In single-cell RNA sequencing, most genes are observed as zero in most cells. Some of those zeros are biological: the gene's promoter switches on and off in bursts, so many cells genuinely contain no transcript. Others are technical: the molecule was present but was not captured. Current tools such as AutoZI decide, gene by gene, whether the zeros exceed what an ordinary count model predicts. When a gene averages less than one count per cell, a single gene carries too little information to make that call, and the tools lose power. We propose to stop deciding one gene at a time. Genes that are co-regulated tend to share bursting behavior, so we couple their decisions in a network model in which "linked genes prefer the same label." For this kind of model, a family of results called correlation inequalities guarantees that a standard message-passing approximation never overestimates the model's total probability mass. That guarantee is what makes the pooled model trustworthy enough to tune and compare.
The second ingredient comes from machine translation. There, a single wrong word-alignment link can block the extraction of many correct phrases, so an aligner with a better intrinsic score can produce worse translations. Our gene network is also an estimated set of links feeding a combinatorial downstream step, and we expect the same failure: a spurious edge into a hub gene can propagate a wrong label across a module. We therefore build the network to favor precision over recall, and we evaluate the method on downstream calls rather than on how tight its bounds are.
Core Claim
Let $z_g \in \{0,1\}$ indicate whether gene $g$ is biologically zero-inflated. Two things feed the model. The first is a per-gene evidence term $h_g$, the log Bayes factor from an AutoZI-style fit of a zero-inflated negative binomial (ZINB) against a plain negative binomial (NB). The second is a gene graph $E$ with nonnegative edge weights $w_{gg'}$, estimated from independent datasets. The posterior over labels is then
$$ p(z \mid \text{data}) \;\propto\; \exp\Big(\sum_g (a_g + h_g) z_g + \lambda \sum_{(g,g')\in E} w_{gg'}\, \mathbb{1}[z_g = z_{g'}]\Big), $$
where $a_g$ is a prior log-odds term and $\lambda \ge 0$ sets the coupling strength.
Because every coupling rewards agreement, the model is attractive. In the language of the theory, this means the model is log-supermodular: the joint probability of the "meet" and "join" of any two labelings is at least the product of their probabilities. Ruozzi's result in The Bethe Partition Function of Log-supermodular Graphical Models then applies. The Bethe approximation, the quantity that loopy belief propagation optimizes, is a guaranteed lower bound on the posterior normalizer $Z_{\text{post}}(\lambda)$.
The prior normalizer $Z_{\text{prior}}(\lambda)$ is the same model with $h_g = 0$. Tree-reweighted belief propagation (TRW) gives a convex upper bound on it. Dividing the first bound by the second yields a certified lower bound on the model evidence $Z_{\text{post}}(\lambda)/Z_{\text{prior}}(\lambda)$:
$$ \log p(\text{data} \mid \lambda, E) \;\ge\; \log Z_B^{\text{post}}(\lambda) - \log Z_{\text{TRW}}^{\text{prior}}(\lambda) + \text{const}. $$
We use this bound to choose $\lambda$ and to compare candidate graphs. Clamping, which means fixing a few hub genes to each label and summing the resulting sub-problems, tightens the Bethe side and the TRW side (Clamping Variables and Approximate Inference, Clamping Improves TRW and Mean Field Approximations).
The claim to be tested is the following. In data downsampled below one count per cell, gene-level labels from this pooled model should recover full-depth labels better than independent AutoZI does. This advantage should appear only when the graph is precision-oriented, and it should vanish on degree-matched random graphs.
Relation to Prior Work
Two single-cell tools motivate the gap. AutoZI (Detecting Zero-Inflated Genes in Single-Cell Transcriptomics Data) places a spike-and-slab prior over NB versus ZINB for each gene. It showed that zero-inflation calls shrink as 10x chemistry improves and that the calls which remain track bursty promoters in mouse embryonic stem cells. It treats genes independently and loses power in the sparse limit. NEBULA (NEBULA is a fast negative binomial mixed model for differential or co-expression analysis of large-scale multi-subject single-cell data) makes multi-subject negative binomial models fast using large-sample approximations. It reports the same sparse-regime weakness and underestimates variance when there are fewer than 30 subjects. We reuse AutoZI's per-gene evidence as the unary term and change only how decisions are shared.
Four lines of work supply the method. The Bethe lower bound for attractive models was conjectured and supported through loop series by Loop Series and Bethe Variational Bounds in Attractive Graphical Models and proved in general by Ruozzi. Clamping was introduced and analyzed by Weller and Jebara. None of these papers applied the bounds to learned biological networks, and none paired a Bethe lower bound with a TRW upper bound to certify a normalizer ratio. The continuous extension, A Lower Bound on the Partition Function of Attractive Graphical Models in the Continuous Case, suggests a later version in which the latent variable is burst frequency rather than a binary label. We defer that version because continuous belief propagation is not generally tractable.
One negative result constrains the design. Weller and Jebara show that in frustrated models, clamping can make the Bethe approximation worse. Frustration arises in models with repulsive couplings arranged so that no labeling satisfies every link. We therefore keep only nonnegative couplings. Repressive regulatory edges are dropped rather than encoded, and we accept the loss of information this causes.
Machine translation supplies the evaluation lens. Going Beyond AER: An Extensive Analysis of Word Alignments and Their Impact on MT showed that high-recall alignment unions reach 22.66 BLEU against 25.38 for a precision-oriented supervised aligner, because spurious links block phrase extraction. Discriminative Modeling of Extraction Sets for Machine Translation responded by scoring the downstream extraction structure directly. We adopt both lessons. We build graphs by intersecting co-expression across independent datasets rather than taking their union, and we score the method on downstream label recovery rather than on bound tightness.
SATURN (Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN) is the closest work touching alignment and single-cell noise. It learns cross-species gene groupings with a deterministic autoencoder. It neither models zero-inflation probabilistically nor measures how alignment errors propagate downstream.
Research Program
We will build the pooled labeler on top of scVI and AutoZI. The per-gene fits are unchanged; only the decision step becomes the network model above.
Graphs come from co-expression estimated on data disjoint from the target dataset, so the prior does not reuse its own evidence. One source is a different 10x chemistry or tissue. Another is NEBULA-style co-expression estimated on the PBMC or Alzheimer's cohorts. We construct two versions of each graph. The "union" version keeps an edge if any source supports it. The "intersection" version keeps an edge only if all sources support it.
Inference runs loopy belief propagation for the posterior marginals and Bethe bound, and TRW for the prior bound. Clamping uses the maxW rule from Weller and Jebara, which fixes the gene with the largest total incident edge weight, and we clamp up to $k$ genes. The coupling $\lambda$ and the choice of graph are selected by maximizing the certified evidence bound.
We will also analyze a question the theory leaves open. How large can a spurious edge's weight be, relative to a gene's own evidence $h_g$, before it flips the label of a well-supported neighbor? This is the network analogue of a blocker link. We will characterize it empirically on real graphs and derive a simple sufficient condition on trees.
Baselines
The naive reference fits a plain NB to every gene, or flags as zero-inflated every gene whose observed zero fraction exceeds the NB prediction by a fixed margin. It anchors how much any modeling buys in the sparse regime.
The strongest existing method is independent AutoZI run through its released scVI implementation, with the same latent dimension, training epochs, and posterior threshold of 0.5 on the same count matrices. AutoZI reports an AUC of 0.9907 on its synthetic data and over 98% of ERCC spike-ins correctly called not zero-inflated (Detecting Zero-Inflated Genes in Single-Cell Transcriptomics Data). No reported result was found on Lacuna for AutoZI accuracy below one count per cell, so we will measure that ourselves. The pooled model beats AutoZI if it improves recovery of full-depth labels at matched specificity on spike-ins, averaged over downsampling seeds, with a confidence interval excluding zero. Holding specificity fixed ensures it cannot win by calling more genes zero-inflated.
A module-aggregation baseline sums counts within each graph community and runs AutoZI on the pooled counts, broadcasting each module's label to its genes. It uses the same graphs, so any difference reflects gene-level versus module-level decisions.
We run mean field and TRW marginals on the same network model with the same unaries, graph, and $\lambda$ search. This isolates the choice of approximation.
Three ablations isolate the key ingredients. A degree-matched random graph tests whether biological edges matter. Union versus intersection graphs tests the precision-over-recall hypothesis. Clamping $k \in \{0,1,5,20\}$ genes tests whether a tighter bound changes decisions at all. All methods receive the same compute budget per dataset and the same hyperparameter search size.
Why not just…?
Why not just aggregate co-expressed genes and call zero-inflation on modules? This is the most serious threat, because aggregation also pushes counts above one per cell. It fails when a module mixes bursty and steadily expressed genes. It then assigns one label to all of them, whereas the network model lets strong individual evidence override the neighbors. We will simulate modules with controlled mixing fractions. If aggregation matches the pooled model even on mixed modules, the network-model framing is unnecessary and we will report that.
Why not just use mean field, which already gives a lower bound? On a 30×30 torus, the mean field bound error fell to -70 as coupling strength grew, while the Bethe gap stayed around 0.69 (Loop Series and Bethe Variational Bounds in Attractive Graphical Models). A bound that loose would make evidence-based selection of $\lambda$ and of the graph unreliable. The mean-field arm tests whether this translates into worse labels on dense gene modules, or only into worse bounds.
Why not just smooth counts across similar cells before calling? Nearest-neighbor smoothing adds cell-to-cell correlation that mimics bursting, and it offers no way to tell whether a call is supported. We will include smoothed-then-AutoZI as an extra arm and check its ERCC specificity. Spike-ins have no biological zero-inflation, so any rise in calls there exposes the artifact.
Why not just fit the coupling strength by cross-validation instead of certified bounds? At sparse depth, held-out likelihood is noisy, echoing the warning in Improved Statistical Alignment Models that end-task-only evaluation is noisy. We will compare the $\lambda$ chosen by the bound against the $\lambda$ chosen by cross-validation, in terms of downstream recovery. If cross-validation is as good, the bound's value reduces to interpretability, and we will say so.
Proposed Experiments
The primary test uses 10x v3 data downsampled to 0.1–1 mean counts per cell. Full-depth AutoZI labels are the reference, and ERCC spike-ins serve as negative controls. The metrics are recovery AUC and precision at fixed spike-in specificity.
A cross-chemistry test treats genes flagged in v2 but not in v3 as probably technical. It asks whether the pooled model called on v2 suppresses these genes.
In mouse embryonic stem cells, we test whether the pooled calls track bursty promoter kinetics in sparse subsamples. The AutoZI synthetic generator is extended with a planted attractive graph, planted spurious edges, and mixed modules. Here we measure the flip threshold directly and compare union with intersection graphs.
A final downstream test asks whether using the pooled labels to choose NB versus ZINB likelihoods changes differential-expression power and false discovery cost. We measure power, true discovery rate, and false discovery cost as defined by cypress: an R/Bioconductor package for cell-type-specific differential expression analysis power assessment, on simulated multi-subject data with NEBULA as the testing engine.
The decisive check has three parts that must hold together. The pooled model must beat independent AutoZI below one count per cell. The gain must disappear on degree-matched random graphs. Intersection graphs must beat union graphs. If a tighter clamped bound improves the certified evidence but never changes labels or downstream power, that is itself the metric disconnect, and we will report it as a finding.
References
- The Bethe Partition Function of Log-supermodular Graphical Models
- Clamping Variables and Approximate Inference
- Clamping Improves TRW and Mean Field Approximations
- Detecting Zero-Inflated Genes in Single-Cell Transcriptomics Data
- NEBULA is a fast negative binomial mixed model for differential or co-expression analysis of large-scale multi-subject single-cell data
- Loop Series and Bethe Variational Bounds in Attractive Graphical Models
- A Lower Bound on the Partition Function of Attractive Graphical Models in the Continuous Case
- Going Beyond AER: An Extensive Analysis of Word Alignments and Their Impact on MT
- Discriminative Modeling of Extraction Sets for Machine Translation
- Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN
- Improved Statistical Alignment Models
- cypress: an R/Bioconductor package for cell-type-specific differential expression analysis power assessment