Soil carbon models fail most badly in the subsoil, the layers below about 30 cm. Carbon there persists for centuries, yet the models that do resolve depth still represent microbial activity as a single fitted knob per layer. Measuring that activity directly means sequencing microbial RNA (metatranscriptomics) from deep soil cores. This is hard for the same reason single-cell RNA sequencing is hard: there is very little starting material, so many genes that are actually expressed come back as zero counts. Single-cell biologists call these false zeros "technical dropout." In soil the problem has a twist. Microbial biomass genuinely falls with depth, so a zero can mean "this pathway is silent here" just as easily as "we failed to capture it."
This note proposes a depth-aware count model for sparse subsoil metatranscriptomes. The model uses external RNA spike-ins to estimate, for each core layer, how much of the RNA that was present was actually recovered. With that estimate it separates lost signal from true silence, and it draws strength from neighboring depths and from cheap, densely measured soil properties without letting those properties invent signal. The resulting depth profiles of carbon-cycling gene activity, with uncertainty attached, are then used to constrain the depth-dependent microbial activity parameter of a 20-layer soil carbon model. That lets us ask a question the field currently cannot answer: does real microbial activity explain subsoil carbon behavior beyond what clay and mineralogy already explain?
Core Claim
Single-cell denoising methods cannot be transplanted to soil as they are. Their usual step of dividing each sample by its total read count assumes every sample has roughly the same total activity. Down a soil profile that assumption is false, and applying it erases the biomass decline that matters for carbon. We claim that a count model with a per-layer capture factor fixed by spike-ins, a smooth prior over depth, and a likelihood tied to the raw counts can recover deep-layer pathway activity reliably at dropout levels comparable to single-cell data (80–95% zeros, per Challenges and emerging directions in single-cell analysis). We further claim this recovered activity is a better constraint on subsoil carbon models than either taxonomic DNA abundance or soil covariates alone.
A secondary, falsifiable hypothesis follows. Beyond microbial carbon use efficiency reports that microbial carbon use efficiency (the fraction of consumed carbon microbes turn into biomass) correlates weakly with soil organic carbon overall ($R^2 = 0.111$) and not at all below 100 cm. Part of that decoupling may be attenuation from noisy low-signal deep measurements rather than an absence of biological control. Our denoised activity estimates let us test this directly. If the coupling stays flat after denoising, that strengthens the mineral-stabilization account.
Positioning Against Prior Work
The soil side motivates the gap. Microbial Activity and Root Carbon Inputs Are More Important than Soil Carbon Diffusion in Simulating Soil Carbon Profiles builds a 20-layer model in which nonlinear (Michaelis–Menten) microbial kinetics are applied layer by layer, following the MIMICS model. It finds the depth-dependent microbial activity parameter $\alpha_v$ to be among the most sensitive inputs, but it calibrates that parameter only against bulk carbon. It leaves a fourfold unexplained variation in desorption and misses deep radiocarbon ages. We keep that model and replace a free parameter with a measured, uncertainty-bearing prior.
Mapping global distributions, environmental controls, and uncertainties of apparent topsoil and subsoil organic carbon turnover times shows that texture and mineralogy dominate subsoil turnover and that Earth system models get turnover times more than twofold too short. Its random-forest maps assume steady state. We treat its covariates-only approach as the main competitor rather than a foil.
Two papers frame the tension we must resolve. The stratified-nitrogen trial in Stratified Nitrogen Application Enhances Subsoil Carbon Sequestration via Enzyme-Mediated Pathways in Straw-Incorporated Croplands of North China Plain attributes subsoil carbon gains of 5.20–8.17% to enzyme activity. The carbon use efficiency meta-analysis attributes deep carbon to minerals instead. We take no side in advance; the experiments decide.
The single-cell side supplies the method. We borrow three components:
- Negative-binomial count modeling, which handles overdispersed counts without a log transform. A fast and efficient count-based matrix factorization method for detecting cell types from single-cell RNAseq data also argues that explicitly modeling an excess of zeros is often unnecessary, and we test that claim rather than assume it.
- Smooth basis functions over an ordered coordinate, which Melissa: Bayesian clustering and imputation of single-cell methylomes uses along the genome and we use along depth.
- A likelihood tied to the raw counts, used by Clustering by Denoising: Latent plug-and-play diffusion for single-cell embeddings to stop its denoiser from inventing structure.
The dense-informs-sparse idea comes from A deep generative model for multi-view profiling of single-cell RNA-seq and ATAC-seq data. It fixes the number of clusters in advance, and depth is a continuum, so we replace clustering with a smooth prior.
The closest microbiome imputation work, DeepMicroGen: a generative adversarial network-based method for longitudinal microbiome data imputation, imputes taxonomic abundances over time in human gut cohorts. It does not handle RNA, depth, capture efficiency, or carbon.
Research Program
The observation model for read count $y_{g,d}$ of pathway $g$ in layer $d$ of a core is
$$ y_{g,d} \sim \mathrm{NB}\big(s_d \exp(f_g(z_d) + \beta_g^\top x_d),\ \phi_g\big). $$
The symbols are as follows:
- $s_d$ is the capture factor, estimated from spike-in recovery and extracted RNA mass rather than from library size.
- $f_g$ is a smooth function of depth $z_d$, built from basis functions shared hierarchically across cores.
- $x_d$ holds the dense covariates of that layer: clay, silt, pH, moisture, and mineralogy.
- $\phi_g$ is a per-pathway dispersion.
The quantity reported downstream is $\exp(f_g + \beta_g^\top x_d)$, an absolute activity per gram of soil. A true zero is inferred only when the posterior for that activity is concentrated near zero despite adequate capture.
Two design rules guard against circularity. First, soil organic carbon and radiocarbon never enter $x_d$, so later tests against carbon are not contaminated. Second, a second variant drops the covariate term entirely, so we can measure how much of the recovered signal is borrowed from soil properties.
Pathway activities for carbon-degrading enzymes (β-glucosidase, xylosidase) and nitrogen-cycling enzymes (urease, N-acetylglucosaminidase), together with genes linked to calcite precipitation, are then mapped to depth-specific priors on $\alpha_v$ in the 20-layer model.
Baselines
The naive reference is ordinary practice. Reads are normalized by library size, zeros are taken at face value, and the result is averaged per layer. This version should make the subsoil look inactive, and the size of that error is itself a result.
The strongest imputation competitors are the following. Each is run on identical count matrices, with the same downsampling seeds and a matched hyperparameter search of fixed size:
- the negative-binomial factorization of A fast and efficient count-based matrix factorization method for detecting cell types from single-cell RNAseq data;
- the anchored diffusion method of Clustering by Denoising: Latent plug-and-play diffusion for single-cell embeddings, which on PBMC CITE-seq reports ARI 0.805 against 0.641 for scVI and 0.674 for MAGIC;
- MICE, the chained-equations imputer used as a baseline by DeepMicroGen, which reports MAE 1.816 against DeepMicroGen's 1.609 on DIABIMMUNE.
No reported result was found on Lacuna for any method on depth-resolved soil metatranscriptomes, so these single-cell and gut numbers only tell us which methods are worth including. They are not targets to transfer. Beating the competitors means lower error in recovering held-out deep-layer pathway activity in the downsampling test described below, plus calibrated 90% intervals. Calibrated means empirical coverage within five points of nominal.
For the carbon question, the strongest competitor is a covariates-only random forest. In Mapping global distributions, environmental controls, and uncertainties of apparent topsoil and subsoil organic carbon turnover times this approach reaches cross-validated $R^2 = 0.70$ for subsoil turnover. Alongside it we include the original calibrated profile model, which explains 40–94% of profile variance, with $r^2 = 0.63$ for deciduous broadleaf sites. Both are refit on our sites with the same covariates.
Two ablations isolate the proposal's ingredients. In the first, library-size normalization replaces the spike-in capture factor; this tests the central claim. In the second, DNA metagenomic abundance replaces RNA activity in the $\alpha_v$ prior.
Why not just…?
Why not just use DNA metagenomics or 16S sequencing, which yield far more material? DNA records which organisms and genes are present, not what is switched on. Ammonia-oxidizing archaea adapted better to the dark, alkaline oligotrophic karst cave than their bacterial counterparts shows that a single marker gene hid metabolic differences between oligotrophic archaea and bacteria. This alternative is a real threat, because dormant cells may be rare enough that DNA tracks activity well. The DNA-substitution ablation settles it. If DNA predicts measured enzyme activity and carbon profiles as well as RNA does, we recommend DNA and drop the RNA program.
Why not just model soil covariates, since minerals control deep carbon anyway? This is the strongest objection, and the carbon use efficiency meta-analysis supports it. Our test is incremental: we ask whether activity priors improve held-out profile $r^2$ and radiocarbon fit beyond the covariates-only forest and the calibrated model. A null result is informative here. It would indicate that mineral stabilization dominates even after measurement noise is removed, which the current literature cannot conclude.
Why not just run scVI or MAGIC, or sequence deeper? Off-the-shelf tools normalize by library size, and depth is exactly where total activity really changes. Smoothing tools also blur distinct states together, a failure documented in Leveraging data-driven self-consistency for high-fidelity gene expression recovery. Deeper sequencing does not help once extracted RNA mass is the limit, because extra reads mostly duplicate the same few molecules. The capture-factor ablation and the duplicate-rate curve per layer will show whether either of these alternatives is enough.
Proposed Experiments
The first test uses ground truth built from real data. We sequence topsoil and shallow-subsoil cores deeply, then subsample reads and spike-ins to match measured deep-layer RNA yields. This follows the synthetic-dropout logic of Modeling expression ranks for noise-tolerant differential expression analysis of scRNA-seq data, which stresses methods at up to 94% zeros. The deep sequencing then serves as truth for every baseline.
The second test checks whether transcripts track biological function. We core the North China Plain stratified-nitrogen plots and compare inferred pathway activity at 0–20 and 20–40 cm against the enzyme assays that trial already measures. We test whether RNA tracks the reported treatment differences better than DNA does.
The third test asks whether activity improves the carbon model. We sample a pilot of roughly 25 profiles drawn from the 93 validation sites of the profile-model study, and from radiocarbon sites where available. On these we compare held-out carbon profile $r^2$, the explained share of the desorption-rate variance, and the mismatch in deep radiocarbon age. The comparison is between the calibrated model, the model constrained by RNA activity priors, the model constrained by DNA priors, and the covariates-only forest.
Finally, we re-estimate the activity–carbon correlation below 30 cm with denoised versus raw activity. We report how much of the decoupling, if any, measurement noise accounts for.
References
- Challenges and emerging directions in single-cell analysis
- Beyond microbial carbon use efficiency
- Microbial Activity and Root Carbon Inputs Are More Important than Soil Carbon Diffusion in Simulating Soil Carbon Profiles
- Mapping global distributions, environmental controls, and uncertainties of apparent topsoil and subsoil organic carbon turnover times
- Stratified Nitrogen Application Enhances Subsoil Carbon Sequestration via Enzyme-Mediated Pathways in Straw-Incorporated Croplands of North China Plain
- A fast and efficient count-based matrix factorization method for detecting cell types from single-cell RNAseq data
- Melissa: Bayesian clustering and imputation of single-cell methylomes
- Clustering by Denoising: Latent plug-and-play diffusion for single-cell embeddings
- A deep generative model for multi-view profiling of single-cell RNA-seq and ATAC-seq data
- DeepMicroGen: a generative adversarial network-based method for longitudinal microbiome data imputation
- Ammonia-oxidizing archaea adapted better to the dark, alkaline oligotrophic karst cave than their bacterial counterparts
- Leveraging data-driven self-consistency for high-fidelity gene expression recovery
- Modeling expression ranks for noise-tolerant differential expression analysis of scRNA-seq data