Rainfall is the main control on how much carbon land plants fix in dry and temperate regions. Satellite studies show that water, more than temperature, sets year-to-year swings in grassland and forest productivity. Belowground, though, the soil microbes that decompose and release that carbon have only ever been studied at single moments. The best existing experiment brought prairie soil into the lab, set it to "dry" or "wet," and sequenced the community's messenger RNA at one endpoint. That design cannot say whether microbes respond to how wet the soil is now, or to the path the soil took to get there.
This note proposes a time-resolved experiment and a matching statistical test for that question. We would sample the same soils densely through controlled drying and rewetting cycles. We would then ask, gene by gene, whether expression at a given moisture level depends on the soil's moisture history: whether it is drying or rewetting, how fast it dried, and whether it has been dried before. The comparison would be made after removing changes that merely reflect which taxa grew or shrank.
The answer matters in practice. If expression depends only on current moisture, a single moisture-response curve per gene is enough to link climate models to soil microbial function. If expression shows hysteresis, meaning a different state at the same moisture depending on direction or history, then climate models need a memory term, and endpoint experiments have been systematically misleading.
Core Claim
The claim is narrow and testable. In mixed soil communities exposed to realistic drying–rewetting trajectories, a measurable fraction of genes will show path dependence: their expression differs at matched soil moisture depending on drying direction, drying rate, or prior drying. The key condition is that this holds only after each gene's counts are corrected for the abundance of the organism that carries it. Without that correction, community turnover will masquerade as regulation.
The new element is not a more complex model. It is a test that separates three things standard pipelines run together:
- Current-state response: expression tracks how wet the soil is now.
- History response: expression also depends on how the soil reached that moisture.
- Composition change: some taxa become more or less abundant.
What Prior Work Establishes and Leaves Open
Motivating the gap. Metaphenomic Responses of a Native Prairie Soil Microbiome to Moisture Perturbations incubated Konza Prairie soils from three hillslope positions under dry, wet, and field-moist conditions and profiled them with metatranscriptomics (sequencing all RNA from the soil community), 16S amplicons, and metabolomics. It found three things:
- Drying, not wetting, drove large coordinated changes, including DNA repair, nucleotide metabolism, and production of protective sugars such as trehalose.
- Community composition also shifted: Actinobacteria and Chloroflexi rose, while Acidobacteria and Bacteroidetes fell.
- Hillslope position conditioned the response.
Because it was a static endpoint design, it cannot separate regulation from regrowth or detect lag and hysteresis. The composition shift is exactly why abundance correction is necessary rather than optional.
The macro-scale literature motivates the stakes but works at the canopy only:
- Impacts of Extreme Precipitation and Diurnal Temperature Events on Grassland Productivity at Different Elevations on the Plateau found extreme drought hurt grassland productivity more than temperature extremes.
- Understanding Drought through Spatial-Temporal Learning forecasts soil-moisture and fluorescence-based drought indices across the contiguous United States. It treats soil as a water store with no biology.
Supplying the method or constraining the design. Two continuous-time papers show that snapshot designs miss cumulative effects:
- Revealing the temporal dynamics of antibiotic anomalies in the infant gut microbiome with neural jump ODEs models gut communities as continuously evolving between sparse samples. It found a "second hit" effect: a second antibiotic course caused disruptions lasting up to four months. We borrow its question, whether a repeated stress erodes resilience, and its idea of encoding perturbation history as model input. It works on diversity indices from 16S data, not gene expression.
- chronODE fits smooth rate curves to time-course chromatin and RNA data in single-organism mouse development. It shows that binning time distorts rate estimates, but it does not handle mixed communities or environmental forcing.
Constraining the design (compositional data).
- Predicting microbial community structure and temporal dynamics by using graph neural network models forecasts taxon abundances in wastewater plants from abundance history alone, deliberately without environmental drivers. Its failure on abrupt, sparse dynamics warns against heavy forecasting models for our short series.
- A novel approach to differential expression analysis of co-occurrence networks for small-sampled microbiome data shows that standard correlation-network tools fail below roughly fifty samples. We therefore avoid inferring interaction networks and represent community context through measured taxon abundances.
Supplying the counting model. For counts we reuse negative binomial models with empirical Bayes dispersion sharing, as in DESeq2 and edgeR (applied, for example, in Yisu Shi et al. and Yuhong Meng et al.). A negative binomial model is a count model whose variance grows faster than its mean, which RNA-seq requires. We add no new estimation machinery; the novelty is in the experimental contrasts and the covariates.
Research Program
Experiment. We would collect soils from two contrasting Konza hillslope positions, matching the two positions sequenced in the metaphenomic study. These would be placed in replicated microcosms with continuously logged moisture under three forcing regimes:
- Single cycle: one drying, followed by rewetting.
- Repeated cycle: two identical cycles, testing memory, the soil analogue of the second hit.
- Rate contrast: fast versus slow drying to the same final moisture, testing rate dependence.
The moisture trajectories would be scaled from observed soil-moisture series for the Kansas cells of DroughtSet (DroughtSet), so the dynamics are realistic rather than arbitrary step changes.
Sampling would be destructive and deliberately irregular, denser around rewetting where fast changes are expected. Each sample would receive metatranscriptomic and 16S sequencing. A pooled deep metagenome would let reads be assigned to genome bins, which are draft genomes reconstructed for individual taxa. A subset of samples would get metabolomics, keeping the trehalose and sugar readouts as an independent check.
Model. For gene $g$ in sample $i$, read count $Y_{gi}$ is negative binomial with mean $\mu_{gi}$ and gene-wise dispersion shrunk across genes:
$$\log \mu_{gi} = \log s_i + \log a_{\tau(g),i} + \beta_{g} + f_g(\theta_i) + \gamma_g^\top h_i$$
The terms are:
- $s_i$ is the library size factor.
- $a_{\tau(g),i}$ is the DNA-based abundance of the genome bin $\tau(g)$ carrying gene $g$. This is the abundance offset that converts changes in reads into changes in per-organism expression.
- $\beta_g$ is the baseline expression of the gene.
- $f_g(\theta_i)$ is a smooth spline in current soil moisture $\theta_i$.
- $h_i$ is a small, fixed set of history features computed from the logged moisture record: moisture averaged over the preceding 6, 24, and 72 hours, the sign of the current moisture change (drying or wetting branch), time since last rewetting, and number of prior dry-downs.
A gene is path-dependent if a likelihood-ratio test of $\gamma_g = 0$ rejects at a Benjamini–Hochberg false discovery rate of 5%. It is state-responsive if $f_g$ is non-flat. Genes falling in neither group are unresponsive.
Running the same model without the abundance offset shows how many apparent differentially expressed genes are really community turnover. Running it at the level of pathways, summed across taxa, asks whether path dependence is a property of the community rather than of particular organisms.
Baselines
Naive reference: pairwise endpoint testing. This reproduces the existing design: DESeq2 contrasts of fully dry and fully wet samples against field-moist samples, with no time or history. It is the reference because it is what the literature currently reports. It will run on the same libraries, restricted to the endpoint samples. The comparison is not "who finds more genes." It is how many of its calls the full model reassigns as pure abundance change or as history effects.
Strongest standard method: time-course negative binomial regression. This is DESeq2 or edgeR with a spline in clock time and a likelihood-ratio test. It is how careful labs analyse time courses today, and it is the real competitor. It gets identical counts, normalization, dispersion shrinkage, and the same spline degrees of freedom.
Our model beats it if two things hold:
- Its held-out log-likelihood is higher when one entire cycle is withheld, especially for the rate-contrast and repeated-cycle regimes, where clock time and moisture history come apart.
- In simulation it recovers planted path-dependent genes at higher power for the same false discovery rate.
Key ablation: moisture state only. This is our model with $\gamma_g$ fixed at zero. It isolates the main ingredient. If it matches the full model on held-out cycles and the history test finds almost no genes, we would report that soil microbial expression is effectively memoryless at these timescales. That is a valid and useful result, and it would justify a single moisture-response curve per gene in climate-linked models.
Composition ablation. Removing the abundance offset isolates the contribution of community complexity to apparent differential expression.
Matrix-decomposition baseline. Robust PCA, from Robust PCA based method for discovering differentially expressed genes, splits a log-transformed expression matrix into a low-rank background plus sparse perturbations. It is included because it is the file's strongest method for drought-stress gene discovery in a single species, where it beat sparse PCA on enriched Gene Ontology terms. It is run on the same matrix, using that paper's reported default regularizer $c=0.3$ plus a small tuning grid equal to the budget given our spline settings.
No published numbers to beat. There is no reported result on Lacuna for time-resolved soil metatranscriptomic differential expression under moisture forcing. The nearest reported figures are on different tasks and serve only as context: a neural jump ODE anomaly macro-F1 of 0.542 against 0.496 for the best baseline on infant-gut 16S data (neural jump ODE paper), and a median MAPE of 14.2–17.7% for abundance forecasting in wastewater plants (mc-prediction). Because of this gap, a simulation benchmark with known ground truth is part of the deliverable.
Why not just…?
Why not just put soil moisture in as a continuous covariate in DESeq2? That is our state-only ablation, and it is a genuine threat. If hysteresis is weak, it will perform as well. The rate-contrast and repeated-cycle regimes are built so that current moisture is identical while history differs. On those matched samples, a state-only model must predict identical expression. Any reproducible difference beyond replicate noise falsifies it, and the absence of such differences supports it. Either outcome is reported.
Why not just use a spline in time, the standard time-course pipeline? Clock time works within a single forcing trajectory. It cannot transfer across regimes, because "day 3" means different moisture in fast and slow drying. The leave-one-cycle-out evaluation tests exactly this. If time-splines still match our model there, the history features add nothing, and we would say so.
Why not just fit a neural ODE or neural jump ODE to the metatranscriptome? With a few hundred libraries and tens of thousands of genes, a flexible latent-dynamics model would be poorly identified. The forecasting paper's breakdown on abrupt, sparse data (mc-prediction) and the finding that classical models stay competitive below about a thousand samples (LLMs Can Learn the Language of the Microbiome) both point the same way. Interpretable per-gene tests are also what the scientific question requires. We will still fit a jump-ODE predictor on pathway-level summaries as a held-out-likelihood comparison. If it clearly beats the regression, that is evidence of dynamics our history features miss.
Why not just interpret the endpoint data more carefully? Endpoint data cannot separate regulation from regrowth. The shift toward Actinobacteria reported in the metaphenomic study alone could generate apparent drought-induced genes. The composition ablation quantifies how large that confound is.
Decisive Checks
The experiment is decisive along four lines.
Simulation recovery. Before any real data, we would simulate negative binomial counts from the real design, with planted abundance-driven, state-driven, and path-dependent genes. The pipeline must hold its stated false discovery rate and recover path-dependent genes at higher power than the time-spline baseline. If it cannot, the method is not ready.
Matched-moisture contrast. On real data, the central test compares samples at matched moisture on the drying versus wetting branch, and fast versus slow drying at the same endpoint. The prediction is a significant set of path-dependent genes enriched in the osmolyte and DNA-repair functions the endpoint study highlighted. The prediction fails if the set is no larger than in label-permuted histories.
Repeated-cycle memory. The second dry-down should either blunt the response (acclimation) or prolong it (resilience erosion, as in the gut second hit). The sign of this effect is itself the finding.
Metabolite validation. Path-dependent genes, and not just state-responsive ones, should improve held-out prediction of trehalose and sugar-alcohol levels. This ties the transcriptional memory to a measured functional consequence.
References
- Metaphenomic Responses of a Native Prairie Soil Microbiome to Moisture Perturbations
- Impacts of Extreme Precipitation and Diurnal Temperature Events on Grassland Productivity at Different Elevations on the Plateau
- Understanding Drought through Spatial-Temporal Learning
- Revealing the temporal dynamics of antibiotic anomalies in the infant gut microbiome with neural jump ODEs
- chronODE : A framework to integrate time-series multi-omics data based on ordinary differential equations combined with machine learning
- Predicting microbial community structure and temporal dynamics by using graph neural network models
- A Novel Approach to Differential Expression Analysis of Co-Occurrence Networks for Small-Sampled Microbiome Data
- Comparative Physio-Biochemical and Transcriptome Analyses Reveal Contrasting Responses to Magnesium Imbalances in Leaves of Mulberry (Morus alba L.) Plants
- Transcriptomic landscape profiling of metformin‐treated healthy mice: Implication for potential hypertension risk when prophylactically used
- DroughtSet: Understanding Drought Through Spatial-Temporal Learning
- Robust PCA based method for discovering differentially expressed genes
- LLMs Can Learn the Language of the Microbiome