The gut microbiome talks to distant organs — brain, liver, kidney — through small molecules: fermentation products like butyrate, bacterial bile-acid transformations, choline-derived trimethylamine oxidized in the liver to TMAO, and amino-acid depletion that starves the brain of neurotransmitter precursors (Decoding TMAO in the Gut-Organ Axis: From Biomarkers and Cell Death Mechanisms to Therapeutic Horizons, Bile acid metabolites predict multiple sclerosis progression and supplementation is safe in progressive disease). Machine learning has started to map this conversation: neural models now predict which metabolites a microbial community produces, and these predictions carry clinical signal — latent features from one such model separate Crohn's disease patients from controls with an AUC of 0.932 (Deep in the Bowel: Highly Interpretable Neural Encoder-Decoder Networks Predict Gut Metabolites from Gut Microbiome, MiMeNet: Exploring microbiome-metabolome relationships using neural networks). But every existing model performs this mapping in flat Euclidean space, even though both sides of the interface are trees: the input side is a phylogenetic taxonomy of microbes, and the output side is a branching chemical taxonomy of metabolites plus directed precursor-to-product cascades (primary bile acids becoming secondary ones, TMA becoming TMAO).
Hyperbolic geometry — a curved space whose volume grows exponentially with radius, exactly matching how tree descendants multiply — lets low-dimensional embeddings represent such hierarchies with far less distortion than flat space: a 5-dimensional Poincaré embedding reached a mean average precision of 0.823 on the WordNet noun hierarchy where a 200-dimensional Euclidean model managed only 0.168 (Poincaré Embeddings for Learning Hierarchical Representations). This matters here because the relevant cohorts are small (150–250 subjects), so parameter efficiency is not a luxury but the difference between a trainable model and an overfit one. I propose to build the first model that maps the microbial phylogeny onto the metabolic chemical hierarchy inside a single shared hyperbolic space, with directed geometric containment representing precursor-product transformations along the gut-to-organ transport chain, and to test whether this geometry improves metabolite prediction, annotation of uncharacterized metabolites, and cross-tissue readout against the strongest flat-space baselines.
Core Claim
The microbe-to-metabolite relationship is not an arbitrary function between two vectors; it is a map between two trees, filtered by a directed transport chain (gut lumen → portal circulation → hepatic transformation → systemic blood → target organ). Encoding both endpoints in a shared Lorentz hyperbolic space — with entailment cones, geodesically convex regions that nest children inside ancestors, enforcing both taxonomic and biochemical containment — should yield more accurate metabolite prediction per parameter, more stable latent features, and geometrically grounded annotation of unannotated metabolomic "dark matter," in the small-sample regime where existing flat models already operate. No published work does this: a search of the literature found no application of hyperbolic embeddings to microbe–metabolite or multi-tissue signaling data. The closest works are either flat multi-omics predictors (Deep in the Bowel: Highly Interpretable Neural Encoder-Decoder Networks Predict Gut Metabolites from Gut Microbiome, MiMeNet: Exploring microbiome-metabolome relationships using neural networks) or hyperbolic embeddings of static taxonomies without any flux prediction (Hyperbolic Multimodal Representation Learning for Biological Taxonomies, Hyperbolic Genome Embeddings). This proposal combines the two, and the combination is not a relabeling: the directed, asymmetric containment structure needed for precursor-product cascades has never been used to model biochemical transport between compartments.
Research Program
The model, which I will call a dual-cone map, has four components. First, each microbial species is embedded in the Lorentz hyperboloid (the upper sheet of a hyperboloid in Minkowski space, chosen over the Poincaré ball because its closed-form geodesics and lack of a boundary singularity make optimization more stable (Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry)), initialized from the known phylogenetic tree with a reduced-learning-rate burn-in phase, which the original Poincaré embedding work found necessary to avoid premature collapse to the boundary (Poincaré Embeddings for Learning Hierarchical Representations). Second, each metabolite cluster is embedded in the same space, with annotated metabolites placed by entailment cones over their chemical taxonomy (superclass → class → subclass) using the stacked entailment loss machinery from Hyperbolic Multimodal Representation Learning for Biological Taxonomies; unannotated clusters start free and are positioned only by their predicted producers. Third, known directed transformations — primary to secondary bile acids, choline to TMA, TMA to TMAO, tryptophan to indole derivatives — add a flux cone loss requiring each product to lie inside the cone of its precursor, extending the asymmetric containment of Hyperbolic Entailment Cones for Learning Hierarchical Embeddings from pure taxonomy to biochemistry. Fourth, a Lorentz multinomial logistic regression head (Hyperbolic Neural Networks) maps species embeddings to predicted centered-log-ratio metabolite abundances, optionally with the non-negative weight constraints that made the Sparse-NED encoder interpretable as an additive microbe-metabolite graph.
The full training objective, with all symbols defined, is
$$\mathcal{L} \;=\; \mathcal{L}_{\text{pred}} \;+\; \lambda_p\, \mathcal{L}_{\text{ent}}^{\text{phylo}} \;+\; \lambda_c\, \mathcal{L}_{\text{ent}}^{\text{chem}} \;+\; \lambda_f\, \mathcal{L}_{\text{cone}}^{\text{flux}},$$
where $\mathcal{L}_{\text{pred}}$ is squared error between predicted and observed log-ratio metabolite abundances, the three $\mathcal{L}_{\text{ent}}/\mathcal{L}_{\text{cone}}$ terms are margin-based containment losses placing species within genus cones, metabolites within chemical-class cones, and transformation products within precursor cones, and the $\lambda$'s are weights tuned by nested cross-validation. Two applications follow directly from the geometry: unannotated metabolite clusters inherit putative chemical classes from the cones that contain them (a geometric version of the guilt-by-association annotation that MiMeNet performs with Euclidean biclustering), and radial depth in the manifold becomes a readable coordinate — general hubs near the origin, specific downstream products near the boundary — which the Lorentz model work showed can recover hidden hierarchy from similarity data alone, as it did for the Enron executive structure (Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry).
A design guard precedes everything: Hyperbolic Genome Embeddings showed hyperbolic networks beat Euclidean ones on 37 of 42 tree-like genomic tasks but gave no gain on promoter detection, because combinatorial regulatory logic is not tree-like, and the Lorentz model paper warns that forcing curvature onto cyclic or grid-like data distorts it. So I will first measure Gromov $\delta$-hyperbolicity (a standard tree-likeness ratio) on the metabolite co-abundance graph, the chemical taxonomy, and the microbe–metabolite bipartite graph. If the co-abundance structure turns out not tree-like, the hyperbolic prior will be restricted to the taxonomy and chemical-hierarchy sides, with a flat Euclidean head for the combinatorial part — a hybrid the field currently lacks (Hyperbolic Genome Embeddings).
Proposed Experiments
The core benchmark is the Franzosa et al. (2018) inflammatory bowel disease cohort: 220 subjects (164 Crohn's and ulcerative colitis patients, 56 controls), 201 microbial species, 8,848 metabolic clusters, the same data on which Sparse-NED was validated (Deep in the Bowel: Highly Interpretable Neural Encoder-Decoder Networks Predict Gut Metabolites from Gut Microbiome). Using identical folds and preprocessing, the decisive metrics are: Pearson correlation for the top 10 best-predicted metabolites (Sparse-NED reports 0.74), feature stability index (Sparse-NED reports 0.79), AUC for Crohn's-versus-healthy classification from latent features (reported 0.932), and AUC for antibiotic-use prediction (reported 0.866). The second benchmark is the IBD PRISM cohort (121 patients, 34 controls), where MiMeNet raised the mean Spearman correlation for annotated metabolites from 0.108 under the linear MelonnPan baseline to 0.309, and the count of well-predicted metabolites from 198 to 366 (MiMeNet: Exploring microbiome-metabolome relationships using neural networks); the targets there are a mean Spearman correlation above 0.309 and more than 366 well-predicted metabolites. Third, a dark-matter annotation task: hide the chemical class labels of annotated clusters, predict them from cone membership, and compare against Euclidean biclustering guilt-by-association run on the same data — no reported result found on Lacuna for annotation accuracy on this task, so this experiment establishes the reference number itself. Fourth, a cross-tissue readout on the major depressive disorder cohort of 111 patients and 102 controls with paired fecal microbiome, fecal metabolome, and multimodal brain MRI (Multi-omics analyses of the gut microbiome, fecal metabolome, and multimodal brain MRI reveal the role of Alistipes and its related metabolites in major depressive disorder): can the hyperbolic microbe and metabolite coordinates predict the MRI-derived measures that the original study linked to Alistipes (abundance difference at $z = 4.845$, $P = 1.27 \times 10^{-6}$) and recover any of its 23 significant mediation pathways, such as cholic acid acting on anxiety through visual-cortex functional connectivity? Success is measured as recovered-pathway overlap against that paper's list. Ablations run throughout: a Euclidean twin with identical architecture and identical hierarchical losses at dimensions 5 to 200 (this isolates curvature from hierarchy supervision, the single most important control), a hyperbolic model with the cone losses removed, and a dimension sweep at 2, 5, and 10 to substantiate the parameter-efficiency claim.
Baselines
The naive reference point is per-metabolite linear regression, which on the Franzosa cohort achieves a Pearson correlation of only 0.46 for the top 10 best-predicted metabolites, with canonical correlation analysis at 0.47 (Deep in the Bowel: Highly Interpretable Neural Encoder-Decoder Networks Predict Gut Metabolites from Gut Microbiome); the dual-cone map must clear this trivially or the geometry is irrelevant. The strongest existing competitor is Sparse-NED, the pruned non-negative neural encoder-decoder that holds the reported record on the Franzosa data (top-10 Pearson 0.74, stability index 0.79, Crohn's-versus-healthy AUC 0.932), with MiMeNet the record holder on PRISM (mean Spearman 0.309, 366 well-predicted metabolites). Both will be run from their published implementations on exactly our folds, with the same centered-log-ratio preprocessing, the same nested cross-validation tuning budget, and the same single-GPU compute; the hyperbolic model's Riemannian optimization adds wall-clock overhead that I will report honestly rather than hide. Beating the baseline means a paired, cross-validated improvement in top-10 Pearson on Franzosa and mean Spearman on PRISM at conventional significance, with stability no worse; if prediction only matches but the model wins decisively on dark-matter annotation and needs an order of magnitude fewer latent dimensions, I will count that as a qualified success and label it as such. The key ablation-as-baseline is the Euclidean twin described above: identical losses, identical architecture, flat latent space. This is the control that decides whether the contribution is curvature or merely hierarchy supervision, and I commit in advance to reporting it with equal prominence; if the Euclidean twin matches the hyperbolic model on every metric, the paper's claim reduces to "hierarchical losses help," which is a weaker but still publishable finding, and I will frame it that way.
Why not just…?
Why not just use the established linear tools (CCA, PLSR, LASSO)? The file's own numbers rule this out as sufficient: on the Franzosa cohort linear regression and CCA reach top-10 Pearson correlations of 0.46 and 0.47 against 0.74 for the neural encoder-decoder, and LASSO showed low feature stability where the constrained network achieved a stability index of 0.79 (Deep in the Bowel: Highly Interpretable Neural Encoder-Decoder Networks Predict Gut Metabolites from Gut Microbiome); on PRISM the linear MelonnPan baseline reached a mean Spearman of only 0.108 against MiMeNet's 0.309 (MiMeNet: Exploring microbiome-metabolome relationships using neural networks). Linear methods will still be included as the naive reference, but they are not competitive on this question.
Why not just scale up Euclidean dimensions? This is the most reasonable objection, and the evidence from other domains is that geometry is not replaceable by dimension: on WordNet, 5-dimensional Poincaré embeddings reached mean average precision 0.823 while a 200-dimensional Euclidean model reached 0.168 (Poincaré Embeddings for Learning Hierarchical Representations), and hyperbolic convolutional networks with about 4.6 million parameters matched DNA language models with up to 2.5 billion on genomic benchmarks (Hyperbolic Genome Embeddings). But those are other domains, and in small omics cohorts the extra capacity of high dimensions is a liability, not an asset. The question is settled empirically inside this proposal: the Euclidean twin at matched architecture and swept dimension is a first-class baseline, and if it performs as well, the hyperbolic machinery was unnecessary and the experiments will say so.
Why not just build a Euclidean graph neural network over the phylogenetic tree? Message passing along tree adjacency does inject the phylogeny, and PathFinder-style graph transformers already infer inter-cellular signaling from graph structure (PathFinder: a novel graph transformer model to infer multi-cell intra- and inter-cellular signaling pathways and communications), so this is a genuine threat. Two things distinguish the proposal: the graph-network node representations still live in flat space, where polynomially growing volume crowds the thousands of leaf metabolite clusters that must stay separated (the volume argument from Poincaré Embeddings for Learning Hierarchical Representations), and adjacency alone does not encode the asymmetric, directed containment that precursor-product cascades require. I am also honest that geometry is domain-dependent: on BIOSCAN-1M, hyperbolic models beat Euclidean baselines on unseen-species DNA retrieval (species-level harmonic mean 70.8%) but lost on image retrieval (Hyperbolic Multimodal Representation Learning for Biological Taxonomies). A Euclidean graph network over the same trees will therefore be added as a fourth baseline, and the δ-hyperbolicity pre-check will tell us in advance whether the data's structure favors the curved model at all.
Why not just build a mechanistic differential-equation model of the gut-organ loop? The closed-loop delay-differential-equation model of the gut-brain axis already delivers what machine learning cannot: bifurcation tipping points (barrier leakage collapse near $1.52\ \text{min}^{-1}$) and channel-capacity analysis (Gut-Brain Axis as a Closed-Loop Molecular Communication Network). But it ingests six lumped state variables, not thousands of taxa and metabolic clusters per patient, and its own authors note that frequency-domain linearization misses the nonlinear circadian dynamics. The two approaches answer different questions — mechanism per patient versus population-scale prediction — and this proposal deliberately stays on the data-driven side while borrowing that paper's framing of the axis as directional compartmental transport, which is exactly what the cone structure encodes. Fully closing the loop (host feedback onto the gut) remains out of scope: the available cohorts are cross-sectional, and I would rather claim the static dual-tree map honestly than promise dynamics the data cannot support.
Decisive Checks
The proposal succeeds if, on identical folds, the dual-cone map beats the reported 0.74 top-10 Pearson and 0.932 Crohn's AUC on the Franzosa cohort and the 0.309 mean Spearman on PRISM, while the Euclidean twin ablation shows the gap is genuinely geometric rather than an artifact of the hierarchical losses. It fails informatively if the Euclidean twin matches it — in which case the fallback contribution is the cone-based annotation protocol for metabolomic dark matter, evaluated on the held-out-annotation task where no prior number exists. Either outcome is decided by experiments already specified, on cohorts already in hand, with a pre-registered geometric sanity check that tells us before training whether hyperbolic space was ever the right home for this data.
References
- Decoding TMAO in the Gut-Organ Axis: From Biomarkers and Cell Death Mechanisms to Therapeutic Horizons
- Bile acid metabolites predict multiple sclerosis progression and supplementation is safe in progressive disease
- Deep in the Bowel: Highly Interpretable Neural Encoder-Decoder Networks Predict Gut Metabolites from Gut Microbiome
- MiMeNet: Exploring microbiome-metabolome relationships using neural networks
- Poincaré Embeddings for Learning Hierarchical Representations
- Hyperbolic Multimodal Representation Learning for Biological Taxonomies
- Hyperbolic Genome Embeddings
- Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry
- Hyperbolic Entailment Cones for Learning Hierarchical Embeddings
- Hyperbolic Neural Networks
- Multi-omics analyses of the gut microbiome, fecal metabolome, and multimodal brain MRI reveal the role of Alistipes and its related metabolites in major depressive disorder
- PathFinder: a novel graph transformer model to infer multi-cell intra- and inter-cellular signaling pathways and communications
- Gut-Brain Axis as a Closed-Loop Molecular Communication Network