Mode-Mixing MCMC Enhances Sampling Efficiency for G-Quadruplex Ligand Binding Pose Sampling under a Walltime-Normalized Benchmark

# MM-MCMC G-Quadruplex Ligand Sampling Benchmark This repository contains the complete codebase, input data, and analysis scripts used in the benchmark study: **“Mode-Mixing MCMC Improves Accuracy-per-Compute for G-Quadruplex Ligand Binding Pose Sampling under a Walltime-Normalized Benchmark”** submitted to the *Journal of Chemical Information and Modeling (JCIM)*. --- ## Overview Efficient sampling of ligand binding poses is a central challenge in physics-based drug discovery.This work introduces **Mode-Mixing Markov Chain Monte Carlo (MM-MCMC)**, an adaptive Monte Carlo method that learns correlated rigid-body proposal moves from the sampling trajectory to accelerate exploration of multimodal binding landscapes. To rigorously evaluate sampling efficiency, we designed a **walltime-normalized benchmark** in which MM-MCMC is compared against standard Metropolis Monte Carlo (MC) under identical computational budgets. This Zenodo archive provides all materials required to **fully reproduce the benchmark results** reported in the manuscript. --- ## Contents The archive includes: - **MM-MCMC implementation** - Adaptive proposal learning with numerical stabilization (ridge regularization, eigenvalue projection) - Rigid-body (6 DOF) move set optimized for ligand–macromolecule sampling - **Standard MC reference implementation** - Identical move set and energy model for fair comparison - **Surrogate potential** - Fast DNA–ligand interaction model (electrostatics + Lennard-Jones) - Flat-bottom distance restraint to maintain a well-defined binding region - **Benchmark scripts** - Preflight runs (10 min walltime) for rapid performance estimation - Production runs (60 min walltime) used in the main manuscript - **Analysis and plotting scripts** - Effective sample size per second (ESS/s) - Speedup statistics and geometric mean - Figure reproduction (main and validation figures) - **Input data** - Docked poses for six chemically diverse G-quadruplex ligands - Configuration files used in the benchmark --- ## Benchmark Design - **Systems**: 6 G-quadruplex ligands (DC-34, PhenDC3, Pyridostatin, BRACO-19, DAPI, Caffeine)- **Replicates**: 3 docking poses × 3 random seeds- **Methods compared**: MM-MCMC vs standard MC- **Compute budget**: Fixed walltime per simulation (10 min preflight, 60 min production)- **Primary metric**: Effective sample size per second (ESS/s) All comparisons are strictly walltime-normalized to isolate the effect of proposal adaptation on sampling efficiency. --- ## Reproducibility All results reported in the manuscript, including figures and tables, can be reproduced using the scripts provided in this archive.Detailed instructions for running the benchmarks and generating the figures are included in the `README`. --- ## Relationship to the Manuscript - **Main manuscript**: Describes the algorithm, benchmark design, and key results- **Supporting Information (SI)**: Provides additional methodological details and supplementary analyses- **This Zenodo archive**: Supplies executable code and data enabling full computational reproducibility --- ## License This project is released under the **MIT License**, permitting reuse and modification with attribution. --- ## Citation If you use this code or data, please cite the associated manuscript: > *Mode-Mixing MCMC Improves Accuracy-per-Compute for G-Quadruplex Ligand Binding Pose Sampling under a Walltime-Normalized Benchmark* > (Journal of Chemical Information and Modeling, under review) A DOI for this archive will be added upon publication.