When microbiologists model how a resident gut community repels an invader, they almost always reach for the generalized Lotka–Volterra (gLV) equations, in which each species' growth rate is a linear function of everyone else's densities. When neuroscientists model how a neural circuit keeps its sensitivity across a huge range of input intensities, they reach for divisive normalization, in which each channel's response is divided by a pooled measure of its neighbors' activity. These are, structurally, the same problem — a bounded system whose members compete for a shared, finite resource — but the two literatures have never been connected. This proposal builds that connection: we replace the linear competition sum in gLV with a divisive normalization pool, producing a "Divisive Lotka–Volterra" (dLV) model in which per-capita growth is automatically scaled down as the total community fills the available resource space. The payoff is threefold: bounded, mostly stable dynamics where standard gLV is catastrophically unstable; a mechanistic handle on the "invasion threshold" that defines colonization resistance; and, we hypothesize, recovery of species-level self-limitation from purely relative (compositional) sequencing data, which existing compositional models provably cannot identify.
The opportunity is sharpened by two failures documented in the literature. First, Stein et al., inferring gLV parameters from longitudinal mouse microbiome data, found that 98% of theoretically possible microbial compositions in the inferred system are dynamically unstable — the linear pairwise form itself, not the inference, is the problem. Second, the iterative Lotka–Volterra work proved that self-interaction coefficients (the terms encoding intra-species carrying capacity) are mathematically unidentifiable from compositional data, forcing them to be set to zero — removing density-dependent self-limitation from every downstream ecological prediction. Divisive normalization enters exactly where both failures originate: its denominator is a saturating, resource-like suppression pool, and its self-gain is governed by a semi-saturation constant rather than a linearly canceling coefficient.
Core Claim
The claim is that a population dynamics model whose per-capita growth is divisively normalized by a weighted pool of community densities,
$$\frac{dx_i}{t} = x_i \left( m_i + r_i \, \frac{\sigma_i^\gamma}{\sigma_i^\gamma + \sum_j w_{ij} x_j^\gamma} \right),$$
predicts compositional microbiome trajectories at least as well as the strongest existing compositional models, while adding three things none of them have. Here $x_i$ is the density of taxon $i$, $r_i$ its maximum growth rate, $m_i$ a density-independent loss rate (flow, immune clearance, or death), $\sigma_i$ a semi-saturation constant setting the community density at which suppression begins to bite, $\gamma$ an expansive exponent, and $w_{ij} \ge 0$ the suppression weight exerted by taxon $j$ on taxon $i$ — the direct analogue of the interaction matrix in gLV and the suppression field in canonical divisive normalization. The additions are: bounded per-capita growth in $[m_i, m_i + r_i]$ regardless of community configuration, which structurally precludes the runaway divergence Stein et al. observed; an explicit, computable colonization-resistance criterion; and a self-limitation parameter $w_{ii}$ that enters the log-ratio dynamics nonlinearly and so may remain identifiable where gLV's $b_{ii}$ provably is not. The last point is a hypothesis we will test rather than assume, since the identifiability argument (nonlinear saturation leaves curvature signatures in log-ratio trajectories that a linear self-term cannot) could still fail when suppressors co-vary with the focal taxon.
The model also has a direct biological reading that gLV lacks. Colonization resistance is understood mechanistically as resource sequestration and niche pre-emption — residents consume nutrients and occupy attachment sites until an invader "encounters a fully booked ecosystem lacking the energetic resources or physical space required to establish a critical population density" (Ren et al.'s framing, Wang et al.'s demonstration that commensal pre-colonization starves Klebsiella). Divisive normalization is precisely the mathematics of a shared depletable pool: it is the same computation that Senn showed arises physically from vesicle depletion at a synapse, where high-frequency firing drains a finite resource pool and scales down subsequent efficacy. Resource competition and divisive gain control are the same equation in different units.
Research Program
The program has three parts: theory, inference, and application to colonization-resistance prediction.
The theory part addresses two open problems flagged in the source literatures. First, stability under asymmetry. Rawat, Heeger, and Martiniani proved unconditional stability of a recurrent divisive-normalization circuit (ORGaNICs) but only for identity recurrent weights, explicitly leaving open whether the guarantee survives asymmetric, competitive interaction topologies — exactly the topologies microbial communities have, since Coprobacillus suppresses C. difficile without reciprocation (Stein et al.). We will derive Lyapunov and boundedness conditions for dLV under asymmetric $w$: the key structural difference from gLV is that suppression saturates, so per-capita growth is bounded above by $m_i + r_i$, and global boundedness follows whenever $m_i < 0$ (every species dies in the absence of resources). We will characterize when asymmetric $w$ additionally yields a unique stable equilibrium versus multi-stability — the multi-stable regime Stein et al. found is clinically important, since antibiotics push the gut into an alternative basin permissive to pathogens, so we want sufficient conditions for both outcomes, not a blanket stability claim. Second, identifiability. We will perform a structural identifiability analysis of $(w_{ii}, \sigma_i)$ in the additive log-ratio (ALR) coordinates that compositional data force on us, following the apparatus Joseph et al. introduced for cLV. The mechanistic reason for hope: in gLV, a self-term $b_{ii} x_i$ contributes a species-specific constant to log-ratio growth that cancels against the reference taxon's tied self-term, whereas in dLV self-limitation contributes $-\log(1 + w_{ii} x_i^\gamma / \sigma_i^\gamma)$, a nonlinear function of the focal taxon's own trajectory whose curvature is not shared with any reference. We will verify this with Fisher-information profiles on synthetic data before trusting it on real cohorts.
The inference part is deliberately conservative: we fit dLV in ALR space with elastic-net regularization on the suppression weights (the recipe cLV validated for high-dimensional, collinear microbiome data), using the two-stage iterative initialization that the iLV framework showed fixes the accuracy loss of one-shot linear fits. Total biomass, when qPCR is unavailable, is treated as a latent variable with the semi-saturation constants absorbing the scale — a cleaner alternative to cLV's requirement of a reference taxon with constant absolute abundance, which the file notes is rarely met under dysbiosis. We will additionally explore constraining $w_{ij}$ by metabolic-overlap priors, motivated by Liu et al.'s finding that suppression weights track functional similarity in sensory circuits, and by Králová et al.'s metabolic microarrays showing that fungal exclusion follows divergent substrate-utilization profiles; where substrate-utilization data exists, similarity-weighted $w$ becomes a testable structural prior rather than a free parameter matrix.
The application part turns the model's fixed-point algebra into a colonization-resistance quantity. At equilibrium, taxon $i$ persists only if the resident suppression field it experiences exceeds a computable threshold:
$$\sum_j w_{ij} x_j^{*\gamma} > \sigma_i^\gamma \left( \frac{r_i}{|m_i|} - 1 \right),$$
which we call the resistance budget for invader $i$: a resident community repels an invader of that growth/loss profile if and only if its pooled suppression exceeds the right-hand side. This converts colonization resistance from a qualitative narrative into a per-invader scalar that can be computed from fitted parameters and, crucially, evaluated against held-out invasion experiments.
Proposed Experiments
The first experimental block uses the synthetic and small-ecosystem benchmarks from the iLV paper, which carry ground truth and reported numbers. On the synthetic oscillating and stabilizing benchmarks at 0%, 5%, and 10% noise, we measure trajectory RMSE and cosine similarity of recovered suppression weights against the true matrices, comparing to iLV's reported cosine similarities exceeding 0.7; the decisive addition is that the synthetic systems will carry planted self-limitation, so we can measure whether dLV recovers $w_{ii}$ where iLV is forced to assume $b_{ii} = 0$. On the snowshoe hare–Canadian lynx census data, the Stylonychia pustula–Paramecium caudatum biculture, and the cheese fermentation community, we compare trajectory RMSE against iLV's reported values of 0.109, 0.0673, and 0.149 respectively.
The second block targets the mouse cohorts. On the Diet dataset (stable biomass, compositional and absolute versions available), we test whether dLV trained on relative data alone matches the cLV result that Joseph et al. established there (interaction correspondence $R = 0.96$ against absolute gLV; predictive advantage over ALR linear models at $p = 0.008$ and over direct linear models at $p = 0.039$). On the antibiotic-perturbed mouse cohort challenged with C. difficile, the hardest test in the file, we ask whether dLV fitted to relative data alone approaches the 62% rank correlation in unseen subjects that Stein et al. achieved with absolute qPCR counts — that number is the absolute-data ceiling, and matching it from relative data would be the strongest single result. The gnotobiotic C. difficile challenge dataset was also used by cLV, but no reported result was found on Lacuna for cLV's numeric performance on it; we will report the first head-to-head numbers there, with the pre-registered prediction that dLV's resistance budget correctly separates pre-antibiotic (resistant) from post-antibiotic (permissive) communities.
The third block tests the resistance budget against clinical engraftment outcomes. The FMT cohort of NPJ (2017) provides labeled outcomes — donor Bacteroides strains persisting up to two years while abundant donor F. prausnitzii is blocked by recipient niche barriers — and the pediatric ASD cohort of Wu et al. provides another: donor Bifidobacterium failing to engraft despite a clinically effective protocol (CARS scores down 14.6%, $p = 0.0089$). We fit dLV to pre-transplant recipient trajectories, compute the resistance budget for each donor strain, and test whether predicted-repellent strains are the ones observed to fail. Success is discrimination materially better than an abundance-based rule (donor strains that are already abundant in the recipient), since the clinical puzzle is precisely that abundant donor strains sometimes fail to engraft.
Baselines
The naive reference point is fitting standard gLV directly to relative read counts (gLV_relative), the do-nothing approach most practitioners actually take; the file records that it yields significantly higher trajectory RMSE than iterative methods on all three iLV benchmarks and cosine similarities far below 0.7 under noise, so it calibrates the floor but cannot be the comparison that matters. The strongest existing competitor is the iterative Lotka–Volterra framework of PLoS Comput. Biol. (2025), which the file shows outperforms both gLV_relative and absolute gLV on the hare–lynx (RMSE 0.109), Stylonychia–Paramecium (RMSE 0.0673), and cheese (RMSE 0.149) benchmarks — it is the incumbent on exactly the compositional data our proposal targets, and beating or matching it on those same datasets with the same train/test splits is the primary success criterion. On the mouse cohorts, the strongest competitor differs by dataset: cLV on the Diet dataset, where it holds the reported advantage over linear alternatives, and Stein et al.'s absolute-data gLV on the antibiotic cohort, where its 62% rank correlation in unseen subjects is the number to approach from relative data alone. All baselines will be run on identical data, splits, and evaluation code, with the same grid-search budget over regularization strength and iteration count, and the same wall-clock cap, so differences reflect model structure rather than tuning effort. Finally, two ablations isolate the proposal's key ingredients: dLV with $w_{ii}$ clamped to zero (removing recoverable self-limitation, the identifiability ingredient) and dLV with the divisive pool replaced by an added quadratic interaction term $\sum_{jk} q_{ijk} x_j x_k$ in otherwise-standard gLV (removing the divisive structure while keeping a saturating-capable nonlinearity). The proposal's claim is specifically supported only if full dLV beats the $w_{ii}=0$ ablation on self-limitation recovery and beats or matches the quadratic variant on trajectory accuracy with fewer parameters and interpretable resistance budgets.
Why not just…?
Why not just use cLV, since it already handles compositional data? cLV solved the algebra of the simplex, and on stable-biomass data it is excellent ($R = 0.96$ against absolute parameters on the Diet dataset). But the file documents two structural weaknesses our proposal attacks: its formulation requires a reference taxon whose absolute abundance stays constant, a condition the authors themselves note is rarely met during dysbiosis or antibiotic intervention, and its fidelity degrades precisely in the regimes where colonization resistance matters most — large biomass swings such as antibiotic depletion followed by pathogen bloom. Moreover, iLV already showed cLV recovers true interaction matrices only moderately under realistic noise. dLV differs by having no reference taxon (the semi-saturation constants carry the scale) and by being the model whose dynamics remain bounded through those swings.
Why not just regularize gLV harder? Regularization improves the inference but not the dynamics: Stein et al. used Tikhonov regularization and still found 98% of configurations unstable in the inferred system, because instability is a property of the linear pairwise form, not of the estimator. A better-fit unstable model is worse for colonization-resistance prediction, since antibiotic perturbation predictions require integrating the dynamics forward through exactly the disturbances where unstable models diverge. Additionally, gLV requires absolute abundance data that most human cohorts lack.
Why not just train an LSTM or neural ODE on the time series? This is a real threat on pure trajectory prediction, and we will run it as a fourth baseline rather than argue it away. But the scientific question is not trajectory interpolation; it is whether the fitted system supports an invasion threshold, a per-strain resistance budget, and a carrying-capacity interpretation — quantities a black box does not expose. The file's own evidence supports this: Rawat et al. note that LSTMs match divisive circuits on sMNIST accuracy (~97.8%) while their parameters map to no biological variables at all, and Gao et al. found standard architectures failed completely to recover an inverse-square physical law from trajectories where divisive gain control succeeded (estimated constant 27.02 vs. true ~39.48). If the LSTM matches dLV's RMSE on the antibiotic cohort and we can extract nothing interpretable from it, that is our thesis confirmed, not refuted; if it also extrapolates correctly through the antibiotic perturbation and dLV does not, the proposal fails on its own terms.
Why not just add quadratic or higher-order interaction terms to gLV? This is the closest genuine alternative, since saturating behavior can be mimicked by polynomial terms. Three things distinguish the divisive form, and we commit to letting the quadratic ablation adjudicate rather than asserting them: the divisive pool has a mechanistic resource reading (a shared depletable substrate, matching how colonization resistance is actually understood and how Senn showed division arises from physical resource depletion), it bounds per-capita growth structurally where polynomials diverge at large densities, and it confers the scale-invariance property that makes the model well-posed on relative data. If the quadratic gLV matches dLV's trajectory accuracy on all benchmarks at similar parameter counts, the honest conclusion is that the identifiability and stability theory — not raw accuracy — is the contribution, and we would report it that way.
Decisive Checks
Three results would decide the proposal. First, on synthetic benchmarks with planted self-limitation, dLV recovers $w_{ii}$ with materially higher cosine similarity than the $b_{ii}=0$ ceiling that iLV is forced into; if Fisher-information profiles show $w_{ii}$ is also unidentifiable in dLV, the core hypothesis is wrong and we will report the negative result, which would itself settle an open question raised by the iLV identifiability proof. Second, on the antibiotic mouse cohort, dLV fitted to relative data alone achieves rank correlation in unseen subjects approaching Stein et al.'s 62% absolute-data result, and its resistance budget correctly flags post-antibiotic communities as permissive to C. difficile before the challenge outcome is shown to the model. Third, on the two clinical engraftment cohorts, the resistance budget separates persisting from blocked donor strains better than any abundance-based rule. Any two of the three, with the theory holding under asymmetric weights, would establish divisive normalization as a quantitative mechanism of colonization resistance — the first dynamics model in this literature whose instability problem and compositional-data problem share a single solution.
References
- Ecological Modeling from Time-Series Inference: Insight into Dynamics and Stability of Intestinal Microbiota
- Quantifying microbial interactions based on compositional data using an iterative approach for solving generalized Lotka-Volterra equations
- Evolving the olfactory system with machine learning
- Multiple-species biofilms as structuralized microbial communities for modulating microbiota homeostasis in human
- Gut–liver translocation of pathogen Klebsiella pneumoniae promotes hepatocellular carcinoma in mice
- A Synaptic Explanation of Suppression in Visual Cortex
- Unconditional stability of a recurrent neural circuit implementing divisive normalization
- Compositional Lotka-Volterra describes microbial dynamics in the simplex
- Target specific functions of EPL interneurons in olfactory circuits
- Competitive fungal commensalism mitigates candidiasis pathology
- Identification of donor microbe species that colonize and persist long term in the recipient after fecal transplant for recurrent Clostridium difficile
- Gut microbiota characteristics and therapeutic effects of fecal microbiota transplantation in children with autism spectrum disorder in central China: a combined cross-sectional and prospective study