Are Microbial Keystone Hubs Artifacts of Mixing Niches?

Microbial ecologists routinely grind up a soil core or a stool sample, sequence it, and correlate taxon abundances across many such samples to draw a "co-occurrence network": a graph whose nodes are taxa and whose edges mark pairs that rise and fall together (read as cooperation) or in opposition (read as competition). The most connected taxa are then called keystones. Each sample, however, is a physical mixture of micro-habitats such as mucosa-attached and luminal gut communities, or distinct soil aggregates. When the share of one micro-habitat varies from sample to sample, every taxon that lives there moves together, even if the taxa never interact. Cancer and single-cell biology already know this problem well. Bulk RNA sequencing averages over cell types, and co-expression computed on bulk tissue largely tracks shifting cell-type proportions. That field built deconvolution methods to undo the averaging. Microbial network analysis has not borrowed that lesson.

We propose a model that treats each sequenced microbial sample as a mixture of a small number of latent niches, where a niche is a micro-habitat with its own composition and its own sparse interaction network. The model is fit directly to the raw read counts, and the mixing happens in absolute abundance before the sequencer reduces everything to proportions. It returns per-sample niche proportions and one network per niche. With these outputs we can test a sharp claim: a large fraction of bulk co-occurrence edges, and a disproportionate fraction of edges attached to "keystone" hubs, come from niche mixing rather than from within-niche coordination.

Core Claim

A simple identity shows why the claim is plausible. Let a sample's absolute taxon abundances be $x = \sum_k a_k y_k$. Here $a_k$ is the (random) mass of niche $k$ in that sample, and $y_k$ is the niche-$k$ abundance profile, with mean $\mu_k$ and within-niche covariance $\Sigma_k$. Assume the profiles are independent of the niche masses and of each other. Then across samples

$$\mathrm{Cov}(x) = \sum_k \mathbb{E}[a_k^2]\,\Sigma_k \;+\; M\,\mathrm{Cov}(a)\,M^\top,$$

where $M$ stacks the niche means $\mu_k$ as columns. The first term holds the interactions ecologists care about. The second term is pure mixing. It adds positive edges between taxa enriched in the same niche and negative edges between taxa enriched in niches whose masses trade off. It adds the most edges to taxa that are abundant and niche-specific, which is exactly the profile of a typical "hub."

Sequencing then closes the vector to proportions, and standard pipelines take log-ratios. Because the logarithm of a sum is not a sum of logarithms, the mixing term is no longer a clean low-rank block in the space where compositional network methods operate. So the two usual fixes each handle only half of the problem. Log-ratio sparse networks correct closure but not mixing. Linear deconvolution corrects mixing but not closure. Our model does both inside one likelihood.

The idea gains weight from a finding in cancer genomics. SubNetDL propagated mutations separately within each tumor subclone and found that the most important functional drivers were inversely related to centrality in a protein interaction network. Heterogeneity-aware analysis there overturned hub-centric conclusions. The microbial network literature, for example the keystone and co-exclusion analyses in pine phyllosphere communities and the sparse log-ratio networks in DSS colitis and colorectal carcinogenesis, has never asked whether its hubs survive the same treatment.

Relation to Prior Work

Two groups of work come closest, and each covers only half of what we need.

The first group infers networks that differ by subpopulation. JRDNN-KM jointly clusters single cells and estimates group-specific gene networks, with shared hidden units that pool strength across groups. DeepCSCN extracts cell-type-specific co-expression from learned embeddings. Both require observations that are already separated into individual cells. A stool sample has no such separation: every observation is itself a mixture. We borrow JRDNN-KM's shared-plus-specific structure and its use of the network to inform grouping. We change the unit of grouping from "which cluster a cell belongs to" to "what fraction of each niche a sample contains."

The second group unmixes bulk measurements without estimating networks. TAPE learns cell-type proportions and cell-type-specific expression with an autoencoder whose decoder is a signature matrix. MEAD gives identifiability conditions and valid inference for proportions under platform scaling. Count Bridges treats single-cell counts as latent variables in an EM procedure, where each E-step sample is projected so that it sums exactly to the bulk count. None of these estimates covariance or precision within a component. All of them assume counts whose total carries meaning, whereas microbial sequencing depth is arbitrary.

We reuse the Count Bridges idea of latent component counts constrained to sum to the observation. We replace its integer birth-death generator with a closed multinomial observation layer, which makes the model indifferent to sequencing depth. Finally, themetagenomics already fits topic models that decompose OTU tables into co-occurring subcommunities. Its topics are mixture components with fixed profiles and no internal covariance, so the model cannot say which edges live inside a topic. It supplies our initialization and one of our baselines.

Research Program

We will build a hierarchical model with three layers. In the first, each sample draws niche masses $a_k$. In the second, each niche draws a log-normal abundance vector whose mean is shared across samples and whose precision matrix is sparse; a shared component plus niche-specific deviations lets rare niches borrow strength. In the third, the observed reads are multinomial draws from the closed sum $\sum_k a_k y_k$.

Fitting uses Monte Carlo EM. The E-step samples latent per-niche read allocations conditioned exactly on the observed reads, a multinomial analogue of Count Bridges' projection-guided sampling. The M-step updates proportion parameters and runs a graphical lasso on each niche's imputed log-abundances.

Identifiability of the mixing and within-niche terms is the main technical risk. We handle it in two regimes. In the reference-guided regime, a few physically fractionated or spatially resolved samples fix the niche means $\mu_k$, as single-cell atlases do in transcriptomic deconvolution. In the unsupervised regime, $K$ is chosen by held-out likelihood, and we report the stability of edges across restarts and across bootstrap resamples of samples.

The model exposes two outputs we will study. The first is a per-edge mixing score: the fraction of an edge's bulk association explained by the second term of the identity above. The second is a hub-survival analysis that recomputes degree and betweenness centrality inside each niche network.

Baselines

The naive reference point is a Spearman correlation network on relative abundances, thresholded by permutation p-value. This is how networks are built in studies such as the germ-free mouse colonization work and the Poyang Lake phytoplankton study. It suffers from both closure and mixing, so it anchors the scale.

The strongest existing microbial method is the sparse log-ratio network estimation used by Zhang et al. and Nakatsu et al.. We will run it through the same graphical-lasso solver our M-step uses, with its penalty tuned by the same stability-selection budget. Beating it means higher area under the precision-recall curve for recovering within-niche edges on data with known ground truth. We will also run it as a robustness check with the heavy-tailed multi-view estimator MVTLASSO, since contamination-robust estimation is a plausible alternative explanation for any gain.

Because the question is really about hidden structure, the most dangerous competitor is a sparse-plus-low-rank graphical model on log-ratios. This model absorbs latent confounders into a low-rank term. The file cites no instance of it, so we implement it ourselves with the same tuning budget.

The two-step competitor fits themetagenomics topics first and then either adds topic proportions as covariates in the log-ratio network or builds a separate network per dominant-topic cluster.

As an oracle upper bound on the spatial benchmark, we run JRDNN-KM on the unpooled single cells. This shows how much of the recoverable network our method reaches from pooled data alone. For JRDNN-KM, the paper reports only clustering accuracy on single cells ("ARI and NMI values close to 1.00"). No reported result was found on Lacuna for network-edge recovery from pooled mixtures by any method, so all network-recovery numbers will come from our own matched runs.

The key ablation removes the closure layer by fitting the same mixture model to log-ratio-transformed data with linear mixing. It isolates whether mixing in count space before closure, the central modeling choice, is what matters.

Why not just…?

Why not just use a sparse-plus-low-rank model, which already soaks up hidden confounders? This is the real threat. If niche mixing were linear in log-ratio space, a low-rank term of rank $K$ would absorb it. The identity above says mixing is linear in absolute abundance, and closure followed by the logarithm makes it curved. We therefore predict that the low-rank term will leak residual mixing edges into the sparse part, especially for niche-specific, abundant taxa. The semi-synthetic experiments settle this directly by varying niche contrast. If the low-rank model matches ours on edge recovery, we will say so. Our remaining contribution would then be the niche-specific networks and mixing scores, which a single pooled low-rank-corrected graph cannot provide.

Why not just cluster samples by topic and build a network per cluster? Samples are not members of one niche; each is a blend. Hard assignment leaves every cluster with a mixture of mixtures and starves small clusters of samples. JRDNN-KM already showed that clustering on mean expression before network estimation misses groups distinguished by wiring (JRDNN-KM). The two-step baseline tests whether this failure carries over to mixtures.

Why not just sequence finer, physically separating mucosa from lumen or sampling micro-aggregates? Where that is affordable it is the gold standard, and we use such samples as references and as validation. But existing cohorts, including bulk soil surveys and stool studies, cannot be resampled. The practical question is how much of their published network structure to trust, and only a computational decomposition can answer that retroactively.

Why not just apply Count Bridges or TAPE to microbial tables? They would estimate niche profiles but not niche networks. They also treat total counts as physical, which sequencing depth is not. Our closure ablation measures the cost of that assumption.

Proposed Experiments

The first experiment is a transcriptomic ground truth that tests the heterogeneity-masking claim with real within-component structure. We pool cells from the MERFISH mouse brain data used by Count Bridges and the Xenium breast cancer data used by GNN-C2L into pseudo-bulk samples whose regional composition varies. Each spatial domain plays the role of a niche and each gene the role of a taxon. We then downsample the pooled counts multinomially to impose arbitrary depth. The ground truth is the sparse partial-correlation network estimated within each domain on the unpooled cells. Metrics are edge AUPRC, sign agreement, and Spearman correlation between true and estimated degree ranks.

The second experiment is a microbial semi-synthetic test. We take the generalized Lotka-Volterra dynamics and 17 interaction modules that MDSINE2 inferred from gnotobiotic mice, assign modules to two to four simulated niches with their own parameters, simulate each niche, and pool with varying niche masses and sequencing depths. This yields known signed interactions under realistic dynamics.

The third experiment uses real cohorts without ground truth: the IBD, oral cancer, and time-series 16S datasets analyzed by themetagenomics, plus the mucosal colorectal cohort of Nakatsu et al.. Here we report the distribution of mixing scores, which published-style hubs lose centrality once niches are separated, and whether the recovered niche proportions predict Crohn's disease at least as well as topic proportions, which themetagenomics reports outperform raw abundances in F1.

Three results would be decisive. First, on the pooled MERFISH and MDSINE2-based data, our method must recover within-niche edges better than both sparse log-ratio networks and the sparse-plus-low-rank model, and the margin must grow with niche contrast, as the covariance identity predicts. If the margin does not grow with contrast, the mechanism is wrong. Second, the closure ablation must lose most of that margin; otherwise the modeling choice that distinguishes this work from Count Bridges and TAPE is not doing the work. Third, in the simulations, bulk hubs whose high degree came from the mixing term must be exactly those the model flags with high mixing scores, and in real cohorts a meaningful share of reported hubs must drop in within-niche centrality. If hubs are stable, that is itself a useful negative result, because it would license the keystone interpretations the field already makes.