inSilico Dynamics · Computational biophysics

EGFR Resistance: MM-GBSA vs Docking (MD Case Study)

A comparison of osimertinib docking and end-point free-energy estimates across EGFR constructs, with unresolved comparisons and WT non-convergence reported alongside the results.

Question

Does end-point MM-GBSA rank the EGFR resistance pattern for osimertinib better than docking scores? Both methods put T790M first. Only the T790M–triple-mutant comparison is statistically resolved; the WT comparisons are not resolved.

Systems and methods

Osimertinib was docked into WT (4WKQ), T790M (6JX0) and the L858R/T790M/C797S construct (6LUD). OpenMM simulations used ff14SB protein, GAFF2/AM1-BCC ligand parameters with a shared charge set, TIP3P water and 0.15 M NaCl.

Minimisation and NVT/NPT equilibration preceded 3 × 10 ns per genotype. The 3 seeds re-draw velocities from one equilibrated state per genotype; they are not independent starting structures. WT was extended by a further 10 ns per seed.

AmberTools MMPBSA.py used single-trajectory MM-GBSA, igb = 5 and 0.15 M salt, with 250 frames from the last 5 ns of the chosen window: WT 15–20 ns, T790M and triple 5–10 ns. All free energies use the common receptor range EGFR 701–989. There is no entropy term.

Docking and MM-GBSA results

Mean docking scores and MM-GBSA estimates (kcal/mol). MM-GBSA uncertainty is SD across 3 MD seeds.
Genotype (start structure)Docking, mean of 3 seedsΔG_GBSA, mean ± SDMM-GBSA window
WT (4WKQ)−7.87−43.5 ± 15.115–20 ns
T790M (6JX0)−8.80−55.8 ± 4.25–10 ns
L858R/T790M/C797S* (6LUD)−8.67−40.8 ± 3.25–10 ns

* The 6LUD construct also carries E865A/E866A/K867A; differences from the other constructs cannot be attributed to L858R and C797S alone.

  • Both methods rank T790M first. MM-GBSA resolves T790M from the triple-mutant construct: −15.0 kcal/mol (95% CI −23.8 … −6.3; Welch p = 0.01; n = 3 seeds per genotype). About two thirds of the gap (−10.1 kcal/mol) is van der Waals.
  • T790M vs WT is not resolved: −12.2 kcal/mol (95% CI −46.6 … +22.1; p = 0.29).
  • WT vs triple mutant is not resolved: −2.8 ± 8.9 SE kcal/mol. Docking ranks the triple mutant above WT; MM-GBSA cannot say.
  • Welch tests and confidence intervals use seed means. No multiplicity correction was applied; the README notes that the T790M–triple p = 0.01 remains below 0.05 after Bonferroni correction.
EGFR MM-GBSA estimates and energy components for WT, T790M and the triple-mutant construct
MM-GBSA estimates and energy components from the public EGFR study; absolute energies are model-dependent.

Residue decomposition

The decomposition uses 3 seeds per genotype and 25 frames per run. The reported T790M gain concentrates on residues 718 (−2.5 ± 0.5 SE kcal/mol), 796 (−1.8 ± 0.5), 793 (−1.3 ± 0.3), 792, 723 and 726. Residue 790 itself contributes −0.1 ± 0.3. Residue 797 (−1.66 ± 0.85) is not resolved.

Per-residue MM-GBSA contributions and differences in the EGFR genotype comparison
Residue contributions from the public study; sparse frames and seed uncertainty limit interpretation.

RMSD, plateau and WT convergence

Core Cα RMSD uses residues 701–989. In the 5–10 ns window, both core and ligand plateau for all T790M and triple-mutant seeds. In WT, osimertinib moves away from the docked starting pose in all three seeds. At 15–20 ns, WT seed 1 reaches ligand RMSD 5.6 Å; seed 3 is still drifting at +0.22 Å/ns.

WT is not converged. Its ΔG reflects different poses across seeds (−27 to −56 kcal/mol), rather than a settled comparison of the same pose.

Interpretation and limitations

This is a computational study, not experimental validation. Absolute ΔG values are model-dependent; only ranking is interpretable within these limitations.

  1. Starting-structure bias. WT starts from a gefitinib-bound crystal; the mutant structures contain the co-crystal ligand YY3. Osimertinib was docked into each. Pocket shapes differ by construction and may favour the mutants, contributing to WT pose drift.
  2. Seeds share one equilibrated state. n = 3 per genotype underestimates structural uncertainty; the T790M–WT comparison is under-powered.
  3. WT not converged at 20 ns. Seed 3 still drifts and seed 1 settles in a displaced pose. Longer WT runs from several starting poses remain an open item.
  4. Covalent binding is not modelled. Osimertinib binds C797 covalently; these calculations are non-covalent, so C797S resistance is not expected to appear in these energies.
  5. No entropy; GB/PB disagreement. The same T790M seed frames give ΔG_GB = −60.7 and ΔG_PB = +0.8 ± 0.3 (SEM) kcal/mol. The roughly 60 kcal/mol difference limits interpretation of absolute values.
  6. Residue decomposition is limited. Only 25 frames per run and three seeds are used. An earlier single-seed comparison against unequilibrated WT suggested residue 797 as the largest contributor (−3.4); with three seeds it is −1.66 ± 0.85 and not resolved, while residue 718 leads.

Source and reproducibility

This page summarizes the public repository README. Read the methods, derived data and limitations in the source repository.