STRC Research

STRC h01 Phase 9a FEP-TI E1659A Alchemical 2026-04-27 SKELETON

5 min read

STRC h01 Phase 9a — FEP / TI Alchemical ΔΔG (SKELETON)

Placeholder. Documents the question, method, smoke-test, production protocol, pass criteria, and known artifacts so the run can fire on a single day’s notice. No production calculation has been performed yet.

Question

For the top-N shortlisted leads from Phase 5q, what is the relative binding free energy difference between E1659A mutant STRC and WT STRC, computed by alchemical λ-mutation rather than by docking score or kinetic proxy?

ΔΔG_bind = ΔG_bind(mutant) − ΔG_bind(WT)

Sign expectation (from existing pipeline): ΔΔG_bind < 0 (mutant binds better).

Method

Two complementary approaches; pick one for production after smoke:

  • (A) Relative binding FEP — same ligand, mutate E1659 → A on protein side via λ-windows in apo and holo states. Closes thermodynamic cycle.
    • Tooling candidates: GROMACS-FEP + pmx (Gapsys), AMBER pmemd.cuda + pmx, OpenMM/Yank, BFEE2.
  • (B) Absolute binding FEP — decouple ligand from solvent and from each protein variant separately; subtract.
    • Tooling candidates: BAT.py (Heinzelmann/Gilson 2021), Loeffler ABFE protocol, Yank.

Recommended default: (A) relative, because protein-side mutation is exactly the variable we care about and the cycle closure cancels most systematic error.

Why orthogonal

  • Different physics from docking: alchemical free energy uses the actual partition function via λ-mutation, no scoring function.
  • Different physics from tauRAMD: equilibrium thermodynamics, no forced pulling.
  • Different from MM-PBSA: explicit solvent, full configurational entropy, no continuum approximation.
  • Same force field as MD pipeline → catches docking/scoring artifacts but does not catch force-field artifacts (see Phase 9d for that).

Inputs needed

  • Phase 5q shortlist top-3 to top-5 ligands (SMILES + parameters in OpenFF or GAFF2).
  • Phase 5d equilibrated holo MD frame for each ligand × {WT, mutant}.
  • Force field: AMBER ff19SB protein + OpenFF-2.x or GAFF2 ligand + OPC or TIP3P water.
  • Hardware: GPU. M5 Max MPS path needs verification (see Apple Silicon first rule); fall back to remote A100 if MPS path crashes for FEP kernels.

Smoke test (1-day, theoretical)

Single ligand, single protein variant, no production statistics:

  1. Build alchemical topology with pmx for one E→A mutation in apo-protein-only (no ligand) — confirms the topology generator works for STRC.
  2. Run 11 λ-windows × 250 ps each, NPT, T=300K — total ~3 ns × 11 = 33 ns wall.
  3. Check dV/dλ time-series finite and bounded; check soft-core potentials don’t explode at λ=0/1.
  4. Estimate apo ΔG_E→A via TI (gmx bar) — value is not the answer, just a pipeline-health signal.

Smoke pass: pipeline runs end-to-end, dV/dλ has no NaN/explosion, BAR analysis returns a finite number with reasonable error bar (<1 kcal/mol on this trivial apo case).

Smoke fail: any of: topology generation crashes, MD simulation explodes, soft-core blows up, BAR gives infinite variance. → debug before committing to production.

Production protocol (theoretical)

  • 21 λ-windows × 5 ns each per leg × 2 legs (apo and holo) × 2 ligand-variant pairs = ~840 ns total per ligand.
  • 3× independent replicas with different random seeds for error bars.
  • Cycle closure check: |ΔG_cycle| < 1 kcal/mol.
  • Per-ligand wall time on single A100: ~3-5 days. For 3 ligands: ~2 weeks GPU or parallelize across multiple GPUs.

Pass criteria

  • PRIMARY: ΔΔG_bind sign matches existing pipeline (negative = mutant binds better) for ≥2 of 3 production ligands.
  • SECONDARY: |ΔΔG_bind| > 1 kcal/mol (above thermal noise) for the lead ligand.
  • TERTIARY: cycle-closure |ΔG_cycle| < 1 kcal/mol — sanity check on numerical convergence.
  • FAIL state: ΔΔG_bind insignificant (|value| < 0.5 kcal/mol) or wrong sign on the lead → drops h01 deliv tier.

Known artifacts and risks

  • Sampling: 5 ns/window may be insufficient for charged-residue mutations (E→A removes a carboxylate); replica-exchange λ-dynamics or Hamiltonian replica exchange might be required.
  • Charge-changing mutation: E→A changes net charge by +1 — must use co-alchemical counterion or background-charge correction (Rocklin, Ross, Pavlova 2013 type artifact).
  • Force-field bias: same FF as MD pipeline → if FF systematically over-favors charged interactions, FEP carries the same bias. Phase 9d (QM/MM) is the disambiguator.
  • Pose dependence: starting pose for holo leg comes from docking; if docked pose is wrong, ΔG is wrong. Mitigate by 3 independent starting poses (Phase 5q replicas).

References (canonical)

  • Mey et al. 2020. LiveCoMS 2:18378 — “Best practices for alchemical free energy calculations”.
  • Wang et al. 2015. JACS 137:2695 — FEP+ benchmark.
  • Cournia, Allen, Sherman 2017. J Chem Inf Model 57:2911 — relative binding FEP.
  • Gapsys et al. 2020. Chem Sci 11:1140 — pmx large-scale validation.
  • Rocklin, Mobley, Dill, Hünenberger 2013. J Chem Phys 139:184103 — finite-size correction for charge-changing FEP.
  • Loeffler et al. 2018. J Chem Theory Comput 14:5567 — reproducibility of ABFE across engines.

Status

  • 2026-04-27 — skeleton created. Software not installed. Smoke not run. Production not run.

Ranking delta

  • No change. Skeleton only.

Connections

Graph View

Backlinks