Code and Data for "Reproducible Adaptive MCMC via Sharing a Pretrained Generator Matrix Across Runs and Structures"

This repository contains all code, data, and figure-generation scripts for reproducing the results in:

"Reproducible Adaptive MCMC via Sharing a Pretrained Generator Matrix Across Runs and Structures"
Masato Tanigawa and Takafumi Iwaki
Journal of Chemical Information and Modeling, 2026

KEY FINDINGS

- Shared-M protocol reduces cross-run variability by 68%
- Two failure modes identified: Freezing (25K steps, Acc < 5%) and Over-adaptation (100K steps, variance doubles)
- Optimal training length: ~50K steps (healthy acceptance 10-30%, lowest cross-seed variance)
- Transferability: Pretrained M transfers across structures (gap < 1 Å)
- Critical insight: Minimizing apparent reproducibility metrics can select for over-adaptation

CONTENTS

- scripts/: Core GM-MCMC implementation and all experiment scripts (Exp1-Exp5)
- data/: Raw results (CSV), including individual seed-level data
- figures/: Publication-quality figure generation scripts (matplotlib)
- structures/: PDB files (1BNA, 1NAJ) used as test systems

REPRODUCTION

pip install numpy matplotlib
python scripts/run_all_experiments_fast.py  # ~5 min
python figures/plot_all_figures.py

LICENSE

MIT License