When clinicians sample bacteria at several places along an anatomical tract — say, the vagina, uterus, fallopian tubes, peritoneal fluid, and endometrial tissue — they get one snapshot per patient and no follow-up. From that snapshot, machine learning classifiers readily flag disease with impressive accuracy, but they cannot say whether the microbes they flag are causing the disease or merely colonizing tissue that the disease has already reshaped. This is the core limitation of cross-sectional biological studies: the diseased body creates new living conditions for microbes, so the observed microbial shifts are often downstream consequences, not causes. The microbiome literature calls this the environmental selection trap, and it recurs across every disease domain in the file: constipation (Characteristics of functional constipation and analysis of intestinal microbiota in children aged 0–4 in Zunyi region), COPD (Koirobi Haldar et al.), COVID-19 (Yang Han et al.), sleep apnea (Tingyuan Zeng et al.), and endometrial cancer (Distinct Microbial Signatures along the Female Reproductive Tract in Endometrial Cancer Patients), whose authors state outright that their five-site cross-sectional design cannot establish causal direction.
This proposal imports a solving idea from a seemingly unrelated field: multi-microphone audio processing. When several microphones hear the same sound, each channel carries the source signal transformed by a known spatial transfer — a delay, an attenuation. Because the geometry between microphones is known, the differences between channels reveal where the source is, separating "what the sound is" from "how space changed it on the way here." I propose to treat anatomical sampling sites as channels of a spatial array: a taxon that grows locally at a site leaves a different signature across sites than a taxon that was physically transported from an upstream site, and the anatomy supplies the known "geometry." A transport-coupled dynamical model fit to a single snapshot can then separate the two — narrowing, though not eliminating, the causal ambiguity that makes cross-sectional microbiome biomarkers scientifically fragile.
Core Claim
The claim is concrete: in multi-site cross-sectional microbiome data, passive translocation and local niche-driven growth are statistically distinguishable, because transport couples sites along known anatomical directions while local growth does not. A taxon appearing at a downstream site in amounts predictable from its upstream abundance plus a transport rate is a passenger; a taxon whose site-local dynamics shift between patients and controls, beyond what upstream changes and transport explain, is a candidate local responder; a perturbation that originates at the most upstream site and propagates downstream is a candidate driver signal. No existing method tests any of this: BEEM-Static infers ecological interactions from cross-sectional data but only within a single, well-mixed compartment (BEEM-Static: Accurate inference of ecological interactions from cross-sectional microbiome data); the Spatial Deconfounder handles multi-location causal inference but assumes smooth confounding over continuous geographic space and has never touched biology (Spatial Deconfounder: Interference-Aware Deconfounding for Spatial Causal Inference); Celcomen disentangles intra- from inter-cellular effects in spatial transcriptomics but enforces acyclic structure and, by its authors' own statement, cannot always resolve causal direction (Estimation of single-cell and tissue perturbation effect in spatial transcriptomics via Spatial Causal Disentanglement).
A secondary claim gives the proposal its methodological spine: the scalar diversity statistics that cross-sectional studies collapse across sites are erratic precisely because they mix these two regimes. The file documents the contradiction directly — gut diversity drops in autism spectrum disorder (Chao1 richness 140 ± 46.1 versus 168 ± 54.8, p = 0.025; Jiangbo Ying et al.), shows no significant change in major depressive disorder (Xuan Wang et al.), and rises in endometrial cancer and pediatric constipation (Distinct Microbial Signatures along the Female Reproductive Tract in Endometrial Cancer Patients; Characteristics of functional constipation). A site-resolved, transport-aware model predicts when each trend should appear, instead of treating the inconsistency as noise.
From Microphone Arrays to Anatomical Tracts
The audio literature supplies the design intuitions, and I want to be explicit that I am importing principles, not equations. First, known geometry turns cross-channel contrasts into identification: the Binaural Angular Separation Network separates a target speaker purely from delay differences between two microphones, suppressing side interference by over 40 dB and gaining 2.4 dB BSS-SDR over a three-microphone baseline while running in 31.8 ms on a phone CPU (Binaural Angular Separation Network). The analog: anatomical adjacency and flow direction play the role of inter-microphone delay, turning cross-site contrasts into a translocation signature. Second, geometry should be an input, not a fixed architecture: PhaseCoder feeds raw microphone coordinates into a transformer and localizes sound with 7.44° mean absolute error on LOCATA (PhaseCoder); the analog is encoding site positions along the tract rather than hard-coding a five-site network. Third, and most cautionary, naive channel handling destroys the signal of interest. Multichannel speech enhancement without cross-channel attention reaches a perceptual quality score (PESQ) of only 2.93 where attention-based integration reaches 3.11, and plain single-channel processing reaches 2.63 (AMDM-SE) — the analog is that concatenating site features into a classifier, without modeling how sites couple, forfeits site attribution. Likewise, optimizing the wrong single-channel objective actively harms the downstream task — single-channel BLSTM enhancement worsens speech-recognition word error rate from 21.40% to 28.99% until multi-channel privileged distillation repairs it to 23.34% (Student-Teacher Learning for BLSTM Mask-based Speech Enhancement), and the standard separation objective SI-SDR erases interaural level cues until a binaural model optimizing plain SNR preserves them to 0.19 dB error while lifting SNR improvement from roughly 10.2 dB to 15.6 dB (Real-time binaural speech separation with preserved spatial cues). These are warnings that a pooled, site-blind objective will quietly discard exactly the information this proposal needs.
Research Program
The model extends the equilibrium machinery of BEEM-Static across sites with a transport layer. Let $x_{s,i}$ be the abundance of taxon $i$ at anatomical site $s$ in one patient's snapshot, $b_{s,ij}$ the effect of taxon $j$ on taxon $i$'s growth at site $s$, $a_{s,i}$ the site-local carrying capacity, and $\tau_{s' \to s}$ a transport rate along the directed anatomical edge from site $s'$ to $s$. At cross-sectional quasi-equilibrium, $\frac{dx_{s,i}}{dt} \approx 0$:
$$x_{s,i}\Big(a_{s,i} + \sum_j b_{s,ij}\, x_{s,j}\Big) + \sum_{s' \neq s} \tau_{s' \to s}\, x_{s',i} \;-\; \sum_{s'' \neq s} \tau_{s \to s''}\, x_{s,i} \;\approx\; 0.$$
The inference procedure alternates, expectation-maximization style, between estimating absolute biomass (the step BEEM-Static performs with per-sample medians) and sparse regression of the local parameters $(a, b)$ and transport rates $\tau$, with $\tau$ constrained to the known anatomical adjacency (vagina to uterus to tubes to peritoneal cavity) but with free magnitudes. Three read-outs follow directly from the fitted parameters. A passenger at site $s$ is a taxon whose equilibrium abundance there is fully explained by inflow terms, with its local growth row near zero. A local responder is a taxon whose local parameters differ between patients and controls beyond what upstream shifts plus fitted transport predict — the statistical fingerprint of the environmental selection trap. A propagating perturbation is a case-control change that appears most strongly at the upstream-most affected site and decays along the transport direction — a candidate driver, weaker than longitudinal evidence but strictly more informative than anything a pooled classifier can say. I will also extend BEEM-Static's equilibrium filter spatially: patients whose full five-site profiles deviate from joint spatial equilibrium are down-weighted, replacing its single-compartment filter, which the authors report tolerates up to 40% non-conforming samples with minimal degradation (BEEM-Static).
The second deliverable is a benchmark. BEEM-Static's synthetic generalized Lotka-Volterra simulator is the existing standard, and its reported results define the bar: interaction-recovery AUC-ROC above 0.88 against below 0.63 for all ten correlation-based methods tested, sign sensitivity above 80% and sign precision near 100% (BEEM-Static). I will extend that simulator to multiple coupled compartments with controllable transport strength, known ground-truth interaction graphs, and taxa injected either as local growers or passive translocators, so that every claim above becomes measurable. Third, the real-data proof of concept is the female genital tract cohort of Xu et al. — 59 women sampled at five anatomical sites, with a random forest diagnostic AUC of 0.88 from vaginal biomarkers (Distinct Microbial Signatures along the Female Reproductive Tract in Endometrial Cancer Patients). It is, per the file, the only multi-site cross-sectional microbiome cohort available, and its authors' explicit statement of unresolved causal direction is exactly the gap this model attacks. Finally, curatedMetagenomicData — the 4,617-sample gut resource BEEM-Static uses for its human validation (BEEM-Static) — serves two roles: pretraining site-local growth priors for the gut, and a negative control, since single-site data cannot identify transport terms and the model should degrade gracefully to a BEEM-Static-equivalent there.
Proposed Experiments
Experiment one is ground-truth recovery on the multi-site simulator. Metrics are interaction-recovery AUC-ROC, sign sensitivity, and sign precision, directly comparable to BEEM-Static's reported 0.88-plus AUC, above-80% sensitivity, and near-100% precision, and to the correlation floor below 0.63. Success means matching those numbers while additionally recovering transport rates, which no prior method attempts — no reported result found on Lacuna for transport-rate recovery in microbiome inference, because the task is new.
Experiment two is the passenger/local classification task on the simulator: for each taxon at each site, predict local-grower versus translocator. There is no external number to beat here — no reported result found on Lacuna for this task — so the decisive comparisons are internal: the same model with transport terms forced to zero, and the same model with transport free of anatomical constraints. If the constrained version does not beat both, the geometry claim is empty.
Experiment three is leave-one-site-out generalization on the Xu et al. cohort: hold out an anatomical site, and predict its microbial profile from the remaining sites through the fitted transport structure. This is the microbiological counterpart of the graph-splitting discipline that SpaCE introduced to prevent spatial leakage from inflating geographic causal estimates (SpaCE: The Spatial Confounding Environment) — standard models look strong under random splits and fail under spatially separated ones. A transport-coupled model should predict a held-out site from its anatomical neighbors; a pooled classifier has no mechanism to.
Experiment four is the diagnostic-accuracy illusion, made empirical. Rebuild the random forest of Xu et al. on pooled site features and confirm its 0.88 AUC under random splits; then evaluate under leave-one-site-out splits and under the case/control contrast decomposed by the transport model. The prediction, grounded in SpaCE's finding that standard learners overfit spatially correlated structure (SpaCE) and in the file's observation that high AUCs of 0.9551 in autism and 0.919 in depression studies mask unresolved causal direction (Jiangbo Ying et al.; Xuan Wang et al.), is that pooled AUC degrades under site-held-out evaluation while the transport model holds its ground and additionally labels which biomarkers are translocated passengers.
Experiment five is the transport-strength sweep that adjudicates the strongest skeptical alternative (below): vary simulated transport rates from negligible to dominant and track every method's error. Experiment six is the single-site negative control on curatedMetagenomicData: with one site, transport is unidentifiable and the model must reduce to BEEM-Static behavior, which is itself the correctness check rather than a failure.
Baselines
The naive reference point is a pooled co-occurrence or correlation network across all sites and patients, the default analysis in the cross-sectional literature. It requires no modeling and is the honest "do nothing" comparator; its expected performance is defined by BEEM-Static's evaluation, where all ten state-of-the-art correlation methods scored below 0.63 AUC-ROC on ecological ground truth and could not infer interaction signs at all (BEEM-Static). Any value added by this proposal must appear above that floor.
The strongest existing methodological competitor is BEEM-Static itself, run two ways: pooled across sites (treating sites as replicates) and per-site with networks merged afterward. It is the strongest because it is the only method with demonstrated directed-interaction recovery from cross-sectional snapshots — AUC-ROC above 0.88, sign sensitivity above 80%, precision near 100% on its own benchmark (BEEM-Static) — and my model is deliberately its extension, so the comparison isolates exactly the new ingredient. It will be run with its published settings on identical simulated and real data, given the same tuning budget, and evaluated on its own metrics plus the new transport and passenger metrics. Beating it means higher interaction-sign precision at matched recall once transport is active, correct recovery of transport rates, and correct passenger labels — none of which its design can output, since it models each community as a closed, well-mixed container.
The applied competitor is the random forest diagnostic pipeline of Xu et al., which reports an AUC of 0.88 using vaginal biomarkers on the endometrial cancer cohort (Distinct Microbial Signatures along the Female Reproductive Tract). It will be retrained with the same features, folds, and tuning budget, under both random and leave-one-site-out splits. Beating it does not mean a higher random-split AUC — that would prove nothing about causality — but matching its diagnostic accuracy while surviving the site-held-out split and producing site-of-origin attributions it cannot.
Two ablations isolate the proposal's key ingredients. Setting all transport rates to zero reduces the model to per-site BEEM-Static-style inference, showing how much the cross-site coupling earns. Freeing the transport matrix from anatomical constraints tests whether known geometry matters or whether any learned coupling suffices — the difference between this proposal and a generic multi-view model. As a further stress test, Celcomen (Estimation of single-cell and tissue perturbation effect in spatial transcriptomics) will be run with sites as nodes, to check whether an off-the-shelf spatial causal disentangler already solves the problem.
Why not just…?
Why not just run BEEM-Static at each site and merge the networks? This is the most serious alternative, and I will not argue around it. If transport between anatomical sites is weak, per-site BEEM-Static is sufficient and this proposal's added machinery is unjustified. The transport-strength sweep in experiment five is designed to settle it: as simulated transport grows, inflow at a downstream site becomes statistically indistinguishable from local growth or local interactions within a single-compartment model, so per-site BEEM-Static should misattribute translocation as interaction and its sign precision should degrade, while the coupled model should not. If per-site BEEM-Static shows no degradation at any transport level, the core claim is wrong and the proposal should be abandoned — I state that as a falsifiable outcome, not a risk to be managed.
Why not just use a standard classifier on pooled multi-site features? Because diagnostic accuracy is not the quantity in question. The file's evidence is consistent on this: pooled classifiers reach 0.88 in endometrial cancer, 0.919 in depression (0.800 on external validation), and 0.9551 when microbiota are combined with questionnaire scores (Distinct Microbial Signatures along the Female Reproductive Tract; Xuan Wang et al.; Jiangbo Ying et al.) — and every one of those papers concedes the causal direction is unresolved. SpaCE demonstrates the mechanism empirically: standard learners overfit spatially correlated noise and fail when evaluated on spatially separated splits (SpaCE). Experiment four makes the distinction concrete rather than rhetorical, and if the pooled classifier survives leave-one-site-out evaluation with attribution intact, that result would genuinely favor the simpler path.
Why not just collect longitudinal data? A longitudinal cohort would solve the ordering problem more directly than any snapshot model, and I concede the point in principle. In practice, the only multi-site cross-sectional cohort available here took five samples per woman across 59 participants (Distinct Microbial Signatures along the Female Reproductive Tract); multi-site serial sampling of internal anatomy is invasive and rare, so the realistic choice is between extracting more from existing snapshots and waiting for data that may never be collected. The method also inherits the equilibrium assumption that BEEM-Static shows is tolerable up to 40% non-conforming samples (BEEM-Static), and the spatial equilibrium filter is designed to extend that tolerance; acute inflammatory states where equilibrium genuinely fails remain out of scope, and I say so rather than claim otherwise.
Why not just apply the Spatial Deconfounder or Celcomen off the shelf? Both are closer than they look and both fall short in documented ways. The Spatial Deconfounder reconstructs latent confounders from neighbor treatment vectors via a conditional variational autoencoder, but it assumes confounding varies smoothly over continuous geographic space, and its own evaluation covers sea ice and semi-synthetic county data, not directed transport along discrete anatomical tracts (Spatial Deconfounder) — anatomy is not smooth, and flow direction is exactly the structure a smooth prior erases. Celcomen separates cell-autonomous from neighbor-driven effects in spatial transcriptomics, but it enforces a directed acyclic graph, cannot represent the reciprocal host-microbe feedback loops that define these ecosystems, and its authors state it cannot always resolve the direction of a causal link (Estimation of single-cell and tissue perturbation effect in spatial transcriptomics). Running Celcomen with sites as nodes is nevertheless a real threat worth one honest run, and it is included among the baselines rather than dismissed.
Decisive Checks
Two results would decide the proposal's fate. The transport-strength sweep is the decisive internal test: the coupled model must beat per-site BEEM-Static in interaction-sign precision at high transport strength and match it at low transport strength, or the added structure buys nothing. The leave-one-site-out prediction on the Xu et al. cohort is the decisive external test: predicting a held-out anatomical site's microbial profile from its neighbors through fitted transport is something no pooled or per-site method can do by construction, and it converts the sites-as-channels analogy into a falsifiable prediction. The honest limitations are the equilibrium assumption shared with BEEM-Static, the small real-data cohort of 59 women, and the fact that even a perfect transport decomposition narrows rather than removes causal ambiguity — a propagating upstream perturbation is a candidate driver, not a proven one. That narrowing, from "these biomarkers exist" to "these biomarkers grow here and were carried there," is what cross-sectional microbiome studies currently lack and what this proposal aims to supply.
References
- Characteristics of functional constipation and analysis of intestinal microbiota in children aged 0–4 in Zunyi region
- The sputum microbiome is distinct between COPD and health, independent of smoking history
- The active lung microbiota landscape of COVID-19 patients through the metatranscriptome data analysis
- Correlation between gut microbiota and metabolomics under intermittent hypoxic conditions
- Distinct Microbial Signatures along the Female Reproductive Tract in Endometrial Cancer Patients
- BEEM-Static: Accurate inference of ecological interactions from cross-sectional microbiome data
- Spatial Deconfounder: Interference-Aware Deconfounding for Spatial Causal Inference
- Estimation of single-cell and tissue perturbation effect in spatial transcriptomics via Spatial Causal Disentanglement
- The Gut Microbiota in Young Adults with High-Functioning Autism Spectrum Disorder and Its Performance as Diagnostic Biomarkers
- Utilizing metagenomic profiling and machine learning model to identify bacterial biomarkers for major depressive disorder
- Binaural Angular Separation Network
- PhaseCoder: Microphone Geometry-Agnostic Spatial Audio Understanding for Multimodal LLMs
- AMDM-SE: Attention-based Multichannel Diffusion Model for Speech Enhancement
- Student-Teacher Learning for BLSTM Mask-based Speech Enhancement
- Real-Time Binaural Speech Separation with Preserved Spatial Cues
- SpaCE: The Spatial Confounding Environment