ESMACS Protocol for Free Energy Estimation

• ESMACS is a physics-based computational workflow that uses ensemble MD and MM/PBSA to estimate protein–ligand binding free energies with quantifiable uncertainties.
• It employs two variants—CG-ESMACS for rapid screening and FG-ESMACS for detailed energetic refinement—tailoring its approach to different stages of drug discovery.
• The protocol integrates advanced force fields, explicit solvation models, and high parallelism to deliver reliable binding affinity predictions and effective compound ranking.

The Enhanced Sampling of Molecular dynamics with Approximation of Continuum Solvent (ESMACS) protocol is a physics-based computational workflow for the estimation of protein–ligand binding free energies. ESMACS employs an ensemble molecular dynamics (MD) approach combined with a Molecular Mechanics/Poisson–Boltzmann Surface Area (MM/PBSA) post-processing framework, leveraging high parallelism to deliver reproducible binding affinity predictions and quantifiable uncertainties. In large-scale drug discovery settings, ESMACS is utilized both for rapid compound prioritization and for high-precision energetic ranking of lead candidates, as exemplified in recent evaluations of AI-driven molecular modeling pipelines (Wan et al., 2 Mar 2026).

1. Ensemble Simulation Structure

ESMACS is built on two operational variants: Coarse-Grained ESMACS (CG-ESMACS) and Fine-Grained ESMACS (FG-ESMACS). CG-ESMACS is designed for throughput, deploying 10 independent replicas per ligand–protein complex within a 10 Å solvent buffer (production length typically 2–4 ns), without entropic correction, and is suitable for initial rank-order screening across extensive compound libraries (16,780 for 3CLPro; 21,702 for TNKS2 in the referenced study). FG-ESMACS, used for energetic refinement of top-ranked candidates, increases the ensemble size to 25 replicas per complex, extends the solvent buffer (standard 12 Å), and employs longer production runs (10 ns per replica), with no entropic component calculated. Each FG-ESMACS campaign comprises three independent pose sets for each ligand: (i) best docking-derived poses, (ii) Boltz-2 co-folded heavy-atom coordinates with geometrically added hydrogens, and (iii) Boltz-2 heavy-atom coordinates with hydrogens assigned via canonical SMILES topology. No restraints or alchemical transformations are introduced; ESMACS remains a fully non-alchemical, one-step post-processing protocol.

2. Force Fields and Solvation Parameters

All MD simulations in ESMACS are executed with NAMD (version 2.14 or later) under the standard AMBER protein force field (such as ff14SB). Ligand parameters are generated using OpenEye workflows: charges assigned via assigncharges.py (AM1–BCC or similar), protonation state fixed at pH 7.4 with FixpKa, and both bonded and nonbonded parameters conforming to GAFF2 conventions. The solvent environment is modeled explicitly using TIP3P water, with addition of neutralizing counter-ions and 0.15 M NaCl, both handled through automated NAMD tools.

3. Statistical Mechanics and Free Energy Estimation

Binding free energy ($\Delta G_{\rm bind}$) is evaluated post hoc on the MD ensemble via MM/PBSA as implemented in AmberTools. The total free energy of each system snapshot is given by $G = E_{\rm MM} + G_{\rm solv}$, with $E_{\rm MM}$ comprising all bonded, van der Waals, and electrostatic contributions, and $G_{\rm solv}$ partitioned into Poisson–Boltzmann (PB) or generalized-Born (GB) polar terms plus a nonpolar surface-area term ($\gamma\,\mathrm{SASA} + b$). $\Delta G_{\rm bind}$ is computed as

$$\Delta G_{\rm bind} \approx \langle G_{\rm complex} \rangle - \langle G_{\rm protein} \rangle - \langle G_{\rm ligand} \rangle$$

No entropic ($T \Delta S$) or interaction-entropy corrections are included, due to prior findings that such terms do not substantially alter compound rankings in this protocol. Statistical uncertainty (standard error of the mean) is obtained directly from the standard deviation ($\sigma$) over the replica $\Delta G$ values, divided by $G = E_{\rm MM} + G_{\rm solv}$0.

4. Protocol Optimizations and Variants

CG-ESMACS achieves high screening throughput at the expense of detailed precision, employing a reduced number of replicas, a smaller solvent buffer, and brief simulation trajectories. FG-ESMACS increases ensemble size and MD duration to support higher accuracy in the top-ranked subset. The use of three independent starting structure protocols for FG-ESMACS enables systematic evaluation of errors arising from initial docked pose and hydrogen assignment (either geometric or SMILES-based). No additional restraints, alchemical $G = E_{\rm MM} + G_{\rm solv}$1-windows, or post hoc entropic corrections are deployed. The entire protocol is engineered for maximal parallelism on large-scale HPC resources, enabling efficient ensemble execution across thousands of ligands.

5. Convergence Criteria, Error Quantification, and Scaling

The ensemble-based nature of ESMACS guarantees stabilization of mean binding free energies as the number of replicas increases, with prior benchmarks reporting root-mean-square (RMS) fluctuations of approximately 0.5 kcal/mol for 25 replicas with 4 ns production per replica. In the fine-grained protocol applied to the top-100 ligands (per target), a total of 7,500 simulations (100 ligands × 25 replicas × 3 pose sets) are launched in parallel. Performance scaling is demonstrated on the Frontier exascale platform (37,888 AMD MI250X GPUs), delivering 5,000 concurrent FG-ESMACS runs and completion in under two hours for both the 3CLPro and TNKS2 ligand sets (112 min and 101 min, respectively). No cross-validation stages are reported beyond ensemble statistics.

6. Implementation and Workflow Integration

MD simulations are performed in NAMD, while MM/PBSA energy calculations utilize AmberTools. Ligand and protein preparations, including protonation, coordinate generation, and charge assignment, leverage the OpenEye software suite (FixpKa, OMEGA, FRED, assigncharges.py). Boltz-2 inference, which supplies alternative starting poses for FG-ESMACS, is executed on Isambard-AI (NVIDIA GH200 G/H GPUs), requiring approximately 75 s per compound per GPU. Overall, FG-ESMACS achieves 10 ns per replica in roughly 4.5 minutes, supporting large-scale, fully parallelized free energy calculations for compound prioritization and AI workflow benchmarking (Wan et al., 2 Mar 2026).

---

In summary, ESMACS provides a rigorously validated, ensemble-based framework for protein–ligand binding free energy estimation that supports both high-throughput screening and precision ranking for drug discovery. The protocol's design, exclusion of entropic contributions, and statistically sound error quantification enable direct benchmarking of AI-based approaches such as Boltz-2, whose energetic predictions, as observed, presently lack the resolution of fine-grained ESMACS on top candidates (Wan et al., 2 Mar 2026).