Alchemical Transfer Method (ATM)

• ATM is a dual-topology, coordinate-perturbation method that computes absolute and relative binding free energies by rigidly translating ligand positions.
• It streamlines calculations by avoiding atom mapping, dummy atoms, and force field parameter interpolation, operating with unmodified molecular mechanics engines.
• Validated on large benchmarks, ATM robustly handles challenging transformations such as charge changes and scaffold hops, making it valuable for drug discovery and molecular design.

The Alchemical Transfer Method (ATM) is a dual-topology, coordinate-perturbation alchemical approach for computing absolute and relative binding free energies (ABFE/RBFE) of ligands to biomolecular receptors in explicit solvent. Unlike classical parameter-interpolation methods, ATM accomplishes the transformation between bound and unbound (or between two bound ligands) by rigidly translating ligand coordinates in Cartesian space, rather than scaling force field parameters. This design eliminates the need for atom mapping, dummy atoms, or the splitting of electrostatic and van der Waals transformations, and is implemented without any modification to core molecular mechanics engines. ATM has been validated on large-scale benchmarks, shows robust performance across diverse chemical transformations—including challenging scaffold hopping and charge-changing perturbations—and provides a general, streamlined, and force-field agnostic workflow for free energy calculation in molecular design (Chen et al., 2023, Azimi et al., 2024, Azimi et al., 2021, Wu et al., 2021, Gallicchio, 2024, Zariquiey et al., 2023).

1. Theoretical Foundations and Statistical Mechanics

ATM generalizes the thermodynamic cycle underlying double-decoupling methods (DDM). It replaces alchemical parameter changes with a physically intuitive coordinate transformation, mapping the bound ligand into bulk solvent (ABFE) or swapping two ligands (RBFE) via rigid-body translation.

The central idea is to construct a λ-dependent potential:

$U_\lambda(x) = U_0(x) + W_\lambda[u(x)]$

where $U_0(x)$ is the potential at the starting configuration, $u(x) = U_1(x) - U_0(x)$ is the total perturbation energy of the coordinate shift, and $W_\lambda(u)$ is a non-linear interpolation ("softplus" or linear) such that $W_0=0$ and $W_1=u$. This results in a continuous alchemical pathway parameterized by λ, connecting the two physical end states through an explicit, engine-agnostic dual-topology transformation (Azimi et al., 2024, Chen et al., 2023, Wu et al., 2021, Gallicchio, 2024).

For ABFE, the free energy of binding is given by:

$$\Delta G^\circ_{\rm bind} = \Delta G_b^* + \Delta G^\circ_{\rm site}$$

with $\Delta G_b^*$ as the excess free energy from the coordinate perturbation (ATM calculation) and $$\Delta G^\circ_{\rm site} = -k_B T \ln (C^\circ V_{\rm site})$$ the standard-state correction. For RBFE, ATM computes:

$$\Delta\Delta G = -k_B T\ln\left\langle e^{-\beta u_{\rm RBFE}}\right\rangle_0$$

where $u_{\rm RBFE}$ is the total perturbation energy associated with swapping two ligands (Azimi et al., 2021, Zariquiey et al., 2023).

Potential Distribution Theory (PDT) provides an exact theoretical underpinning, showing that the probability density of perturbation energies along the ATM path is a convolution of the corresponding densities from the DDM, and the resulting free energy is formally identical regardless of the chosen path, provided it is a function of $u(x)$ only (Azimi et al., 2024).

2. Methodology and Workflow

ATM operates by preparing a single simulation box containing the receptor and either one (ABFE) or two (RBFE) ligands. Key steps include:

• System setup: Precise alignment and placement of ligands in binding pocket and bulk, using automated pipelines for explicit solvation, protein and ligand parameterization (FFEngine, GFN2-xTB/BCC, GAFF2, QM geometry optimization).
• Coordinate displacement: For ABFE, the ligand is shifted from the binding site to bulk solvent. For RBFE, two entire ligands are swapped via translation defined by a displacement vector $h$ or $d$.
• Alchemical path: Lambda windows are defined using a softplus or linear mixing schedule, typically with 11-22 λ-states per leg, to interpolate smoothly from end states through a symmetric intermediate ($U_{1/2} = (U_0 + U_1)/2$).
• Replica-exchange: Hamiltonian replica-exchange enhances sampling across λ, with exchanges attempted every 5–40 ps; production trajectories aggregate to 5–40 ns per replica depending on system size (Chen et al., 2023, Gallicchio, 2024, Azimi et al., 2021).
• Restraints: Flat-bottom harmonic restraints on ligand site, protein atoms, and alignment references are employed to maintain positional and orientational integrity.
• Free energy estimation: Multistate reweighting methods (UWHAM, MBAR) are applied to extract $\Delta G$ and associated uncertainties (Zariquiey et al., 2023).

No force field parameter interpolation, dummy atoms, or splitting of interaction types is needed; all nonbonded and bonded interactions are evaluated natively by the molecular dynamics engine.

3. Extensions and Specializations: ATS and Selectivity Calculations

ATM has been extended to address the scaling limitations in RBFE with the Alchemical Transfer with Coordinate Swapping (ATS) method (Gallicchio, 2024). ATS implements the coordinate-swap only for the region of chemical difference between two ligands or protein/peptide mutants, leaving the common core unchanged. This approach:

• Drastically reduces the variance in the perturbation energy u for large ligands or macromolecular mutations.
• Enables RBFE calculations on peptides, nucleic acids, and mutants by restricting the swap to variable sidechains or sequence positions.
• Maintains the "drop-in" usability and engine agnosticism of ATM.

For binding selectivity (e.g., between two receptors), ATM protocols for receptor hopping and swapping enable direct calculation of selectivity free energies by simultaneous alchemical transfer of ligands between multiple receptors (identity-agnostic, handles arbitrary sequence or structural variation), with DiffNet providing network-based self-consistent inference of selectivity profiles (Azimi et al., 2024).

4. Implementation and Computational Performance

ATM is implemented as an OpenMM plugin—the ATM MetaForce and AToM-OpenMM packages—which call unmodified OpenMM energy/force routines. All routines are compatible with any chemical topology and force field (classical, polarizable, QM/MM, or ML-derived), and straightforwardly generalize to RBFE, ABFE, selectivity, and mutational scanning.

• Force fields: Automated with FFEngine (GAFF2, GFN2-xTB/BCC, TeraChem, fallback to quantum-derived or bespoke).
• Sampling: Typically 11–24 λ-windows, with 5–40 ns per window, resulting in total aggregate sampling of ~110 ns per ligand pair in benchmark studies.
• Performance: Empirically, ATM achieves per-target RBFE Pearson r ≈ 0.44–0.63, AUE ≈ 1.1 kcal/mol, comparable to FEP+ and other commercial pipelines, at similar computational cost (1.5–2 days per ΔΔG on modern GPUs) (Chen et al., 2023, Zariquiey et al., 2023).
• Convergence: Enhanced by replica exchange and smooth alchemical scheduling; challenging perturbations (large R-group, charge change, scaffold hops) may require intermediate states or metadynamics.

5. Validation, Benchmarking, and Performance Analysis

ATM has been validated against a series of host–guest and protein–ligand benchmark sets:

• Merck RBFE benchmark: 8 protein targets, 550 ligand pairs (384 small R-group, 43 large R-group, 66 charge switches, 60 scaffold hops). Pearson r = 0.60, Kendall τ = 0.46, AUE = 1.11 kcal/mol versus experiment, with larger fluctuations for highly challenging transformations (Chen et al., 2023).
• SAMPL host–guest challenges: Absolute and relative binding free energies agree within RMSD ≈ 0.8 kcal/mol, cycle closures < 0.25–0.6 kcal/mol, and statistical uncertainties competitive with best published physical-pathway and alchemical methods (Azimi et al., 2021, Wu et al., 2021).
• Selectivity protocols: Direct receptor hopping/swapping yield binding selectivity RMSD ≈ 0.3–0.5 kcal/mol versus reference ABFEs and experiments (Azimi et al., 2024).
• ATS: For large ligands and peptides, variance of RBFE is reduced, with comparable or slightly better convergence than whole-ligand ATM; mean pairwise hysteresis ≈ 0.6 kcal/mol in peptide mutagenesis (Gallicchio, 2024).
• Charge changes and scaffold hops: ATM handles net-charge changes and major topology transformations natively, though very large perturbations can produce high energy barriers or sampling gaps and require additional methodological care.

6. Strengths, Limitations, and Future Perspectives

ATM offers significant simplifications and broad applicability:

Strengths

• Unified treatment of all perturbation types (no special handling for charge changes or scaffold hops).
• Simultaneous dual-topology transformation with coordinate swaps independent of atom mapping or parameter interpolation.
• Compatible with any unmodified force field or potential.
• No requirement for custom soft-core pair potentials or energy routines.
• Open-source implementations and streamlined automated workflows.

Limitations

• Statistical fluctuations are larger for very large structural changes, especially those involving dramatic charge reorganization or conformational reorganization.
• For strongly size-mismatched or flexible systems, intermediate states, enhanced sampling, or alternative λ-schedules may be necessary.
• Sampling demands remains high for transformations with high intermediate energy barriers (>40 kcal/mol), and certain charge-changing moves can induce slow conformational transitions or multiple poses.

Future developments include automated path optimization, intermediate state insertion, sampling acceleration (metadynamics, longer exchange windows), and extension to systematic mutational scanning and binding selectivity profiling (Chen et al., 2023, Gallicchio, 2024, Azimi et al., 2024).

7. Comparative Context and Practical Applications

ATM and its ATS extension redefine the landscape of alchemical free energy methods by providing a simple, general-purpose, and robust protocol for both small-molecule and large macromolecular binding free energies:

• Compared to DDM/FEP: Comparable accuracy (~1 kcal/mol), less setup complexity, and broader applicability to topologically divergent transformations or macromolecular mutations (Zariquiey et al., 2023, Azimi et al., 2021).
• Practical utility: Routine production-level RBFE for lead optimization, scaffold‐hopping, charge-changing transformations, selective binding analysis for receptor panels, and computational optimization of large ligands or protein/nucleic acid mutations (Gallicchio, 2024, Azimi et al., 2024).
• Force-field flexibility: No dependence on hard-coded or patched force fields; plug-and-play with new energy models, including neural network and quantum mechanical potentials.

ATM thus enables a broad spectrum of scientific and drug discovery applications while reducing the technical burden typically imposed by traditional alchemical protocols. Its robust theoretical foundation, operational simplicity, and validated performance underpin its growing adoption in computational chemistry and molecular design (Chen et al., 2023, Azimi et al., 2024, Gallicchio, 2024).