Umibato: estimation of time-varying microbial interaction using continuous-time regression hidden Markov model

Shion Hosoda, Tsukasa Fukunaga, Michiaki Hamada

Abstract

Accumulating evidence has highlighted the importance of microbial interaction networks. Methods have been developed for estimating microbial interaction networks, of which the generalized Lotka-Volterra equation (gLVE)-based method can estimate a directed interaction network. The previous gLVE-based method for estimating microbial interaction networks did not consider time-varying interactions. In this study, we developed unsupervised learning based microbial interaction inference method using Bayesian estimation (Umibato), a method for estimating time-varying microbial interactions. The Umibato algorithm comprises Gaussian process regression (GPR) and a new Bayesian probabilistic model, the continuous-time regression hidden Markov model (CTRHMM). Growth rates are estimated by GPR, and interaction networks are estimated by CTRHMM. CTRHMM can estimate time-varying interaction networks using interaction states, which are defined as hidden variables. Umibato outperformed the existing methods on synthetic datasets. In addition, it yielded reasonable estimations in experiments on a mouse gut microbiota dataset, thus providing novel insights into the relationship between consumed diets and the gut microbiota.

Introduction

Microbes in complex environments like the human gut do not exist in isolation; they form intricate networks of competition and cooperation. Understanding these directed interactions is crucial for ecological and medical research. Traditionally, the generalized Lotka-Volterra equation (gLVE) has served as the mathematical backbone for modeling these dynamics. However, standard gLVE-based methods typically assume that interaction parameters remain constant over time.

In reality, the gut is a volatile environment. External perturbations, such as dietary shifts or antibiotic administration, can fundamentally alter the "rules of engagement" between microbial species. Existing tools often struggle to capture these transitions, tending to average dynamics into a single, potentially misleading snapshot. Furthermore, biological data is rarely collected at uniform intervals—human sampling, for instance, is often dictated by defecation timing. This irregularity creates significant mathematical hurdles for standard discrete-time models.

Methods: The Umibato Framework

To overcome these limitations, we introduce Umibato (unsupervised learning based microbial interaction inference method using Bayesian estimation). Umibato operates through a two-stage computational pipeline designed to handle both data noise and temporal variability.

1. Gaussian Process Regression (GPR)

The first stage utilizes GPR to estimate microbial growth rates from raw abundance data. GPR is particularly effective in this context because it naturally handles measurement noise and provides a variance estimate, offering a built-in measure of uncertainty for the derived growth rates.

2. Continuous-Time Regression Hidden Markov Model (CTRHMM)

The core innovation of Umibato is the CTRHMM. This model posits that a microbial community exists in one of several discrete "interaction states." Within each state, the microbes follow a specific, state-dependent gLVE network.

Probabilistic graphical model of the Continuous-Time Regression Hidden Markov Model (CTRHMM).

Unlike discrete-time HMMs, the CTRHMM uses a continuous-time Markov chain to govern transitions between states. This allows the model to account for the exact elapsed time between samples, making it robust to irregular sampling intervals. Parameter estimation is performed using Variational Inference (VI), representing a novel application of VI to continuous-time hidden Markov models in microbial ecology.

Key Results and Performance

The researchers evaluated Umibato using both synthetic benchmarks and real-world longitudinal data. On synthetic datasets designed to simulate multi-state dynamics, Umibato demonstrated superior accuracy compared to existing methods like MDSINE (including MLRR and MLCRR variants).

Performance comparison of Umibato against existing methods on synthetic datasets.

As shown in the performance metrics, Umibato achieved a mean Pearson correlation of 0.99 in the true model case and 0.90 in practical applications for multi-state data. In contrast, methods that assume static interactions saw their accuracy drop significantly (to approximately 0.53–0.56) when faced with state changes. This highlights Umibato's unique ability to correctly identify and model shifting ecological networks.

Biological Insights: Mouse Gut Microbiota

The practical utility of Umibato was further demonstrated using a dataset tracking 13 Clostridium strains in mice subjected to diet shifts. Umibato identified five distinct interaction states, providing a more nuanced view of the gut's response to fiber intake:

  • Diet-State Mapping: The model successfully mapped "State 5" almost exclusively to days when mice were on a low-fiber diet.
  • Parasitism and Competition: In State 5, Umibato revealed parasitic relationships—where one strain benefits at another's expense—that were invisible in static models.
  • Transient Dynamics: States 1 and 2 were identified as unstable, transient states occurring immediately after the initial introduction of microbes, capturing the "settling" period of the community.
  • Irreversible Transitions: Simulations based on the estimated parameters suggested that prolonged exposure to a low-fiber diet could lead to the permanent loss of certain beneficial strains, even if a high-fiber diet is subsequently restored.

Conclusion and Future Directions

Umibato represents a significant step forward in microbial community modeling by treating ecological "noise" as meaningful state transitions. By leveraging continuous-time modeling and Bayesian inference, it provides a flexible tool for analyzing complex, irregularly sampled longitudinal data.

Future work will focus on automating the selection of the number of states ($K$) and moving toward a fully Bayesian estimation of the transition rate matrix ($Q$). The C++ and Python source codes for Umibato are available at https://github.com/shion-h/Umibato. As high-quality quantitative abundance data becomes more prevalent, Umibato is poised to provide deep insights into the stability and resilience of human and environmental microbiomes.

Summary updated

BibTeX
@article{hosoda2021umibato,
  author = {Shion Hosoda and Tsukasa Fukunaga and Michiaki Hamada},
  title = {{Umibato: estimation of time-varying microbial interaction using continuous-time regression hidden Markov model}},
  journal = {Bioinformatics},
  year = {2021},
  doi = {10.1093/bioinformatics/btab287},
}
Figure