Bayesian Multinomial Logistic Normal Additive Regression for Microbiome Data Analysis
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Bayesian Multinomial Logistic Normal Additive Regression for Microbiome Data Analysis".
Jane: Here's the gist of the paper. Statistical analysis of microbiome data is challenging because existing approaches are either computationally intractable or restricted to purely parametric or non-parametric methods, which limits their flexibility and scalability.
Tom: This work introduces MultiAddGPs, a novel semi-parametric framework that integrates additive Gaussian Process (GP) regression within a Bayesian MLN model to disentangle linear and non-linear covariate effects, including non-stationary dynamics.
Jane: Our approach builds on the computationally efficient Collapse-Uncollapse (CU) sampler and additive GP regression, introducing a novel back-sampling algorithm and marginal likelihood approximation for efficient inference and hyperparameter estimation.
Tom: Model Formulation The MultiAddGPs model has the following form: Y·n ∼ Multinomial(Π·n), where Π·n = ϕ−1(H·n).
Title and authors: Tom: So we're looking at this paper now "Bayesian Multinomial Logistic Normal Additive Regression for Microbiome Data Analysis". The title itself tells you a lot about what they're trying to do with microbiome data.
Jane: Exactly. It’s saying they are using additive regression inside a Bayesian model, which is pretty smart because it lets them look at both the linear and the non-linear parts of how different factors influence community structure.
Lu: What’s interesting here is that they're tackling those complex non-stationary dynamics mentioned in the title by integrating Gaussian Process regression into this framework. That suggests they aren't just looking for simple linear relationships; they are trying to model things that change over time or depend on location.
Meng: I mean, from an engineering standpoint, the challenge with microbiome data is that it’s compositional and zero-laden, which means standard methods often struggle with the noise. Does this approach actually handle those zeros well at scale?
Tom: That's a fair question, Meng. The paper points out that counts do not represent true abundance; they are just samples from a larger pool, and the measurement process limits us to relative abundances, which is why they use MLN models in the first place >
Jane: And what I find fascinating is their method for handling those zeros. They aren't relying on log-ratio transformations alone, which is a common way people try to deal with this kind of data >
Lu: They are using Bayesian MLN models because they handle the count compositional nature of these data directly, which addresses the limitations of traditional methods that typically involve analyses of log-ratio transformed data >
Tom: Right. And the core idea is that they build this on a computationally efficient Collapse-Uncollapse sampler and additive GP regression, which sounds like they’ve done some heavy lifting on the computational side >
Meng: Computationally efficient is one thing, but what about deployment? If we’re running this on massive sequencing datasets in production, how fast is the inference actually going to be when you introduce these GPs?
Jane: Well, the authors claim that their methods are over two hundred forty thousand times faster than alternatives while simultaneously producing more accurate posterior estimates >
Tom: two hundred forty thousand times faster. That’s a big number. But accuracy is everything in this field, especially when we’re talking about biological insights >
Lu: They support that claim by using the Laplace approximation to the collapsed form of Multinomial Logistic Normal Posterior Transforms, and they state that this approximation is computationally efficient and extremely accurate in the context of microbiome data analysis >
Meng: So it’s fast and accurate. But what’s the catch? What is their actual error bound when you use that Laplace approximation? I need to know how reliable those posterior estimates are when we actually deploy them on real traffic >
Title and authors: Jane: They give a specific way to bound that error, which they state is bounded in probability by Op(D minus one times P N n equal one lambda n minus one where lambda n equals the PD d equals one Y dn >
Tom: Okay, so they are giving us a mathematical guarantee on the error rate. That’s good because it moves this out of just being a promising idea and into something that has some quantifiable reliability >
Lu: And beyond the efficiency, they introduce a novel back-sampling algorithm and a marginal likelihood approximation specifically for efficient inference and hyperparameter estimation >
Meng: Marginal likelihood approximations are often tricky to get right in complex models like this. How do they ensure these approximations stay stable when you’re estimating hyperparameters using Maximum Marginal Likelihood or penalized MLN estimation >
Jane: They show how the same Laplace approximation used in the collapse step of the Collapse-Uncollapse sampler can also be used to approximate marginal likelihoods, which allows them to estimate those hyperparameters via Maximum Marginal Likelihood or penalized MLN estimation >
Tom: That’s a clever way to reuse their existing computational structure, so it saves them from having to develop entirely new inference machinery for every part of the model >
Lu: And they even show practical application in simulation and real data studies. In Simulation two MultiAddGPs produced effective samples at over two hundred forty thousand times the rate of Lgpr >
Jane: That comparison shows the speed advantage is significant when compared to other methods in their internal testing environment >
Meng: So, if we think about real-world deployment, what does that simulation result actually mean for us on a production server? Does that speed translate into usable throughput for a large community dataset >
Tom: It means the inference time drops dramatically. If an alternative method takes ten hours and this one takes one minute, you can run more experiments in the same time window >
Lu: They also showed strong evidence in their application to artificial human guts that the Rikenellaceae relative abundance decreases drastically during an initial starvation event, followed by a later overcorrection where their abundance ultimately increases before trending back to pre-starvation levels >
Jane: That specific finding about the Rikenellaceae is really interesting because it shows they can actually find novel biological insights from data that was originally published in a human gut study >
Tom: So they're not just doing math; they're modeling these kinds of complex microbial community shifts influenced by environmental factors like starvation >
Meng: That moves it beyond just fitting a curve to understanding a dynamic process. But what about the flexibility of the model itself? Can it handle things we haven't thought of, like specific treatment interventions over time >
Title and authors: Lu: The paper also suggests incorporating non-stationary kernel functions designed to model treatment interventions and disease effects, which means the framework can capture complex microbial community dynamics in response to external stimuli >
Jane: That capability is key because it lets researchers infer how specific treatments alter microbial communities over time, not just snapshot changes >
Tom: So we have a framework that’s fast, relatively accurate with error bounds, and capable of modeling non-linear interactions and temporal dependencies like those in treatment response >
Meng: The software is available through the fido package on CRAN one <ref:2410.03911#pg2>. That means it's accessible, which is crucial for adoption, but what about the complexity of setting up those prior distributions for the Gaussian Process terms >
Jane: They state that they use priors like a Normal distribution with a mean of zero and a variance gamma distribution for B, and GP kernels are defined by an Inverse Wishart distribution ζ >
Tom: Those priors need careful tuning, obviously. But the authors claim that using Bayesian Optimization with the upper confidence bound acquisition function helps guide this hyperparameter estimation process efficiently >
Lu: They use that same Laplace approximation for the collapse step to estimate marginal likelihoods, and then they perform Bayesian Optimization with ParBayesianOptimization R package using an upper confidence bound acquisition function >
Meng: It sounds like a self-reinforcing inference loop. If the approximation works well for estimating hyperparameters, it helps refine the model structure, which improves the sampling efficiency for the main inference >
Jane: It does sound like a system designed to be robust across different scales of complexity and computational demands in microbiome analysis >
Tom: So to wrap up on this "Bayesian Multinomial Logistic Normal Additive Regression for Microbiome Data Analysis" paper, it’s a framework that tackles the hard problems of compositional data with additive modeling, offering significant speedups while maintaining rigorous error bounds derived from Laplace approximations >
Lu: It provides a way to disentangle linear and non-linear covariate effects, including non-stationary dynamics, which is what we really needed for complex biological signals >
Meng: From a deployment standpoint, the speed advantage is substantial enough that it could make these kinds of analyses feasible on high-throughput sequencing data rather than being confined to small sample sizes >
Jane: It’s a powerful tool because it respects the counting noise inherent in the data while still allowing for deep modeling of complex biological processes like community shifts during starvation or treatment effects >
Tom: That’s what we were looking for. It brings flexibility and speed to an area that has historically been too computationally heavy or too restricted to simpler methods >
The paper's summary: Tom: So, this paper is essentially saying that when you look at microbiome data—that messy count data—you can use this Bayesian model to separate what’s linear and what’s curved in how different factors affect the community structure >
Jane: That's right. They are using additive Gaussian Process regression inside a Bayesian multinomial logistic normal model, which lets the AI handle those complex non-linear ways things change, like when you introduce a new drug or there's a change in environment >
Lu: The core innovation here is how they handle the complexity of the data itself. They aren't just looking at simple log ratios; they use Bayesian MLN models to directly address the count nature and the zero-filled problem inherent in sequencing data >
Meng: I'm still wondering about the speed. If we’re dealing with real, large sequencing runs, how does this framework actually hold up on a production server compared to existing methods?
Tom: Well, they claim it's over two hundred forty thousand times faster than alternatives because they use a clever sampling method called the Collapse-Uncollapse sampler >
Jane: And they back that up by showing that even when you’re estimating those complex settings—the hyperparameters—they can do it efficiently by reusing the same approximation technique for marginal likelihoods >
Lu: They're using this Laplace approximation, which is usually for a collapsed form of a different model, and they show it works extremely well here to get accurate posterior estimates >
Meng: But accuracy matters. What's the actual risk when you rely on an approximation like that instead of just running the full sampler?
Tom: They provide a way to bound that error in probability by giving you a formula based on the dimensions of your data and some scaling factor lambda, which shows they have quantified what they don't know >
Jane: It’s pretty reassuring because it moves this from just an idea to something with a mathematical guarantee on its reliability >
Lu: Beyond the speed, they showed that this model can handle non-stationary dynamics, meaning it can model how things change over time in response to external stimuli like a treatment >
Tom: So we’ve got a fast framework that respects the counting noise and lets us look at both linear and curvy effects on community shifts >
Jane: And they even found real biological signals, like how a specific group of bacteria changes dramatically during an initial starvation event in artificial human guts >
Lu: That application is pretty compelling because it shows the model can actually discover new patterns from existing research data >
Meng: It sounds like a powerful tool for analyzing large-scale sequencing datasets that would otherwise be too slow or too restrictive for us to even look at properly >
Tom: It’s definitely pushing the boundaries on how fast and how deep we can analyze these complex biological systems without just drowning in computational time >
Jane: This opens the door to asking much more nuanced questions about what drives community shifts, whether it's a simple linear trend or a very specific non-linear response to an intervention >
Lu: And if we can get better at modeling those dynamics, we start understanding how targeted interventions actually reshape microbial communities over time >
The paper's improvements: Tom: So, they aren't just happy with their initial setup; they suggest ways to make the model even more flexible and powerful >
Jane: What’s new is that they can incorporate non-stationary kernel functions, which means you can design these kernels specifically to model things like a treatment intervention or a disease effect over time >
Lu: That capability is key because it lets the AI capture complex microbial community dynamics that aren't constant; it lets you see how specific external stimuli alter the system >
Meng: So if we have an experiment where we change the environment mid-way through, this framework can actually model that change dynamically rather than just taking a snapshot >
Tom: Exactly. It moves us away from static models and towards understanding processes that evolve >
Jane: And they also addressed how to estimate those complex settings more reliably. They show you how to use the same Laplace approximation they used in the collapse step for estimating marginal likelihoods >
Lu: That’s a smart way to reuse their existing computational structure; it lets them estimate hyperparameters using maximum marginal likelihood or penalized maximum marginal likelihood estimation >
Meng: I still have my concerns about those approximations being stable when we’re dealing with very large datasets and complex prior setups, but the method for estimating those settings seems sound >
Tom: They are using Bayesian Optimization with an upper confidence bound acquisition function to guide that hyperparameter estimation process, which helps them tune the model structure efficiently >
Jane: So it’s a self-improving loop where the inference step helps you figure out the best settings for the model itself >
Lu: And they make it accessible by releasing all their code through an R package called fido, which is implemented in optimized C++ code, making it ready for actual use >
Meng: Availability is good, but what about the computational cost of running that whole thing when you try to scale it up to millions of samples? I need to know what breaks at scale >
Tom: They claim the speed advantage still holds even with these extensions because they built the model on a foundation that was already computationally efficient >
Jane: It seems they’ve managed to build something that is both flexible enough for complex biology and fast enough for real-world analysis, provided you use their sampling techniques correctly >
Lu: They also showed how to estimate the count nature of data directly using these Bayesian MLN models, which is a big step away from just using log-ratio transformations >
Tom: So we've got a model that respects the counting noise while still allowing us to do deep modeling of complex biological processes like community shifts during starvation or treatment response >
Jane: This really opens the door to asking much more nuanced questions about what drives community shifts, whether it's a simple linear trend or a very specific non-linear response to an intervention >
Lu: And if we can get better at modeling those dynamics, we start understanding how targeted interventions actually reshape microbial communities over time >
Conclusion: Tom: So to wrap up, this paper on "Bayesian Multinomial Logistic Normal Additive Regression for Microbiome Data Analysis" is basically saying we can finally handle complex, non-linear factors in microbiome data without getting bogged down by computational roadblocks >
Jane: They showed us how to use additive models within a Bayesian framework to separate the simple linear effects from those more intricate non-linear relationships that cause community structure changes >
Lu: It’s about making sure we model both the direct influence and the indirect, curvy influence of different factors simultaneously, which is a big step for understanding complex biological signals >
Meng: From a deployment standpoint, it means that when we look at real sequencing data in production, we can expect inference to be way faster than what's currently available for these kinds of analyses >
Tom: Yeah, the speed improvements are substantial enough that this could make running these kinds of deep statistical analyses feasible on larger community datasets >
Jane: And they gave us mathematical guarantees on the error rates using those approximations, which gives us confidence in the results we get out of it >
Lu: That rigor is important because it shows that we aren't just guessing at the accuracy; there’s a bound on how much error we can expect >
Meng: I still have to worry about setting up those priors correctly for real-world data, but the framework itself seems solid for what it aims to do >
Lalam: I am confident in this result because it directly addresses the core limitation of current models: they either ignore the count nature or they only look at linear relationships. >
Tom: Exactly. It's about respecting how counts work and being able to model those non-linear dynamics we talked about earlier >
Jane: This paper on "Bayesian Multinomial Logistic Normal Additive Regression for Microbiome Data Analysis" gives us a much better toolkit for interpreting these complex biological shifts >
Lu: The implication is that we can start inferring how specific interventions actually reshape microbial communities over time, not just taking snapshots of them >
Meng: If we can model those dynamics better, it means targeted treatments could be designed with much more precision to affect the microbiome in a desired way >
Pennsylvania State University
stat.ME, stat.AP
Submitted: 2024-10-04
Updated: 2026-11-10
Code: https://github.com/Silverman-Lab/MultiAddGPs
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 87/100
The gist: Statistical analysis of microbiome data is challenging because existing approaches are either computationally intractable or restricted to purely parametric or non-parametric methods…
Key concepts
- MultiAddGPs
- This is a novel semi-parametric framework that integrates additive Gaussian Process regression into a Bayesian MLN model. It allows AI to disentangle linear and non-linear covariate effects in microbiome data, effectively modeling complex community structure changes influenced by multiple factors, including non-stationary dynamics.
- Collapse-Uncollapse (CU) Sampler
- This is a two-step inference method used for sampling from the posterior distribution of the model. It involves first 'collapsing' to get a sample from the simpler marginal distribution of H given Y, and then 'uncollapsing' by conditioning on that sample to get a sample from the conditional distribution of Phi given H and Y.
- Additive Gaussian Process (GP) Regression
- This technique is used within the model to capture non-linear relationships in the data. It allows for flexible modeling of complex biological signals by adding various kernel functions, enabling AI to decompose community dynamics into interpretable additive components like batch effects or trends.
Terminology
Summary
Statistical analysis of microbiome data is challenging because existing approaches are either computationally intractable or restricted to purely parametric or non-parametric methods, which limits their flexibility and scalability. p. 0 Silverman∗
College of Information Science and Technology,
Department of Statistics,
Department of Medicine,
Pennsylvania State University
May 27, 2025
Abstract
Statistical analysis of microbiome data is challenging. Bayesian multinomial
logistic-normal (MLN) models have gained popularity due to their ability to account
for the count compositional nature of these data, but existing approaches are either
computationally intractable or restricted to purely parametric or non-parametric
methods, which limit their flexibility and scalability. Bayesian multinomial
logistic-normal (MLN) models have gained popularity due to their ability to account
for the count compositional nature of these data, but existing approaches are either
computationally intractable or restricted to purely parametric or non-parametric
methods, which limit their flexibility and scalability. In this work, we introduce MultiAddGPs, a novel semi-parametric framework that integrates additive Gaussian Process
(GP) regression within a Bayesian MLN model to disentangle linear and non-linear
covariate effects, including non-stationary dynamics. In this work, we introduce MultiAddGPs, a novel semi-parametric framework that integrates additive Gaussian Process
(GP) regression within a Bayesian MLN model to disentangle linear and non-linear
covariate effects, including non-stationary dynamics. Our approach builds on the computationally efficient Collapse-Uncollapse (CU) sampler and additive GP regression,
introducing a novel back-sampling algorithm and marginal likelihood approximation
for efficient inference and hyperparameter estimation.Where the paper says it
Where the paper says it
Where the paper says it
Model Formulation
The MultiAddGPs model has the following form: Y·n ∼ Multinomial(Π·n), where Π·n = ϕ−1(H·n). H·n ∼ N (F·n, Σ). F = BX + ΣX K k=1 f(k)(Z(k)), with priors B ∼ N (Θ(0), Σ,Γ(0)), f(k) ∼ GP(Θ(k), Σ,Γ(k)), and Σ ∼ IWΞ, ζ. p. 5 ,K})
·n
a Qk-vector of covariates to be modeled with distinct non-linear
functions. The MultiAddGPs model have the following form:
Y·n ∼ Multinomial(Π·n) (1)
Π·n = ϕ
−1
(H·n) (2)
H·n ∼ N (F·n, Σ) (3)
F = BX +
X
K
k=1
f
(k)
(Z
(k)
) (4)
with priors B ∼ N (Θ(0)
, Σ,Γ
(0)), f
(k) ∼ GP(Θ(k)
, Σ,Γ
(k)
), and Σ ∼ IW(Ξ, ζ). In this work, we introduce MultiAddGPs, a novel semi-parametric framework that integrates additive Gaussian Process
(GP) regression within a Bayesian MLN model to disentangle linear and non-linear
covariate effects, including non-stationary dynamics. Our approach builds on the computationally efficient Collapse-Uncollapse (CU) sampler and additive GP regression,
introducing a novel back-sampling algorithm and marginal likelihood approximation
for efficient inference and hyperparameter estimation. Silverman∗
College of Information Science and Technology,
Department of Statistics,
Department of Medicine,
Pennsylvania State University
May 27, 2025
Abstract
Statistical analysis of microbiome data is challenging. Bayesian multinomial
logistic-normal (MLN) models have gained popularity due to their ability to account
for the count compositional nature of these data, but existing approaches are either
computationally intractable or restricted to purely parametric or non-parametric
methods, which limit their flexibility and scalability.Where the paper says it
Where the paper says it
Where the paper says it
Inference Efficiency
The inference for MultiAddGPs is developed using a two-step Collapse-Uncollapse (CU) sampler. This sampler involves obtaining a sample from the marginal p(H Y) (the collapse step) and then conditioning on that sample to obtain a sample from the corresponding conditional p(Φ H, Y) (the uncollapse step). p. 6 2 Posterior Estimation in MultiAddGPs
We develop scalable inference for MultiAddGPs using a two-step Collapse-Uncollapse (CU)
sampler (Silverman et al. Any multi-parameter Bayesian model p(H, Φ | Y) can
be sampled in two-step: (1) obtain a sample from the marginal p(H | Y) (the collapse
step); (2) conditioned on that sample, obtain a sample from the corresponding conditional
p(Φ | H, Y) (the uncollapse step). , S} do
2: Sample H[s] ∼ p(H|Y, ∆) where p(H, Y, ∆) is an LTP; ▷ Collapsed Step
3: Sample {Σ
[s]
, F
[s]} ∼ p(F, Σ | H[s]
, ∆); ▷ Uncollapsed Step (1)
4: Sample B[s]
,f
(1)[s]
, . ,f
(K)[s] = Back-Sampling(Y, F
[s]
, Σ
[s]
, ∆); ▷ Uncollapsed Step (2)
5: end for
the CU sampler can be found in Supplementary Section 1. (2022), while simultaneously enhancing those
methods. We introduce a novel back-sampling algorithm for efficient posterior simulation
from multivariate additive Gaussian process models as well as non-stationary kernels for
estimation of non-linear intervention effects on complex microbial communities.Where the paper says it
Where the paper says it
Where the paper says it
Computational Performance and Accuracy
The authors claim that their models are over 240,000 times faster than alternatives while simultaneously producing more accurate posterior estimates. The Laplace approximation to the collapsed form of MLTPs is computationally efficient and extremely accurate in the context of microbiome data analysis. p. 9 Building on these results, we
then compute the Hessian ∇−2
vec(Hˆ )
in closed form (also see Supplementary Section 3). This approximation is computationally efficient and extremely accurate. This approximation is computationally efficient and extremely accurate. The error in this
approximation is bounded in probability by Op([D − 1] PN
n=1 λ
−1
n
) where λn =
PD
d=1 Ydn
(See Proposition 11 of Silverman et al.Where the paper says it
Where the paper says it
Application and Validation
The methods are validated through simulation and real data studies. In Simulation 2, MultiAddGPs produced effective samples at over 240,000 times the rate of Lgpr. p. 0 Our approach builds on the computationally efficient Collapse-Uncollapse (CU) sampler and additive GP regression,
introducing a novel back-sampling algorithm and marginal likelihood approximation
for efficient inference and hyperparameter estimation. Our models are over 240,000
times faster than alternatives while simultaneously producing more accurate posterior
estimates. In an application involving an ex vivo artificial
human gut study, our findings offer the first evidence that microbiota communities display
oscillatory-like behavior in response to short-term dietary interventions. MultiAddGPs MLTPs Lgpr LongGP
Model Linear and Nonlinear Linear or Nonlinear Nonlinear Nonlinear
Flexible Kernels ✓ ✓ ✕ ✕
Scalable ✓ ✓ ✕ ✕
Hyperparameter Estimation ✓ ✕ ✓ ✓
Count Compositional ✓ ✓ ✕ ✕
Table 1: A comparison of MultiAddGPs and existing alternatives.Where the paper says it
Where the paper says it
Hyperparameter Estimation
The authors demonstrate how the same Laplace approximation used in the Collapse step of the CU sampler can be used to approximate marginal likelihoods and thereby estimate hyperparameters via Maximum Marginal Likelihood (MML) or penalized MML estimation. In practice, Bayesian Optimization is performed using the ParBayesianOptimization R package with the acquisition function set to the upper confidence bound.
Software Availability
MultiAddGPs are made available through the R package fido (Silverman & Nixon 2021), where it can be accessed via the basset function. All code required to reproduce the results of this article is available at https://github.com/Silverman-Lab/MultiAddGPs. p. 3 Nevertheless, there is still a
need for more flexible approaches that can effectively disentangle linear and non-linear effects
as existing methods do not allow for additive linear and non-linear modeling (Silverman
et al. In this study, we introduce MultiAddGPs, a versatile and interpretable semi-parametric
framework aimed at addressing key challenges in microbiome data analysis. simulated and real data studies and produce novel biological insights from a previously
published human gut microbiome study. Our methods are publicly available as part
of the fido software package on CRAN 1
.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Scalable Bayesian Semiparametric Additive Regression Models for Microbiome Studies allow AI systems to disentangle linear and non-linear covariate effects in microbiome data by integrating additive Gaussian Process regression within a Bayesian multinomial logistic normal model. This enables the AI to model complex community structure changes influenced by multiple factors, including non-stationary dynamics, with high computational efficiency.
-
The MultiAddGPs framework incorporates a novel back-sampling algorithm and marginal likelihood approximation for efficient inference and hyperparameter estimation, making it over 240,000 times faster than alternatives while producing more accurate posterior estimates. This allows AI systems to perform complex statistical analysis on large-scale microbiome datasets much faster than existing methods.
-
The model can incorporate non-stationary kernel functions designed to model treatment interventions and disease effects, enabling the AI to capture complex microbial community dynamics in response to external stimuli. This capability allows for the inference of how specific interventions alter microbial communities over time.
-
AI systems can estimate count compositional nature of data by using Bayesian MLN models, which address limitations of traditional approaches that typically involve analyses of log-ratio transformed data or assume Gaussian distributions. This ensures that the analysis respects the zero-laden nature and counting noise inherent in sequencing data.
-
The model facilitates additive modeling of both linear and non-linear processes, contrasting with methods like MLTPs which consist of either a single linear or a single nonlinear component. This flexibility allows AI to decompose complex biological signals into interpretable additive components like batch effects, periodic components, and long-term trends.
Abstract
Longitudinal microbiome studies often seek to separate background temporal variation from nonlinear perturbation-associated dynamics while accounting for the count-compositional nature of sequencing data. We introduce MultiAddGPs, a semi-parametric Bayesian multinomial logistic-normal framework that combines linear effects with additive Gaussian process components. We first prove that MultiAddGPs remain within the Marginally Latent Matrix-T Process class (MLTP) and derive the collapsed and uncollapsed representations needed for scalable inference with the Collapse--Uncollapse (CU) sampler. We then develop a back-sampling procedure for efficient posterior inference on the individual additive components. We further develop approximate maximum marginal likelihood and penalized maximum marginal likelihood procedures for hyperparameter estimation in this broader class of models. In simulations, MultiAddGPs provides accurate posterior inference and calibrated uncertainty while scaling more effectively than existing GP-based additive alternatives. Applications to two longitudinal microbiome studies demonstrate how the model separates background temporal variation from intervention-associated dynamics, including travel-associated perturbations in a human gut time series and starvation effects in an artificial gut system. Our methods are implemented in the fido package on CRAN.
Sources
- Scale Reliant Inference
- Scalable Inference for Bayesian Multinomial Logistic-Normal Dynamic Linear Models
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes