Microbiome sequencing tells us which microbes are present in a sample; metabolomics tells us which small chemicals are present. Linking the two is how we learn who produces or consumes what, and nearly all of that linking is done with a single correlation number computed on a one-time snapshot: if a microbe and a metabolite rise and fall together across patients, they get an edge in a network. The trouble is that both layers move. A patient's gut community and its chemistry shift over weeks, samples arrive at irregular intervals because patients miss visits, and medications — antibiotics, steroids, biologics — are given precisely to the patients whose systems are already disturbed, so the correlations we measure are tangled up with why the drug was prescribed. Three bodies of work each hold one piece: cross-domain correlation analysis builds microbe–metabolite networks but only statically (MiMeNet); longitudinal biomarker tracking shows that a biomarker's trajectory, not its current level, carries prognostic information that single snapshots miss (Pan et al., Zhang and Yang et al.); and retrospective cohort methodology supplies the tools — propensity scores — for untangling confounding by indication in already-collected records (Tewari et al.).
This proposal puts those pieces together in one model. The core claim: by learning a coupled, continuous-time dynamical system over both the microbial layer and the chemical layer, fit to already-collected longitudinal cohorts with medication-confounding removed by inverse-probability weighting, we can (a) predict future metabolite levels better than static network methods, (b) recover known ulcerative colitis biology that snapshot correlation finds, and (c) expose whether past dysbiosis leaves a measurable "legacy" on disease course that current levels hide — the microbial analogue of the legacy effect Pan et al. found for inflammation, where patients whose hsCRP was falling still carried elevated metabolic-syndrome risk (adjusted HR 1.408, 95% CI 1.115–1.779) relative to the low-stable group.
What exists and what is missing
The closest prior work brackets this idea from both sides. MiMeNet (Reiman et al.) trains a neural network to predict the whole metabolome from a static microbial snapshot, and it beats linear models handily — on the PRISM inflammatory bowel disease cohort its mean Spearman correlation across metabolites rose from 0.108 (linear MelonnPan) to 0.309, and the count of well-predicted metabolites nearly doubled from 198 to 366 — but it has no concept of time. BGOB (Lyu et al.) handles time well: it is a continuous-time model built on neural ordinary differential equations that interpolates irregularly sampled microbiome trajectories, improving RMSE by 28.07% over standard GRU-ODE-Bayes at initial time points and, more strikingly, rescuing statistical power — standard differential-abundance tests on the raw irregular IBDMDB and early-caries data found zero significant species, while the same tests on BGOB-smoothed trajectories recovered real disease drivers like Streptococcus mutans. But BGOB models the microbial layer alone; it never touches metabolites. Snyder et al. did collect both layers longitudinally (106 participants, ~4 years, 16S sequencing plus LC-MS/MS metabolomics, distinguishing infection from health with 88.6% accuracy), but analyzed them with linear mixed models and pairwise co-associations, in a small protocolized prospective cohort, with no treatment-confounding adjustment. Zakerska-Banaszak et al. mapped a rich Spearman network in ulcerative colitis — lithocholic acid and indoles in remission, succinate in flare — but on a cross-sectional group of 65 people, comparing different individuals rather than tracking transitions, and without balancing medication use across subgroups. Nothing in the file couples the two layers dynamically, and nothing applies retrospective causal machinery to microbe–metabolite association discovery. That is the gap, and it is exactly the tension Lyu et al. exposed: discrete, snapshot-style statistics lose power on irregular real-world sampling.
Research Program
The object to be built is a two-layer extension of the BGOB machinery, which I will call a coupled GRU-ODE. Let $z_b$ be a latent state summarizing a patient's microbial community and $z_c$ a latent state summarizing their metabolite profile; each evolves in continuous time, and each layer's rate of change reads the other layer's state, plus a treatment channel:
$$\frac{dz_b}{dt} = f_\theta(z_b, z_c, u_i(t)), \qquad \frac{dz_c}{dt} = g_\phi(z_b, z_c, u_i(t))$$
Here $u_i(t)$ is a weighted medication indicator for patient $i$ — which drug class (antibiotic, corticosteroid, 5-ASA, biologic) was active at time $t$ — weighted by the inverse of the propensity to receive it given the patient's recorded clinical state, exactly the confounding-by-indication logic that Tewari et al. applied to prostate cancer treatment comparisons and that Propensity Score Alignment repurposed as a cross-modal coordinate system (we reuse the weighting idea, not the optimal-transport machinery, which requires discrete perturbation labels our clinical cohorts lack). The Bayesian "sureness" weighting from BGOB carries over so that predictions from sparse histories are appropriately uncertain, and the non-negativity constraint stays, since both abundances and concentrations are positive quantities. The coupling terms are the scientific payload: their magnitudes, delays, and signs are the dynamic, directional version of what a static Spearman edge asserts without evidence.
The primary testbed is IBDMDB — 130 individuals with Crohn's disease and ulcerative colitis tracked for up to 57 weeks with longitudinal profiles and clinical records, already collected, hence analyzable retrospectively without the cost and dropout that prospectively serializing deep multi-omics would incur. The Snyder prediabetes cohort (106 participants, ~4 years, paired 16S and metabolomics) serves as the second, denser dataset; where the two cohorts' metabolite platforms disagree, MIRTH (Freeman et al.), which harmonizes and imputes metabolites across heterogeneous mass-spectrometry studies via rank transformation and matrix factorization, provides a principled bridge. The cross-sectional Polish UC cohort is not a training set; it supplies ground-truth biology (lithocholic acid, indoles via microbial tryptophanase, succinate in flare, Akkermansia positively correlated with short-chain fatty acids) against which recovered network edges are checked.
The third deliverable is the legacy-effect analysis. Following the trajectory logic of Pan et al. and Zhang and Yang et al., who showed that albumin's in-hospital change (an S-shaped mortality curve with inflection points at −3.27 and 0.92 g/L, peak OR 2.12) outperformed admission levels, we derive trajectory features from the fitted dynamics — velocity of each microbe–metabolite module, cumulative area under the curve, past dysbiosis spikes — and test whether they predict flare onset or remission failure beyond what the current-snapshot levels predict.
Baselines
The do-nothing reference point is pooled-snapshot Spearman correlation, the method used by Zakerska-Banaszak et al. and the wider correlation literature (Feng et al., Wang et al.): correlate taxa and metabolites across all sampled timepoints as if they were independent observations. For the outcome analysis the naive reference is a baseline-levels-only survival model, mirroring what Pan et al. showed to be insufficient. Beating this baseline means recovering, with held-out predictive validity, associations that pooled correlation either misses or gets backwards; given that BGOB's authors found zero significant species with exactly this style of testing on raw irregular data, the naive baseline is expected to be weak, but it must be run to make that concrete on our data rather than assumed.
The strongest existing single method is MiMeNet, the published state of the art for predicting metabolites from microbes; on PRISM it reports a mean Spearman correlation of 0.309 and 366 well-predicted metabolites, against 0.108 and 198 for the linear alternative. Those numbers are on PRISM, not IBDMDB or the Snyder cohort — no reported result found on Lacuna for MiMeNet on either of our datasets — so we re-run MiMeNet ourselves on identical splits, and it also enters as a per-timepoint dynamic baseline (MiMeNet applied independently at each observed visit). The comparison is kept fair by fixing the evaluation protocol to MiMeNet's own (mean Spearman correlation across held-out metabolites, count of well-predicted metabolites), using the same train/test partitions, and giving every method the same tuning budget in random-search configurations and the same compute. Beating MiMeNet means a higher mean held-out Spearman correlation on future timepoints — predicting the metabolome ahead, which a static model cannot do without degradation, not merely reconstructing the concurrent snapshot.
The strongest combined competitor, and the most dangerous one, is the two-stage pipeline: run BGOB (its codebase is public) on the microbiome stream, then fit MiMeNet on the interpolated trajectories. This composes the best temporal tool with the best cross-domain tool using no new ideas, and if it matches the coupled model, the coupling was unnecessary. It gets the same data, tuning budget, and protocol. Two ablations isolate the proposal's key ingredients: the same architecture with the cross-layer coupling severed (each layer runs its own ODE, reading only its own history and the treatment channel), which isolates the value of joint dynamics; and the same model with propensity weights removed (raw medication indicators only), which isolates the value of the retrospective causal adjustment by measuring how many edges flip or vanish when confounding by indication is left in.
Why not just…?
Why not just correlate snapshots, as the field already does? The literature is unusually direct on this. Pooled or single-timepoint statistics on irregularly sampled longitudinal data lost essentially all power in BGOB's evaluation — zero significant species on raw IBDMDB data — and the prognostic literature shows snapshots miss structure that trajectories reveal: the legacy effect in Pan et al. (HR 1.408 for a decreasing inflammatory trajectory) and the S-shaped albumin curve in Zhang and Yang et al. are both invisible at any single timepoint. Experiment 1 below includes the pooled-Spearman baseline precisely so this is demonstrated, not asserted.
Why not just run a two-stage pipeline — BGOB for the microbiome, then MiMeNet on the smoothed output? This is the real threat, and I will not argue around it. Three things argue the coupled model earns its keep: the metabolite stream is itself irregularly sampled with dropouts, which single-modality BGOB cannot smooth, so the second stage inherits noisy inputs; coupling lets metabolite observations inform microbial trajectories (useful when stool sequencing has gaps but serum chemistry was drawn); and only a joint system estimates lead–lag delays between the layers, which is the directional information static correlation famously lacks. But the honest position is that this is an empirical question, and the two-stage baseline is run under identical conditions on all datasets with a pre-registered criterion: if its held-out metabolite prediction and association recovery match the coupled model within a pre-specified margin everywhere, the coupled architecture is not justified and we will say so.
Why not just use linear mixed effects models, as Snyder et al. did on data with both layers? Snyder's linear mixed models worked on a small, protocolized cohort with quarterly scheduled visits; on real-world irregular sampling, the discrete-method failure documented by BGOB is the relevant caution, and MiMeNet showed that per-metabolite linear models forfeit cross-metabolite information transfer (0.108 versus 0.309 mean correlation on PRISM). We still include an LME-plus-pairwise-correlation baseline in every experiment so the comparison is made rather than cited.
Why not just use routine clinical labs and skip sequencing altogether? The strongest evidence for the marginal value of omics here is negative evidence from the clinical side: on 3,509 ACCORD participants, machine learning on exhaustive baseline clinical records left most lipid volatility unexplained — HDL-C change at $R^2 \approx 0.040$, triglycerides at $\approx 0.188$ — and those authors explicitly pointed to the gut microbiome as the missing driver (ACCORD ML study). Clinical panels alone demonstrably leave most of the interesting variance on the table.
Proposed Experiments and Decisive Checks
Experiment 1 is held-out metabolite prediction. On IBDMDB and the Snyder cohort, mask metabolite observations at the most recent visits, and score every method by mean Spearman correlation and count of well-predicted metabolites under MiMeNet's protocol. The reference number to contextualize (not to match, since it is from a different dataset) is MiMeNet's 0.309 mean SCC and 366 well-predicted metabolites on PRISM. Decisive check: the coupled model must beat its own decoupled ablation and the two-stage pipeline on future-timepoint prediction; if the coupled and decoupled models tie, the cross-layer reading of states is doing nothing.
Experiment 2 is association recovery. Test whether the fitted coupling terms recover the UC biology established by Zakerska-Banaszak et al. — the lithocholic-acid and indole signatures of remission, succinate in flare, the Akkermansia/short-chain-fatty-acid positive edges — and whether the recovered network assigns direction (producer versus consumer) where snapshot correlation cannot. Decisive check: IPTW versus unweighted edges; a large shift in the edge set under weighting means confounding by medication was materially distorting the naive network, which is itself a publishable finding.
Experiment 3 is the legacy-effect test. Fit survival models for flare onset and remission failure on trajectory features (module velocity, cumulative exposure, past dysbiosis spikes) versus current-snapshot levels alone, on IBDMDB. Decisive check: a trajectory model that significantly improves prediction over snapshot levels, in the manner of Pan's legacy effect and Zhang and Yang's albumin curve, establishes that microbe–metabolite dynamics carry prognostic information that current association practice discards; if snapshot levels saturate prediction, the legacy hypothesis fails for this system and the proposal's third leg does not hold.
References
- MiMeNet: Exploring microbiome-metabolome relationships using neural networks
- Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome:A Retrospective Cohort Study
- Association between albumin changes and prognosis in older patients with acute myocardial infarction
- Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach
- Deep-learning-based interpolation of longitudinal microbiome data powers biologically informative discovery
- Longitudinal multi-omics of host–microbe dynamics in prediabetes
- New potential biomarkers of ulcerative colitis and disease course — integrated metagenomic and metabolomic analysis among Polish patients
- Propensity Score Alignment of Unpaired Multimodal Data
- Integrated metabolomics and metagenomics analysis of plasma and urine identified microbial metabolites associated with coronary heart disease
- The Effect of Salt Reduction on the Microbial Community Structure and Metabolite Composition of Cheddar Cheese
- A Machine Learning Approach to Predicting Heterogeneous Lipid Responses to Statins in Type 2 Diabetes