of the binding environment. Table 1 and Fig. 3 collectively display the comprehensive performance of the Dual-LAO method, covering the key statistics (R2 and RMSE) across all four chemical systems, including both edgewise (ΔΔGCAL) and pairwise (ΔGCAL) RMSE values. Table 1Summary of relative binding free energy (RBFE) results for all chemical systemsSystemPDBNo. compoundsNo. edgesR2Edgewise RMSEPairwise RMSEPWWP16G2B12300.620.680.46BRD45I888110.760.420.64P381W7H570.770.420.82CHK13U9N580.480.310.58Total-30560.900.520.56Performance statistics of the Dual-LAO method compared to experimental data across all four chemical systems: PWWP1, BRD4, P38, and CHK1. The table includes the protein data bank (PDB) identifier for the structure used, the total number of compounds and transformations (edges) in the set, the coefficient of determination (R2) computed on ΔΔG, and the RMSE calculated on both edgewise (ΔΔGCAL) and pairwise (ΔGCAL) free energy differences. All RMSE values are given in kcal.mol−1. Summary of relative binding free energy (RBFE) results for all chemical systems Performance statistics of the Dual-LAO method compared to experimental data across all four chemical systems: PWWP1, BRD4, P38, and CHK1. The table includes the protein data bank (PDB) identifier for the structure used, the total number of compounds and transformations (edges) in the set, the coefficient of determination (R2) computed on ΔΔG, and the RMSE calculated on both edgewise (ΔΔGCAL) and pairwise (ΔGCAL) free energy differences. All RMSE values are given in kcal.mol−1. Experimental vs. calculated absolute binding free energies derived from the RBFE network fit for all studied systems. The calculated ΔGCAL results and associated errors represent the mean ± standard error derived from the weighted least-squares (WLS) fitting of the RBFE network. The dark and light shaded regions indicate ± 1 kcal.mol−1 and ± 2 kcal.mol−1 deviations from the experimental values, respectively. The color bar displays the absolute difference between the experimental and computed values, ∣ΔGEXP−ΔGCAL∣. Multi-site hydration free energies for benzene derivatives We applied our multiple-topology approach to the computation of hydration free energies (HFE) of six molecules deriving from benzene, as a proof of concept. Similarly to the difference in binding affinity as calculated in Eq. (1), here the difference in hydration free energy is computed as the difference between the free energies associated to the transformation of L1 into L2 in the solvent phase and in vacuum, such that: HFEs were computed on all fifteen possible pairs of molecules drawn from the chosen pool of six. Inputs were prepared by aligning molecules along the six carbons of the conjugated cycle. Multiple-topology transformations were explored over the course of 15 ns simulation. As reference, HFEs were computed using the dual-topology method, with Dual-LAO. For each transformation, simulations of 2 ns were used for the solvent phase alchemical transformation, and 1 ns for the molecule in vacuum. For all transformations (dual and multiple topology), three replicas were run. A representation of the molecules (Supplementary Fig. S26) as well as a comparison of the multi-site results with dual-site values (Supplementary Fig. S27) are available in the Supplementary information. Exploration of the expanded λ-space is shown in Supplementary Fig. S28, and demonstrates the ability of the method to navigate throughout the various edges of the alchemical transformations graph. The RMSE (dual vs. multiple topology) for all RBFE values was 0.22 kcal.mol−1, and the mean average error (MAE) 0.14 kcal.mol−1, with an excellent correlation coefficient R2 of 1.00. This preliminary result indicates the fidelity of the λ-jumping scheme that was used, and is intended to serve as a proof-of-concept for future methodological developments. In future work, we plan to apply this method to binding free energies and will assess in particular its potential gain in efficiency (we expect exploration to be enhanced by the jumps) and in accuracy (the explicit exploration of cycles during such simulations should help more rigorous cycle closure when compared to sequences of independent simulations). Conclusion We presented Dual-LAO, a unified framework that addresses several long-standing bottlenecks in alchemical free-energy calculations by integrating Lambda-ABF-OPES, dual-topology, and dual-DBC restraints. The methodology’s success and accelerated convergence across diverse benchmarks arise from a synergistic mitigation of sampling and structural instabilities. By utilizing a dual-topology architecture, we circumvent the bonded-term complexities and numerical collapses typical of dummy-atom mapping, providing a smoother alchemical landscape for continuous λ evolution. This structural stability is complemented by dual-DBC restraints, which prevent entropic sampling divergence in decoupled states, and the high-efficiency ABF+OPESe sampling engine. By adaptively biasing the configurational probability distribution, the framework effectively treats solvation dynamics and protein plasticity as orthogonal degrees of freedom. This allows for a concerted relaxation of the system that resolves the hysteresis traditionally associated with complex active-site rearrangements. Ultimately, Dual-LAO offers a numerically robust and rapidly converging approach, providing a practical path toward high-precision binding affinity predictions in challenging drug discovery targets. Used in conjunction with the AMOEBA polarizable force field, this strategy achieves chemical accuracy (overall RMSE of 0.51 kcal.mol−1) with a performance gain of more than an order of magnitude (over 15-fold acceleration) compared to state-of-the-art techniques. We demonstrated that the method is able to robustly tackle complex situations for which traditional approaches struggle. For example, we show that Dual-LAO manages to reversibly rehydrate binding pockets upon ligand change (buried water displacement) without any specific tuning of the method. It also accurately handles other difficult perturbations including scaffold-hopping transformations, charge-changing, and ring-opening events with a marked improvement compared to state-of-the-art techniques. These advances make Dual-LAO a robust and highly efficient tool to drive hit-to-lead and lead-optimization in drug discovery. The multiple topology Dual-LAO extension was successfully implemented and validated, confirming its potential for network-based RBFE evaluations. Future work will focus on the systematic study of the sampling improvement obtained thanks to the multi-site extension (especially regarding cycle-closure), as well as the application of Dual-LAO to machine-learning foundation models such as FeNNix-Bio160. Methodology The RBFE is calculted following Eq. (1) (see ref. 20 for the original thermodynamical cycle) which requires the computation of alchemical free energy components (ΔGcomplex and ΔGsolv). These components correspond to the transformation of a starting ligand L1 into a target ligand L2. Alchemical parameter λ is used to scale key interactions in the system’s Hamiltonian such that λ = 0 corresponds to L1 fully interacting and λ = 1 corresponds to L2 fully interacting (with L1’s intermolecular interactions fully deactivated). Enhanced alchemical sampling: Dynamic λ and Lambda-ABF-OPES In contrast to traditional discrete sampling techniques such as TI or FEP, utilizing fixed λ-points, this work treats λ as a continuous dynamical variable within an extended Lagrangian framework. In this approach, λ is allowed to explore the alchemical space [0, 1] dynamically as a formal degree of freedom. This alchemical coordinate defines the transformation pathway through which electrostatic and van der Waals interactions are decoupled. Thus, while the evolution of λ is continuous and dynamic, it follows a predefined functional mapping (interactions are scaled according to a specific schedule) to maintain numerical stability and ensure a singularity-free transformation across the entire range. The resulting free energy profile is subsequently reconstructed from the potential of mean force (PMF) calculated along this continuous trajectory. To ensure the required accuracy and rapid convergence, it is computed with the Lambda-ABF-OPES framework47, which combines: Adaptive biasing force (ABF): The Lambda-ABF scheme applies an adaptive biasing force to the fictitious particle λ. This force enables proper sampling of the λ space by flattening the free energy landscape along the alchemical path. The resulting PMF is estimated by integrating the average force accumulated in 100 bins (Δλ = 0.01, see the alchemical sampling parameters and discretization section (Supplementary methods) for further detail) using the TI estimator. Adaptive biasing force (ABF): The Lambda-ABF scheme applies an adaptive biasing force to the fictitious particle λ. This force enables proper sampling of the λ space by flattening the free energy landscape along the alchemical path. The resulting PMF is estimated by integrating the average force accumulated in 100 bins (Δλ = 0.01, see the alchemical sampling parameters and discretization section (Supplementary methods) for further detail) using the TI estimator. OPES-explore (OPESe): The Lambda-ABF scheme does not explicitly encompass capabilities to overcome energetic barriers orthogonal to the chosen collective variable (λ). This is addressed by the addition of the exploratory version of the on-the-fly probability enhanced sampling (OPES)16,17 method, known as OPESe, a metadynamics-based method that aims at enhancing configurational sampling by adding a penalty to the already explored regions of the λ interval (See the alchemical sampling parameters and discretization section (Supplementary methods) for further detail). OPES-explore (OPESe): The Lambda-ABF scheme does not explicitly encompass capabilities to overcome energetic barriers orthogonal to the chosen collective variable (λ). This is addressed by the addition of the exploratory version of the on-the-fly probability enhanced sampling (OPES)16,17 method, known as OPESe, a metadynamics-based method that aims at enhancing configurational sampling by adding a penalty to the already explored regions of the λ interval (See the alchemical sampling parameters and discretization section (Supplementary methods) for further detail). To further accelerate convergence, Lambda-ABF-OPES is implemented using a multiple-walker parallelization scheme. In this approach, several independent replicas (walkers) of the system are simulated simultaneously. Crucially, these walkers share and contribute to a global bias potential and force-accumulation grid. This collective information exchange allows the algorithm to map the PMF more efficiently than a single-trajectory approach, as different walkers can explore distinct regions of the configurational space concurrently. In this study, we typically employed 4 walkers for the complex phase and 4 for the solvent phase. Consequently, a reported convergence time of 4 ns/walker corresponds to a total aggregate sampling of 16 ns for complex and 8 ns for solvent. However, because these walkers run in parallel, the total wall-clock time remains equivalent to that of a 4 ns MD simulation, representing a significant throughput advantage over serial alchemical methods. This combination of efficient λ-space sampling from Lambda-ABF and enhanced exploration across orthogonal barriers (such as complex solvent or water exploration47,55) from OPESe constitutes the Lambda-ABF-OPES method. Dual-topology scheme and restraints The transformation employs a dual-topology scheme20,29 where both ligands (L1 and L2) are simultaneously present. This approach simplifies input preparation (as no dummy atoms are required) and the computational cost of the extra atoms is negligible. MCS restraints: The Maximal common substructure (MCS) is the substructure, common to both ligands, with the largest number of atoms or bonds. To ensure both ligands maintain a consistent conformation relative to their common parts, the MCS is identified using the FMCS algorithm61, and a set of dual-topology restraints (harmonic wall restraints) is applied to corresponding atoms in the MCS. MCS restraints: The Maximal common substructure (MCS) is the substructure, common to both ligands, with the largest number of atoms or bonds. To ensure both ligands maintain a consistent conformation relative to their common parts, the MCS is identified using the FMCS algorithm61, and a set of dual-topology restraints (harmonic wall restraints) is applied to corresponding atoms in the MCS. Binding pose restraints: To prevent the ligands from escaping the binding pocket while its intermolecular interactions are scaled down, a Dual-DBC scheme was implemented to positionally restrain the ligands, as detailed below. Binding pose restraints: To prevent the ligands from escaping the binding pocket while its intermolecular interactions are scaled down, a Dual-DBC scheme was implemented to positionally restrain the ligands, as detailed below. Alchemical schedule The alchemical transformation from L1 to L2 is controlled using scaling factors for the intermolecular interactions (λELEX and λV DWX). The continuous λ schedule is divided into three steps, as illustrated in Fig. 4a: Scaling of the intermolecular interactions and DBC restraint as a function of λ. a Scaling of the intermolecular interactions for ligand L1 (λELEL1, λV DWL1) and ligand L2 (λELEL2, λV DWL2) as a function of λ. b Dual-DBC target profiles (dtarget1 and dtarget2). Electrostatic decoupling (λ = [0, 0.2]): Electrostatic interactions of ligand L1 are turned off by scaling down linearly its permanent multipoles and polarizability. Electrostatic decoupling (λ = [0, 0.2]): Electrostatic interactions of ligand L1 are turned off by scaling down linearly its permanent multipoles and polarizability. VDW switch (λ = [0.2, 0.8]): Van der Waals interactions for L1 are linearly turned off, and the same interactions for L2 are linearly turned on. VDW switch (λ = [0.2, 0.8]): Van der Waals interactions for L1 are linearly turned off, and the same interactions for L2 are linearly turned on. Electrostatic coupling (λ = [0.8, 1.0]): Electrostatic interactions of ligand L2 are activated by scaling up linearly its permanent multipoles and polarizabilities. Electrostatic coupling (λ = [0.8, 1.0]): Electrostatic interactions of ligand L2 are activated by scaling up linearly its permanent multipoles and polarizabilities. It can be noted that over the “VDW switch" section, all electrostatic interactions of both ligands are turned off. If ligands bound pose is mainly stabilized through electrostatics, positional drift can then be observed. This caveat is addressed through the use of DBC restraints, as was the case in the original Lambda-ABF-OPES approach47, described in the following paragraphs. Dual-DBC restraint scheme To counteract positional drift during the intermediate λ states, we apply a flat-bottom harmonic restraint on the DBC of each ligand, using two independent, λ-dependent targets, dtarget1 and dtarget2, as shown in Fig. 4b. The restraint acts only if DBC is higher than the target. The target are designed to have a minimum Dmin=2.0 Å and a maximum Dmax=10.0 Å. Restraints are maintained at Dmin, but are gradually released near the end point corresponding to the decoupled ligand of interest by increasing the target to Dmax. As explained in a further section, the value of Dmin is chosen so that the binding mode of the fully coupled ligand of interest is not biased. This ensures that the final (fully coupled and fully decoupled) states are unbiased (see Drivers of Protocol Efficiency and Stability section (Supplementary methods) for further detail). Implementation The Dual-LAO methodology was implemented within the Tinker-HP molecular dynamics software package50, leveraging the capabilities of the Colvars library62 for managing the enhanced sampling Collective Variables and the specialized restraints. Simulation details Plain MD simulations In this work, all molecular dynamics (MD) and advanced sampling simulations were performed using the Tinker-HP molecular dynamics package (Version 1.2)50,63, which includes the integrated Colvars64 library for Dual-LAO RBFE calculations. Simulations utilized the polarizable AMOEBA force field65–68 for the protein, solvent (water), and counter-ions. Ligand parameters were derived using the Poltype package69 (see Computational Cost and Performance Metrics section (Supplementary methods) for further detail). Each system was constructed within a cubic water box, sized to ensure a minimum separation of 12 Å between the ligand and any box boundary. The solvent was set to an aqueous NaCl solution at a concentration of 0.15 M to ensure physiological relevance and system neutrality. A multi-stage protocol was employed to prepare the systems for production dynamics. First, the system underwent energy minimization via the Tinker-HP minimize utility, followed by a two-step heating phase from 200 K to 298 K under the NVT (canonical) ensemble. The first two steps utilized the RESPA integrator with a 2.0 fs time step: Step 1 (4 ns) involved applying positional restraints to all protein, ligand, and structural X-ray water atoms; in Step 2 (4 ns), these restraints were relaxed to act only on the heavy backbone atoms of the protein, maintained with a force constant of 10 kcal.mol−1. Å−2. Next, the system underwent a two-step relaxation phase. Step 3 (5 ns) involved further equilibration in the NVT ensemble using the three-level BAOAB-RESPA1 multiple-timestep integrator70, where the outer time step was 10 fs, with intermediate and inner steps of 10/3 fs and 1 fs, respectively. For this step, restraints were applied only to Cα atoms using a reduced force constant of 1 kcal.mol−1. Å−2. Finally, Step 4 (5 ns) conducted the final equilibration under the NPT (isothermal-isobaric) ensemble at 1 atm pressure, maintaining the Cα restraints from the previous step. Following equilibration, a 20 ns production run was executed for each ligand complex in the NPT ensemble. The BAOAB-RESPA1 Langevin integrator was used with a 10 fs outer time step. Temperature control in non-Langevin steps was ensured by the Bussi thermostat71, while pressure was regulated by the Berendsen barostat72. Van der Waals interactions employed a 12 Å cutoff. Electrostatic interactions were computed using the Particle Mesh Ewald (PME) method73 with a 7 Å real-space cutoff. Induced dipoles were determined via a Preconditioned Conjugate Gradient (PCG) solver74 to a convergence tolerance of 1 × 10−5 Debye. The computational cost associated with force field parameterization and production simulations is detailed in the Computational cost and performance metrics section of the Supporting information. For representative systems ( ≈ 60 000 atoms), production throughput reached up to 55 ns per day depending on the hardware architecture (see Table S4 in the SI). Dual-LAO simulations For all alchemical simulations, including R-group, buried water displacement, scaffold-hopping, and fragment-like transformations, we employed the Dual-LAO method. We ran four walkers for each transformation, with 4 ns per walker for the complex phase and 2 ns per walker for the solvent phase. All other simulation parameters matched those of the plain MD simulations, with one exception: we used the BAOAB-RESPA integrator70 with a 2 fs time step. DBC selection As described in Dual DBC Restraint Scheme section, we employed the DBC49 to restrain each L1 and L2 within its respective binding pose during complex alchemical transformations. The DBC coordinate quantifies the ligand’s deviation from its target pose by calculating the RMSD of a selected set of ligand atoms (noted LA) after optimal superposition onto a reference set of receptor binding site atoms (noted RB). This single variable efficiently captures positional, orientational, and conformational fluctuations of the ligand relative to its binding environment. The selection of LA and RB atoms is crucial for defining a meaningful DBC restraint. To identify the least mobile parts of the ligand, we monitored the Root Mean Square Fluctuation (RMSF) of all heavy ligand atoms over a 20 ns conventional MD simulation of each individual complex. For the selection of the ligand atoms (LA) used in the DBC restraint, we applied a uniform Root Mean Square Fluctuation (RMSF) cutoff of 1.0 Å across all ligands and targets under study. This ensures consistency in restraining the least flexible portions of the ligands undergoing alchemical coupling/decoupling. The RB set consisted of the Cα atoms of the protein that fell within 6 Å of the ligand and exhibited an intrinsic flexibility below approximately 0.8 Å during the production MD. The use of these uniform RMSF cutoffs serves to standardize the protocol and minimize operator bias across diverse targets. By selecting atoms based on a fixed flexibility threshold rather than manual inspection, the process becomes largely automated and reproducible. The sensitivity of the resulting free energies to the exact composition of the LA and RB sets is minimized by the nature of the DBC variable itself; as a collective RMSD-based restraint, it is inherently more stable than restraints based on individual atom pairs. Consequently, provided that the selection captures the relatively rigid core of the binding interface, minor variations in the atom sets do not significantly alter the sampled phase-space volume. To ensure that the DBC restraint does not introduce artificial bias during the alchemical simulation, we implemented a dual-DBC restraint strategy as explained earlier. Specifically, a minimum DBC value (DBCmin) was set at 2 Å. This DBCmin value was chosen to be higher than the upper bound of DBC coordinate value observed over the plain MD production run for each ligand, which was monitored using the Colvars library. A flat-bottomed harmonic restraint was applied to the DBC coordinate of both L1 and L2 during the Dual-LAO simulations, following the DBC restraint schedule. This restraint employed a high force constant of 100 kcal.mol−1. Å−2 and was active only when the calculated DBC value exceeded the predetermined DBC cutoff. Multiple topology RBFEs The dual-topology framework is readily extensible to a multiple-topology framework (akin to the separated topologies approach75), which allows several alchemical transformations to be explored over the course of a single simulation. In a typical drug discovery campaign, the goal is to explore a network of transformations connecting a series of ligands (L1, L2, L3, … ). This network is defined by the user (e.g., using methods like LoMap54) and is naturally represented as a graph. In this representation, each node denotes a molecule, and each edge represents the alchemical transformation between two connected molecules (not all pairs of molecules must be connected). Crucially, while the dual-topology framework explores each edge of this graph during an independent simulation, the multiple-topology framework allows all connected transformations to be explored during a single, extended simulation, following the procedure detailed below. Any alchemical state of the multiple-topology simulation can be characterized with the triplet (Li, Lj, λ), where Li and Lj are the two molecules undergoing alchemical transformation, and λ, comprised in [0; 1], describes the evolution of the transformation (Following this notation, the dual-topology case is simply described as the triplet (L1, L2, λ)). If λ = 0, then Li is fully coupled (and Lj fully decoupled), and if λ = 1, Li is decoupled, Lj coupled. The Dual-LAO simulation is started on an (arbitrarily chosen) initial edge (Li, Lj) of the graph. Just like in the dual-topology case, the value of λ is used to explore the alchemical schedule and define scaling terms for the VDW and ELE interactions of Li and Lj, while all other ligands remain fully decoupled. In the dual-topology scheme, the value of λ is kept within the [0; 1] segment using reflexive boundary conditions. In the multiple-topology scheme, after applying reflexive boundary conditions, a special procedure is applied to enable exploration of other edges: (i)The node corresponding to the reflection value is identified. The node corresponding to the reflection value is identified. (ii)All edges connected to this node are identified. All edges connected to this node are identified. (iii)A random integer between 1 and the number of connected edges is drawn. This chooses the next edge of the network to explore, and defines a new triplet-state to explore. A random integer between 1 and the number of connected edges is drawn. This chooses the next edge of the network to explore, and defines a new triplet-state to explore. This process allows the simulation to explore a new alchemical state whenever λ reaches (or exceeds) 0 or 1, which corresponds a “fully coupled” alchemical state. Let us illustrate the process with an example: the simulation starts at (L1, L2, λ = 0). After some time, λ reaches 1, which corresponds to the ligand L2 being fully coupled. Let’s assume L1, L3, L4 are connected to L2 in our graph. A random integer is drawn between 1 and 3 (say, 2). This chooses the second connected edge to continue exploration, here L2 to L3. The new triplet-state becomes (L3, L2, λ = 1), and now alchemical transformation from L2 to L3 is explored. With minimal extra input from the user (namely, the choice of a network via a connectivity matrix), this method allows to compute several RBFEs over the course of a single simulation. Supplementary information Supplementary Information for Dual-LAO for calculating fast and robust relative binding free energies of simple and complex transformations Description of Additional Supplementary Files Supplementary Data 1 Acknowledgements The authors would like to extend their gratitude to Oliver Adjoua and Florent Hédin for their unwavering support in all matters of code development, as well as Chengwen Liu for his help with AMOEBA parametrization. This work was supported by computing grants from the Grand Équipement de Calcul Intensif (GENCI), Institut du Développement et des Ressources en Informatique (IDRIS), and Centre Informatique de l’Enseignement Supérieur (CINES), France, under grant numbers AD010715770 and AD010716167R1. Author contributions Designed research: L.L.; Performed research: N.A., F.A., L.L.; Performed simulations: N.A., F.A.; Contributed analytic tools: N.A., F.A.; Contributed new code: F.A., L.L.; Analyzed data: N.A., F.A., J.H., L.L., J.-P.P.; Wrote the paper: N.A., F.A., J.H., L.L., J.-P.P. Peer review Peer review information Communications Chemistry thanks Marcus Wieder and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Data availability The Tinker-HP and colvars input files used, and the structures are available on GitHub at: https://github.com/ansarinarjes/RBFE-L-ABF-OPES.git. Code availability The code used during the study is available on GitHub: https://github.com/TinkerTools/tinker-hp; https://github.com/TinkerTools/poltype2; https://github.com/Colvars/colvars. Competing interests L.L. and J.-P.P. are co-founders and shareholders of Qubit Pharmaceuticals. The remaining authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Narjes Ansari, Félix Aviat. Contributor Information Narjes Ansari, Email: narjesa@qubit-pharmaceuticals.com. Félix Aviat, Email: felix.aviat@qubit-pharmaceuticals.com. Louis Lagardère, Email: louis.lagardere@sorbonne-universite.fr. Supplementary information The online version contains supplementary material available at 10.1038/s42004-026-02022-6. References Associated Data Supplementary Materials Supplementary Information for Dual-LAO for calculating fast and robust relative binding free energies of simple and complex transformations Description of Additional Supplementary Files Supplementary Data 1 Data Availability Statement The Tinker-HP and colvars input files used, and the structures are available on GitHub at: https://github.com/ansarinarjes/RBFE-L-ABF-OPES.git. The code used during the study is available on GitHub: https://github.com/TinkerTools/tinker-hp; https://github.com/TinkerTools/poltype2; https://github.com/Colvars/colvars.