Sampling Budget as a Confound in Dimensionality-Reduction Analyses of Molecular Simulation
Abstract
A common claim in molecular simulation is that projecting a trajectory reveals N metastable states. Here N is operationally the number of Gaussian-mixture components selected in a low-dimensional embedding, without a kinetic model it should be read as an apparent-basin count, not a validated metastable-state count. Across two potentials with basin counts fixed by construction and two independent 100 ns alanine-dipeptide simulations, the BIC-selected component count rises strongly with sampling budget (Spearman rho = 0.87-0.97 across systems and methods) although the reference landscape is unchanged. The trend persists under uniform thinning of a common reference trajectory, where coverage is approximately matched, and under IID Boltzmann sampling in the exact synthetic latent coordinates, which removes time correlation and projection learning. Other criteria exhibit different failure modes: AIC is frequently ceiling-censored or non-monotone, ICL never reaches the tested ceiling but remains method-dependently inaccurate, and silhouette and the simple elbow diagnostic are largely insensitive. BIC-selected recovery often peaks at an intermediate budget and then declines even while recovery at the reference component count is flat or improves. One exception is explicit, alanine TICA seed 1 shows no decline within the tested range because its high-recovery crossover occurs at the last budget. We recommend reporting sampling-budget curves, treating mixture components as candidate states until kinetically validated, and using ICL only as a less-ceiling-prone sensitivity analysis rather than a general default.
Related
- Interpretable spatial multi-omics data integration and dimensionality reduction with SpaMV
- Shedding light on Gaussia Luciferase conformational space using multi-conformation enhanced sampling simulation
- Deciphering molecular mechanisms of synergistic growth reduction in kinase inhibitor combinations
- Designing Cooperative Molecular Glues for FKBP12-MAPRE1 with Enhanced-Sampling Molecular Dynamics Simulations.
- Light Martini water accelerates sampling in coarse-grained molecular dynamics simulations
BibTeX
@misc{bassi2026sampling,
title = {{Sampling Budget as a Confound in Dimensionality-Reduction Analyses of Molecular Simulation}},
author = {Jaiveer Bassi},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008954/v2},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008954/v2},
note = {CC BY},
}
RIS
TY - UNPB
TI - Sampling Budget as a Confound in Dimensionality-Reduction Analyses of Molecular Simulation
AU - Jaiveer Bassi
PY - 2026
DA - 2026/09/20
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008954/v2
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008954/v2
AB - A common claim in molecular simulation is that projecting a trajectory reveals N metastable states. Here N is operationally the number of Gaussian-mixture components selected in a low-dimensional embedding, without a kinetic model it should be read as an apparent-basin count, not a validated metastable-state count. Across two potentials with basin counts fixed by construction and two independent 100 ns alanine-dipeptide simulations, the BIC-selected component count rises strongly with sampling budget (Spearman rho = 0.87-0.97 across systems and methods) although the reference landscape is unchanged. The trend persists under uniform thinning of a common reference trajectory, where coverage is approximately matched, and under IID Boltzmann sampling in the exact synthetic latent coordinates, which removes time correlation and projection learning. Other criteria exhibit different failure modes: AIC is frequently ceiling-censored or non-monotone, ICL never reaches the tested ceiling but remains method-dependently inaccurate, and silhouette and the simple elbow diagnostic are largely insensitive. BIC-selected recovery often peaks at an intermediate budget and then declines even while recovery at the reference component count is flat or improves. One exception is explicit, alanine TICA seed 1 shows no decline within the tested range because its high-recovery crossover occurs at the last budget. We recommend reporting sampling-budget curves, treating mixture components as candidate states until kinetically validated, and using ICL only as a less-ceiling-prone sensitivity analysis rather than a general default.
ER -