When to Sequence: Event-Triggered 16S Sampling from Cheap Proxies

Tracking a microbial community over time, whether a bioreactor, a shrimp hatchery tank or an animal gut, usually means reading its composition with 16S rRNA gene sequencing. This assay amplifies and reads a bacterial "barcode" gene and reports what fraction of the community each taxon makes up. It is slow, destructive and costly per sample, so longitudinal studies sequence on a fixed calendar or whenever a clinic visit happens. Event-triggered control is a branch of control theory with a different habit. It sends a measurement over a network only when the receiver's estimate has drifted far enough from reality to matter, and it can prove that this cuts communication without losing stability. This note borrows that logic for a different kind of decision: whether today's banked sample should be sequenced.

The core idea fits in two sentences. Between sequencing runs we keep two cheap estimates of community composition: one comes from a fitted ecological model run forward from the last sequenced sample, the other from a fast physical readout such as flow cytometry or total cell density. We trigger a sequencing run only when these two estimates disagree by more than a threshold, measured in log-ratio coordinates that respect the fact that compositions sum to one. That threshold carries a slowly refilling budget reserve borrowed from dynamic event-triggering, plus guaranteed minimum and maximum gaps between runs. The claim to be tested is that, at a matched number of sequencing runs, this disagreement trigger reconstructs community trajectories more accurately and catches regime shifts sooner than calendar sampling, than triggers that watch the proxy alone, and than triggers that watch model uncertainty alone.

Core Claim and What Is New

The file contains no work that schedules 16S sequencing with an event-triggered rule. The closest ingredients sit separately. Event-triggered learning watches the gaps between communication events in an event-triggered estimator. It retrains the model when those gaps stop matching what the model predicts, using concentration bounds that need linear Gaussian dynamics with independent, identically distributed gaps. We reuse the principle that a measurement event is also a model-validity signal. We drop the linear Gaussian assumption, because microbial dynamics are nonlinear and live on the simplex, and we spend the event on an expensive destructive assay rather than on a network packet. Predicting the Presence and Abundance of Bacterial Taxa in Environmental Communities through Flow Cytometric Fingerprinting supplies the proxy. It maps cytometric fingerprints to per-taxon presence and relative abundance with a two-stage random forest and boosted-regression pipeline, but it uses this map only as a passive substitute and never decides when ground truth is worth buying. Compositional Lotka-Volterra describes microbial dynamics in the simplex supplies the model. It recasts Lotka-Volterra interaction dynamics in additive log-ratio coordinates, so trajectories can be fitted from relative abundances alone, and it stays open-loop and retrospective. The new element is the trigger. It fires on the innovation between model and proxy, it is defined in log-ratio geometry rather than on Euclidean proportions, and it carries a budget-aware dynamic threshold whose gaps between runs can be bounded below.

Concretely, let $x(t)$ be the composition of $p$ taxa and $z = \mathrm{alr}(x)$, the vector of $\log(x_i/x_p)$ against a reference taxon. Let $\hat z_m(t)$ be the compositional Lotka-Volterra forecast propagated from the last sequenced sample, and $\hat z_c(t)$ the proxy-derived estimate with calibration covariance $\Sigma_c$. A run is triggered when

$$\| \hat z_c(t) - \hat z_m(t) \|_{\Sigma_c^{-1}} \ \ge\ \delta + \eta(t), \qquad \dot\eta = -\lambda\,\eta + \rho \ \text{between events}, \quad \eta \leftarrow \eta - \kappa \ \text{at events},$$

where $\delta$ is a fixed floor, $\eta$ is a reserve that grows at rate $\rho$ while no runs are spent and shrinks by $\kappa$ after each run, and $\lambda$ sets how fast the reserve forgets. The Mahalanobis weighting means the trigger discounts disagreement in taxa the proxy predicts poorly. That matters because the proxy's average $R^2$ is modest. The minimum inter-event time follows the dwell-time argument of A unifying Lyapunov-based framework for the event-triggered control of nonlinear systems. On a compact set of log-ratio states, the model's vector field is Lipschitz and the proxy's rate of change is bounded, so the disagreement grows at a bounded rate. It therefore needs at least $\delta/L$ time to cross the floor, where $L$ is that rate bound. We add a hard maximum gap so a proxy that is blind to some shift cannot silence sequencing forever.

Research Program

The work has three parts. First, we build the two estimators on each dataset: compositional Lotka-Volterra fitted on whatever samples have been sequenced so far, and a proxy map recalibrated every time a new sequenced point arrives. Recalibration is the event-triggered learning step, so each trigger buys both a state reset and a model update.

Second, we handle the assay delay. Results arrive days after sampling, which breaks the instant-delivery assumption that Event-triggered Control and Online Learning for Networked Systems under Computational Delays flags as damaging in networked loops. In a monitoring setting, though, samples can be frozen at every proxy reading. When a run is triggered at time $t$, we may sequence the banked sample from $t$, or from earlier, and then re-propagate the model from that point once the data arrive. We test whether sequencing the banked sample where the disagreement first began to rise, rather than the one at the crossing time, improves change localization.

Third, we analyze what the trigger costs and guarantees: the dwell-time bound, the maximum detection delay for a step change in the interaction matrix given the hard maximum gap, and the sensitivity of the run count to $\delta$, $\rho$ and $\kappa$.

Closed-loop actuation, meaning dosing or dilution driven by barrier functions as in Safety-Critical Event Triggered Control via Input-to-State Safe Barrier Functions, is deliberately out of scope for this first study. No dataset in the file contains interventions chosen by a controller, and claims about it would be untestable.

Baselines

The naive reference point is calendar sequencing. The same number of runs $K$ that the trigger ends up using is spread evenly over the horizon, so both policies spend an identical budget. The trigger has to beat this to justify any machinery. For context only, the control literature reports large savings over periodic schemes on mechanical plants: 54.34 versus 13.31 executions on a scalar nonlinear benchmark in A unifying Lyapunov-based framework for the event-triggered control of nonlinear systems. Those numbers say nothing about biology. No reported result was found on Lacuna for event-triggered or adaptive 16S sampling of any kind.

The strongest competitor on proxy quality is the cytometric fingerprinting pipeline of Predicting the Presence and Abundance of Bacterial Taxa in Environmental Communities through Flow Cytometric Fingerprinting used alone, with no further sequencing after initial calibration. On the hatchery data it reports 88.13% ± 4.78% presence/absence accuracy and an average $R^2$ of 0.35 ± 0.24 for relative abundance across the top 50 taxa. Our method beats it if, after spending $K$ extra runs, held-out abundance $R^2$ rises clearly above 0.35 on the same taxa and splits. The comparison must also be made per run spent, against calendar-spent runs.

Two ablations isolate the key ingredient, which is disagreement between model and proxy. The proxy-change trigger fires when the cytometric fingerprint, or total cell density, moves more than a threshold from its value at the last run. The model-uncertainty trigger fires when the forecast's predictive variance, estimated by bootstrap over fitted parameters, exceeds a threshold. All triggers share the same estimators and the same hard minimum and maximum gaps, and each threshold is tuned on a held-out split to hit the same expected budget $K$. A further ablation computes our trigger with a Euclidean norm on raw proportions, to test whether log-ratio geometry matters. A last one runs our trigger without recalibrating the proxy at events, to measure the learning component.

Why not just…?

Why not just sequence on a fixed schedule? This is the null hypothesis, and it may win when community change is spread evenly over time. Our trigger only pays off when change is bursty, as in antibiotic onset or pathogen invasion. The budget-matched comparison on the antibiotic and C. difficile datasets against the diet dataset tests exactly this. If the trigger shows no advantage on bursty data, the idea fails.

Why not just trigger on the cheap proxy changing? This is the most serious threat, because it is simple and needs no ecological model. Its weakness in principle is that the proxy moves for reasons that leave composition unchanged, such as total biomass swings or phenotype shifts. Rubbens and colleagues note that cell optical properties change in mixed communities. The proxy also stays still when optically similar taxa swap. A disagreement trigger ignores proxy motion the model already expects. If the proxy-change ablation matches our method at equal budget, we will report that the simpler trigger suffices and that the model is only useful for reconstruction.

Why not just sample where the model is most uncertain? Uncertainty-driven sampling is standard in experimental design. It spends runs early, while the model is poorly fitted, and then goes quiet. It cannot notice a regime shift the model never anticipated, which is exactly the failure that Identifying stationary microbial interaction networks based on irregularly spaced longitudinal 16S rRNA gene sequencing data concedes for stationary network methods after antibiotics. The detection-delay metric on known perturbation times measures this directly.

Why not just use flow cytometry and never sequence again? Proxy-only $R^2$ of 0.35 ± 0.24 leaves most abundance variance unexplained. The proxy also cannot separate strains that look alike optically. The baseline comparison above shows how much each additional run buys.

Proposed Experiments

All main experiments are retrospective subsampling studies. Densely sequenced series serve as ground truth. Each policy chooses which time points to "sequence" from the bank, and every unchosen point is predicted.

The hatchery dataset from Rubbens and colleagues is the primary testbed, because it pairs cytometry with 16S over 18 days. The three mouse datasets from the compositional Lotka-Volterra paper (antibiotic, diet and C. difficile infection) provide perturbations with known timing. There the proxy is total bacterial concentration, the only cheap signal they contain. Synthetic communities simulated from generalized Lotka-Volterra dynamics, with an interaction matrix that switches at a known time and a simulated proxy of tunable quality, let us vary proxy $R^2$ and see where the trigger stops helping. The irregular cystic fibrosis cohort is too sparse for subsampling. We use it only to check that the hard maximum gap does not distort network estimates relative to its reported five communities.

The metrics are Aitchison-distance error at unsequenced times, per-taxon $R^2$ against the 0.35 proxy baseline, delay in detecting known perturbations, and the number of runs used.

The decisive checks are three. First, at matched budget, does the disagreement trigger lower reconstruction error and detection delay compared with calendar sampling on the bursty datasets, while roughly tying on the diet data? Second, does it beat the proxy-change and uncertainty-only ablations? This is the result that justifies the idea. Third, in simulation, below what proxy quality does the advantage disappear? That threshold tells practitioners whether their proxy is good enough to bother.