henhaobio
Sign in

Gut MicrobesMicrobiologyPeer-reviewed journal article

LGTM: Gaussian Process Modulated Neural Topic Modeling for Longitudinal Microbiome

Xiao YuaniD, Ádám AranyiD, András FormanekiD, Yves MoreauiD, Harri J LähdesmäkiiD, Tommi VataneniD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Longitudinal microbiome data are key to understanding the dynamics of microbial communities and their relationships with the host and environment. However, analysis of such data is challenging due to high dimensionality, compositionality, irregular sampling and temporal dependencies on external covariates. Existing analytical approaches typically address only subsets of these challenges, limiting their ability to yield biologically interpretable insights. We introduce LGTM, a probabilistic modeling framework that combines flexible non-linear longitudinal modeling with interpretable topic-based representations of the microbiome. LGTM simultaneously identifies microbial co-abundance patterns ("topics") and models how their proportions change over time and in relation to host and environmental covariates. Using multiple longitudinal human gut microbiome datasets, we demonstrate that LGTM identifies diverse microbial topics whose major patterns are reproducible across runs, while achieving competitive performance in imputation and forecasting tasks. A key strength of the framework is its interpretability: LGTM yields microbial topics with biologically interpretable taxonomic compositions and directly quantifies associations between covariates and microbial dynamics. LGTM is available at https://github.com/yuanx749/lgtm.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (3 Sept 2026) · CC BY 4.0

LGTM couples a neural topic model with additive basis-function Gaussian processes, learning microbial co-abundance topics and how covariates relate to their trajectories in longitudinal gut microbiome data.

Longitudinal gut microbiome data are high-dimensional, compositional and irregularly sampled, and prior tools each handle only part of this: mixed models treat taxa separately, Gaussian processes ignore inter-taxon structure, neural imputers lack interpretability, and topic models are static. LGTM combines covariate-aware temporal modeling with interpretable topic decomposition in one end-to-end framework.

Key results

  1. ShownOn the two infant cohorts, LGTM beat recurrent and attention-based imputers at predicting relative abundances at missing and future timepoints. It trailed the black-box DGBFGP only slightly, and in exchange it has an interpretable linear topic decoder.

    Table 2 compares negative log-likelihood (cross-entropy) on Dhaka and DIABIMMUNE against BRITS, SAITS and DGBFGP, averaged over five random splits with five seeds each. Imputation used 70/10/20 sample splits; forecasting held out 10%/20% of subjects and scored their second-half timepoints. Models trained on CLR-transformed data with MSE loss gave worse likelihood (Table E1). · 3.2. LGTM Provides Accurate Imputation and…; E.2. Experimental details; E.3. Experiment with CLR transform; Tables

  2. ShownAcross all three cohorts, LGTM topics were less redundant than those from NMF, LDA and covariate-free VAE-LDA at similar reconstruction likelihood. Initializing the decoder with NNDSVD substantially improved topic stability from run to run.

    Minimum pairwise Jensen-Shannon divergence and Bray-Curtis dissimilarity were higher over a wide range of L (Figure 2, Table 3), and 5-fold cross-validated NLL was comparable across methods. Topic matrix distance across five seeds improved with NNDSVD ('nnd'). Alignment of models with L = 2 to 20 kept the dominant topics. · 3.3. LGTM Discovers Diverse and Stable Microbial…; F.3. Experimental details; Figure legends; Tables

  3. SuggestedIn Dhaka, Sobol' decomposition attributed the most topic variation to age and tied a Bifidobacterium-dominated topic to breastfeeding. In DIABIMMUNE, country of birth explained much of the variation for several topics, and Bacteroides patterns tracked country and caesarean delivery.

    Dhaka (L = 4): breastfeeding showed contrasting age interactions with Topic 1, while antibiotics and season had milder effects; Topic 2 tracked falling B. longum subsp. infantis and rising Prevotella copri. DIABIMMUNE (L = 5): Bacteroides was higher in Finnish and Estonian children and missing early after C-section; Russian children had more early bifidobacteria. · 3.4. LGTM Identifies Interpretable Microbial…; D.1. Dhaka [52]; D.2. DIABIMMUNE [53]; Figure legends

  4. SuggestedIn the HMP2 inflammatory bowel disease (IBD) metagenomes, differences between subjects dominated topic variance and time contributed little. Dysbiosis state outweighed diagnosis, separating a health-associated topic from one enriched for dysbiosis-associated species; diet covariates also mapped onto topics.

    L = 6. Topic 1 (Bacteroides uniformis, Alistipes putredinis) was also tied to probiotic intake and to drinks with artificial sweeteners, dairy and red meat. Topic 5 held Escherichia coli and Ruminococcus gnavus. Topic 2 (B. vulgatus) was enriched in both UC and CD, Topic 3 (F. prausnitzii) in UC, and Topic 6 was linked to whole grains. · 3.4. LGTM Identifies Interpretable Microbial…; D.3. HMP2 [54]; Figure legends

  5. ShownLGTM training time scaled roughly linearly with sample count and depended only weakly on the number of taxa and subjects. The number of GP covariate components is the main practical cost driver.

    Synthetic relative-abundance benchmark, 100 epochs on one CPU thread (AMD EPYC 7763), L = 5. Log-log runtime slopes: 1.02 for samples, 0.09 for taxa, 0.04 for subjects and 0.64 for GP components (Figure F1). · F.4. Computational efficiency; Figure legends

Methods, data and reproducibility

A one-hidden-layer MLP encoder maps relative abundances to logistic-normal topic proportions, which an NNDSVD-initialized, simplex-constrained topic-taxon matrix decodes. In parallel, additive basis-function Gaussian processes (GPs) on time, subject and time-by-covariate interactions give softmax-averaged proportions, KL-aligned to the encoder. Variance-weighted Sobol' indices score covariates.

  • Shotgun fecal metagenomes profiled with MetaPhlAn2 from Bangladeshi children (Microbiota and Health Study), first two years of life; diarrheal-episode samples excluded · human (children, Dhaka, Bangladesh) · n = 222 children · PRJNA806984 (NCBI SRA BioProject)
  • 16S rRNA amplicon stool profiles, merged to genus level, from DIABIMMUNE children sampled monthly from birth over three years · human (children from Finland, Estonia and Russia) · PRJNA497734 (NCBI SRA BioProject)
  • Stool shotgun metagenomes from the Integrative Human Microbiome Project IBD cohort (IBDMDB), covering Crohn's disease, ulcerative colitis and non-IBD participants over one year · human (IBD patients and non-IBD controls)
  • Synthetic relative-abundance datasets varying samples, taxa, subjects and GP components, used for runtime scaling
Methods & tools
LGTM, MetaPhlAn2, Hilbert-space basis-function GP approximation, NNDSVD decoder initialization, DGBFGP, BRITS and SAITS (via PyPOTS), NMF and LDA (scikit-learn), VAE-LDA, PyTorch, Optuna, Variance-weighted total Sobol' indices, Optimal transport topic alignment
Software & environment
Implemented in PyTorch; BRITS/SAITS via PyPOTS, NMF/LDA via scikit-learn, Optuna for hyperparameter search; Adam, M = 5 basis functions, GP KL coefficient 0.001, learning rate 0.05, batch size 64, MLP hidden size 64; no package versions given.
Assay conditions & compute
Scalability runs used AMD EPYC 7763 CPUs on a single thread for 100 epochs with L = 5. Hyperparameter search ran in parallel via Optuna on a computing cluster, one thread per process; no GPU use is reported.
Reproducibility gaps
No package versions, lock file or container are given. The text does not say whether the processed abundance tables, covariates and split definitions are shared; it points only to raw sources (SRA, the DIABIMMUNE site, IBDMDB).
Limitations (6)
  • The cross-entropy reconstruction on relative abundances is a multinomial-style working likelihood. It does not model sample-specific read depth, overdispersion or abundance-level zeros; the authors propose count-aware likelihoods as extensions. · stated by the authors
  • The squared exponential kernel assumes gradual trends, so abrupt changepoints that are unrecorded or poorly sampled may be smoothed away. · stated by the authors
  • Additive GP components and Sobol' indices describe associations, not causal effects, so correlated metadata need care. Topics are statistical co-abundance patterns, not ecological guilds or pairwise microbial interactions. · stated by the authors
  • Models are fit to one cohort at a time, and transferring topics across cohorts is unresolved. How to choose the number of topics remains open, and initialization can still change the detailed topic composition. · stated by the authors
  • Imputation and forecasting were benchmarked only on the two infant cohorts, which have strong age signals. Predictive accuracy on HMP2, where time explained little variance, was not evaluated. · noted in this summary
  • Covariate-topic links rest on point Sobol' indices from fitted models, without significance tests or replication in independent cohorts; only the latent functions carry one standard deviation uncertainty bands. Agreement with prior literature is qualitative. · noted in this summary

Read this if you analyze irregularly sampled longitudinal gut microbiome cohorts and need interpretable taxon topics whose trajectories are tied to age, feeding, delivery mode, diet or disease metadata.

Citations
0
FWCI
0.00
Field percentile
top 32%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISQu '25Shenhav '18Woloszynek '17Lugo-Martin '18Alvernaz '25Kodikara '25
66% semantic matchMicrobiology
Yixiang Qu, Ruiqi Lyu, Duan Wang+13 more·bioRxiv·Feb 17, 2025· 4 citations