WHAM / MBAR · проверочные наборы / benchmarks · 15.10.2026

1. Synthetic harmonic and double-well examples.
Generated locally by prepare_examples.py, seed 145; not biological data.
Harmonic physical U(x)=0.4*x^2 kcal/mol; AR(1) rho=0.7 samples.
Double well physical U(x)=2.5*(x^2-1)^2-0.4*x kcal/mol; inverse-CDF samples.
Two independent synthetic replicas, 15 windows each, 2000 observations per window.
Known answers are integrated over bins, rather than evaluated only at bin centers.
MBAR RMSE on supported bins: 0.05089 and 0.02710 kcal/mol respectively.
Full double-well input exceeds the online decorrelated-sample limit; use local toolkit
for full reruns. Precalculated results remain available in the browser.

2. Lysozyme L99A / benzene.
Mobley et al. J Mol Biol 371(4):1118–1134 (2007), doi:10.1016/j.jmb.2007.06.002.
Official PyMBAR example of a valine side-chain chi torsion, 26 windows, 501 records/window.
https://github.com/choderalab/pymbar/tree/ed40ec3bbef03bb08938ad1a74d459b0d1ab81f7/examples/umbrella-sampling-fes
Frozen source commit: ed40ec3bbef03bb08938ad1a74d459b0d1ab81f7
Each original XVG and centers.dat was hashed before conversion; download-manifest.json
records source URLs and checksums. XVG time/CV records converted to CSV without thinning.
This is an openly published simulation example, not wet-lab validation or troponin data.
Rerun of official histogram script: RMSE 0.000527 kcal/mol versus our profile;
with R aligned: 0.000181 kcal/mol, close to printed-reference rounding.

3. Method references:
https://doi.org/10.1063/1.2978177 — Shirts & Chodera, MBAR, 2008.
https://pymbar.readthedocs.io/en/stable/fes_with_pymbar.html
https://pymbar.readthedocs.io/en/stable/timeseries.html
Software: PyMBAR 4.0.3. Own discrete WHAM iteration; equal-width bins.
Ten numerical validation groups: analytic truth, periodic seam, unit invariance,
disconnected windows, drift/correlation and input validation. Numerical validation
is not a general percentage of biological accuracy.
