Free Energy Perturbation (FEP)

FEP exploits the fact that free energy is a state function. Instead of calculating the absolute binding energy directly, it computes the difference along a non-physical 'alchemical' path.

• Closed Cycle: ΔG_binding(A→B) = ΔG_solvated(A→B) - ΔG_bound(A→B)
• Cancellation of Errors: The largest systematic errors cancel out, yielding highly accurate relative predictions.
• Pathway: Ligand A is gradually 'morphed' into Ligand B through a series of intermediate, non-physical λ-windows.

The alchemical transformation is discretized into a series of intermediate states defined by a coupling parameter λ (0 to 1). At each window, extensive molecular dynamics sampling is performed.

• λ = 0: System represents the initial state (Ligand A).
• λ = 1: System represents the final state (Ligand B).
• Soft-Core Potentials: Special functional forms prevent 'end-point catastrophes' where atoms overlap and energies diverge.
• Replica Exchange: Often used to enhance sampling by allowing neighboring λ-windows to swap configurations.

The most common estimator used to stitch together the energy differences collected at each λ-window into a single free energy value.

• Statistical Rigor: BAR minimizes the statistical variance of the free energy estimate by optimally combining forward and reverse work distributions.
• Multistate Bennett Acceptance Ratio (MBAR): A modern extension that uses all data from all λ-windows simultaneously, providing the lowest possible variance estimator.
• Overlap Requirement: Accurate results require sufficient phase space overlap between adjacent λ-windows.

While standard FEP calculates relative binding energies, Absolute Binding Free Energy (ABFE) calculations compute the energy to transfer a ligand from bulk solvent to the binding pocket.

• Restraint Strategy: A series of conformational, orientational, and translational restraints are applied and analytically corrected for to prevent the ligand from wandering away during decoupling.
• Double Decoupling: The ligand is alchemically annihilated in both the binding pocket and in bulk solvent.
• Computational Cost: ABFE is significantly more expensive than relative FEP but requires no congeneric series.

FEP calculations are rarely run in isolation. They are organized into perturbation networks to connect a chemical series.

• Star Maps: A central compound is perturbed to all others. Simple but prone to hub-centric errors.
• Cycle Closures: Running perturbations in closed cycles (A→B, B→C, C→A) provides an internal consistency check; the sum of ΔG should be zero.
• Optimal Edge Selection: Algorithms select the minimal set of perturbations with the highest predicted overlap to map a full library efficiently.

Historically limited by CPU clusters, modern FEP workflows leverage GPU-accelerated MD engines to achieve throughput suitable for drug discovery timelines.

• Engine Examples: OpenMM, Amber (pmemd.cuda), and Desmond.
• Throughput: Modern platforms can evaluate hundreds of compound pairs per week.
• Cloud Integration: Orchestration tools like Orion or AWS ParallelCluster enable elastic scaling of FEP campaigns, turning a supercomputer problem into an on-demand service.

A non-physical thermodynamic path that mutates one ligand into another through a series of intermediate, non-physical states. In FEP, this is achieved by scaling the non-bonded interactions of a perturbation region while leaving the common scaffold unchanged.

• Single topology: A unified hybrid representation where atoms appear, disappear, or change atom type.
• Dual topology: Both end-state ligands are present simultaneously, with their interactions scaled inversely.
• Avoids the physical unbinding/rebinding barrier, enabling convergence in finite simulation time.

A closed path of state functions that exploits the fact that free energy is a state function independent of path. The binding free energy difference (ΔΔG) between two ligands is computed via the non-physical alchemical legs rather than the direct physical binding processes.

• Horizontal legs: Physical binding of ligand A and ligand B to the target.
• Vertical legs: Alchemical mutation of A→B in solvent and A→B in the protein complex.
• Cycle closure: ΔΔG = ΔG_bind(B) - ΔG_bind(A) = ΔG_protein(A→B) - ΔG_solvent(A→B).

A continuous coupling parameter (λ) that interpolates the Hamiltonian between the initial (λ=0) and final (λ=1) states. The alchemical path is discretized into λ-windows, with independent equilibrium simulations run at each window.

• Soft-core potentials: Modified Lennard-Jones and Coulomb functions that prevent endpoint singularities when atoms are created or annihilated.
• Lambda scheduling: Non-uniform spacing of windows, concentrated in regions of high dH/dλ curvature.
• Hamiltonian replica exchange: λ-windows periodically attempt to swap configurations, enhancing phase space sampling across the entire alchemical path.

The minimum-variance estimator for computing the free energy difference between two overlapping thermodynamic states. BAR iteratively solves a self-consistent equation using the work distributions from forward and reverse perturbations.

• Requires potential energy differences between neighboring λ-windows, not just the endpoints.
• Multistate Bennett Acceptance Ratio (MBAR) extends BAR to use all data from all λ-windows simultaneously, maximizing statistical power.
• MBAR provides a rigorous framework for estimating statistical uncertainty through asymptotic covariance matrix analysis.

An end-point free energy method that approximates binding free energy from only the unbound and bound state ensembles, bypassing intermediate alchemical states. Combines molecular mechanics energy, an implicit solvation model (GBSA or PBSA), and an entropy estimate.

• Computationally cheaper than FEP by 1-2 orders of magnitude.
• Lacks the rigorous conformational sampling of the alchemical path, leading to higher systematic error.
• Often used as a triage step before committing to full FEP calculations on a prioritized subset of compounds.

Methods that accelerate the exploration of the potential energy landscape to overcome kinetic trapping in local minima, which is critical for converging FEP calculations on systems with slow degrees of freedom.

• Replica Exchange with Solute Tempering (REST2): Selectively heats the ligand and binding pocket region without disrupting the bulk solvent.
• Metadynamics: Adds a history-dependent biasing potential to discourage revisiting previously sampled collective variable regions.
• Self-Adjusted Mixture Sampling (SAMS) : Adaptively adjusts biases to achieve a uniform distribution across a predefined order parameter, optimizing λ-window overlap.