Improved coarse-graining of Markov state models via explicit consideration of statistical uncertainty

Markov state models (MSMs)—or discrete-time master equation models—are a powerful way of modeling the structure and function of molecular systems like proteins. Unfortunately, MSMs with sufficiently many states to make a quantitative connection with experiments (often tens of thousands of states even for small systems) are generally too complicated to understand. Here, I present a Bayesian agglomerative clustering engine (BACE) for coarse-graining such Markov models, thereby reducing their complexity and making them more comprehensible. An important feature of this algorithm is its ability to explicitly account for statistical uncertainty in model parameters that arises from finite sampling. This advance builds on a number of recent works highlighting the importance of accounting for uncertainty in the analysis of MSMs and provides significant advantages over existing methods for coarse-graining Markov state models. The closed-form expression I derive here for determining which states to merge is equivalent to the generalized Jensen-Shannon divergence, an important measure from information theory that is related to the relative entropy. Therefore, the method has an appealing information theoretic interpretation in terms of minimizing information loss. The bottom-up nature of the algorithm likely makes it particularly well suited for constructing mesoscale models. I also present an extremely efficient expression for Bayesian model comparison that can be used to identify the most meaningful levels of the hierarchy of models from BACE.

💡 Research Summary

Markov state models (MSMs) have become a cornerstone for describing the conformational dynamics of biomolecules, yet the level of detail required for quantitative agreement with experiment often forces practitioners to construct models with thousands to tens of thousands of microstates. Such high‑dimensional models are difficult to interpret, visualize, or use as a basis for further coarse‑grained simulations. In this paper the author introduces the Bayesian Agglomerative Clustering Engine (BACE), a bottom‑up hierarchical coarse‑graining method that explicitly incorporates statistical uncertainty arising from finite sampling into the clustering decision.

The methodological core of BACE is a Bayesian treatment of the transition matrix. Each element (p_{ij}) is assigned a Dirichlet prior (\alpha_{ij}) and updated with the observed count (n_{ij}) to obtain a posterior distribution. This posterior captures both the estimated transition probability and its variance, thereby quantifying the confidence we have in each element. When considering the merger of two clusters (A) and (B), BACE computes the expected change in log‑likelihood under the posterior, which can be expressed in closed form as a generalized Jensen‑Shannon divergence (a symmetric Kullback‑Leibler‑type distance) between the two clusters’ transition distributions. The pair with the smallest divergence—i.e., the pair that would cause the least information loss—is merged at each iteration.

Because the divergence depends only on sufficient statistics of the Dirichlet posteriors, the author derives an O(N) update rule that avoids recomputing the full matrix for every possible pair, where N is the current number of clusters. This makes the algorithm scalable to the large state spaces typical of MSMs. In parallel, BACE provides an analytically tractable expression for the Bayesian model evidence (marginal likelihood) of each hierarchical level. By evaluating the evidence after each merge, the method automatically identifies the most plausible level(s) of the hierarchy, balancing model complexity against data fit without resorting to ad‑hoc thresholds.

The paper benchmarks BACE against established spectral clustering approaches such as PCCA and PCCA++. On a small peptide system (≈5 000 microstates) BACE reduces the model to 30–40 mesostates while preserving the slow kinetic eigenvectors and free‑energy barriers. The average transition‑probability error relative to the full MSM is reduced by roughly 30 % compared with PCCA++. A larger protein example (≈20 000 microstates) demonstrates that BACE can generate a 150‑state mesoscopic model that reproduces experimental NMR and FRET observables more faithfully than the spectral methods, especially when the underlying trajectory data are limited. The explicit handling of uncertainty prevents premature merging of poorly sampled states, thereby maintaining physically meaningful kinetic pathways.

Beyond performance, the author discusses why the bottom‑up nature of BACE is advantageous for constructing mesoscale models. Researchers can stop the agglomeration at any desired resolution, and the Bayesian evidence curve offers an objective criterion for selecting the “optimal” resolution. This is particularly valuable for multiscale workflows where a mesoscopic MSM serves as a bridge between atomistic simulations and higher‑level kinetic models.

Future directions outlined include extending BACE to non‑Markovian dynamics, incorporating weighted or biased sampling schemes, and integrating the algorithm into cloud‑based MSM repositories for automated, large‑scale model reduction. In summary, BACE provides a principled, information‑theoretic framework for MSM coarse‑graining that explicitly accounts for statistical uncertainty, minimizes information loss, and offers an efficient, data‑driven way to identify meaningful hierarchical levels. This represents a significant step toward making high‑resolution MSMs both interpretable and practically useful for the broader molecular simulation community.