Bulk RNA sequencing of a tumor biopsy pools transcripts from thousands of cells into one vector of gene counts. Computational deconvolution then tries to recover which cell types were present and in what proportions. These methods fail most in the cases clinicians care about. When a biopsy is mostly cancer cells, the dominant tumor signal absorbs evidence that belongs to similar but rarer cell types. A benchmark of nine leading methods on 38,000 simulated breast-cancer mixtures found that as tumor purity rises toward 95%, every method increasingly mistakes normal epithelial cells for malignant ones. The same benchmark found a hard trade-off between false positives and false negatives: the method with the fewest false calls of absent cell types missed many real ones (Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures).
We propose to borrow one mechanism from spiking neuroscience, the dynamic firing threshold, and use it as an inference rule rather than as a neural network. A dynamic firing threshold is a neuron's firing level that jumps up each time the neuron fires and then decays back. The neuron becomes temporarily less excitable after heavy use, but it keeps accumulating input while it is suppressed. In our version, each reference cell type is a "unit" that repeatedly claims small integer chunks of the bulk counts it can explain. A unit that has just claimed a lot becomes temporarily reluctant to claim more. This gives units for minority cell types, which have been accumulating evidence on genes specific to them, a chance to claim it before the dominant type absorbs the remainder. The question is whether this history-dependent self-suppression gives a better false-positive versus false-negative trade-off at high purity than simpler ways of discouraging a dominant component.
Core Claim
We claim that adaptive-threshold explain-away inference should give a strictly better operating curve at high purity, measured as missed minority cell types against falsely called ones, than both the same inference with fixed thresholds and convex regression with a tuned penalty on component size. We expect this only when the reference profiles are misspecified, meaning the patient's tumor cells differ from the reference tumor cells.
The scope is deliberately narrow. If the reference is exactly right, the problem is a convex fit with a unique answer, and no ordering trick can improve on it. The masking problem arises because the patient's tumor deviates from the reference tumor. That deviation leaves leftover signal, and greedy or least-squares fits push it into whichever component already dominates or into a correlated neighbor such as normal epithelium. We expect the mechanism to matter in exactly this regime.
The concrete inference dynamics, run over iteration steps $t$ for each cell type $k$, are:
$$V_k(t+1) = \lambda V_k(t) + g_k(t) - \Theta_k(t) S_k(t), \qquad \Theta_k(t+1) = \Theta_0 + (\Theta_k(t) - \Theta_0)e^{-1/\tau} + \eta S_k(t).$$
The symbols are as follows. $V_k$ is the unit's accumulated evidence. $\lambda \le 1$ is a leak. $g_k(t) = \sum_g x_{kg}(y_g/\mu_g(t) - 1)$ is the Poisson-likelihood gradient for cell type $k$, computed from its reference profile $x_{kg}$, the observed bulk count $y_g$ of gene $g$, and the current reconstruction $\mu_g(t)$. $S_k(t) \in \{0,1\}$ records whether the unit fires, which happens when $V_k > \Theta_k$. $\Theta_0$ is the baseline threshold, $\tau$ is how long suppression lasts, and $\eta$ is how much it jumps per firing.
When a unit fires, a fixed integer quantum of reads is assigned to cell type $k$, distributed across genes in proportion to the profile $x_k$ and capped by the unclaimed counts. The subtraction reset keeps leftover evidence instead of zeroing it. Inference stops when all reads are claimed or no unit fires within a window. Two properties follow from this design. First, the bulk counts are conserved exactly, the same constraint that Count Bridges enable Modeling and Deconvolving Transcriptomic Data enforces with projection-guided sampling. Second, there is nothing to train: one deterministic pass per sample, with three hyperparameters.
Relation to Prior Work
Three lines of work motivate the design. The breast-cancer benchmark above documents the purity trap and the false-positive versus false-negative tension, and it supplies our main test bed. Count Bridges enable Modeling and Deconvolving Transcriptomic Data builds discrete Poisson birth-death generative bridges, uses expectation-maximization to deconvolve bulk counts into integer single-cell profiles, and reports beating CIBERSORTx and MuSiC. We share its view that counts should be treated as integers, but we perform a deterministic claiming process rather than stochastic sampling, and we focus on the high-purity failure mode, which that paper does not isolate. Cross-modal transfer learning for mapping bulk transcriptomes at cellular level (POPPY) aligns bulk and single-cell embeddings with ontology-aware contrastive learning and cross-attention. It addresses a different question, patient-level representation and attribution, and gives no mechanism for stopping one dominant population from absorbing attention.
The method itself comes from the spiking literature. Long short-term memory and Learning-to-learn in networks of spiking neurons showed that thresholds which rise after each spike and decay slowly act as a memory trace. Incorporating the Refractory Period into Spiking Neural Networks through Spike-Triggered Threshold Dynamics (RPLIF) showed that raising the threshold after a spike while integration continues suppresses over-excitation without discarding evidence. RPLIF reports nearly 20% higher noise robustness and nearly 30% lower firing rates than standard spiking neurons with fixed thresholds. We take two elements directly from these papers: the subtraction reset and continued integration from RPLIF, and the decaying jump from the LSNN work. However, we use them inside an inference loop, not a trained network.
Cellular-Resolution Population Imaging Reveals Robust Sparse Coding in the Drosophila Mushroom Body found that a population of olfactory neurons keeps roughly 5–6% of its cells active regardless of odor concentration, through adaptation and sublinear mixture suppression. We use this only as a design target, the analog of an output that does not change with purity, not as evidence that our method works. We found no prior work applying threshold adaptation to deconvolution. The closest relative conceptually is greedy pursuit, which repeatedly picks the reference profile that best explains the remaining signal and subtracts it. Our method is greedy pursuit with history-dependent refractoriness, and the ablations are designed to test exactly that difference.
Research Program
The work proceeds in three steps.
First, we implement the inference loop on raw counts. We use per-type reference profiles built from single-cell atlases, a hierarchical variant that first allocates broad lineages and then splits them into subtypes, and an optional per-sample re-estimation of the tumor profile from the reads assigned to it, run between passes.
Second, we characterize when adaptation matters. We vary two kinds of reference misspecification on controlled synthetic mixtures and trace each method's full curve of minority-type recall against false calls of absent types as $(\eta, \tau)$ or the competing method's penalty varies. Reference profiles are built from held-out patients, and tumor profiles are perturbed along realistic patient-to-patient axes. Our prediction is that adaptation leaves accuracy unchanged when the reference is correct and improves it more as misspecification grows. If the improvement does not grow with misspecification, the mechanism is not doing what we think.
Third, we test transfer to real profiles and analyze the failure mode. Here we check whether refractoriness causes minority units to absorb leftover tumor signal, which would produce new false positives.
Baselines
The naive reference is non-negative least squares, a linear fit of the bulk vector to the reference profiles with proportions constrained to be non-negative, together with MuSiC. Both are run on the same reference and the same genes. These establish how much any method gains over plain unmixing.
The strongest existing competitor is BayesPrism. The breast-cancer benchmark identifies BayesPrism and DWLS as having the lowest combined false-positive and false-negative rates and the lowest error on fine immune subsets, and BayesPrism as among the most stable across purities (Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures). BayesPrism also already addresses misspecification by re-estimating the tumor profile for each sample, which makes it the most dangerous comparison. The file gives no numeric value for its error or false-positive and false-negative rates, so we report no reported result found on Lacuna for BayesPrism's purity-stratified error. We will rerun it ourselves with default settings on identical mixtures and references. Beating it means lower minority-type false-negative rates at matched false-positive rates in the 70–95% purity bins, without worse error at low purity. DWLS is run alongside it. Count Bridges is included if code is available, at matched wall-clock time.
Two ablations isolate the key ingredient, and both receive the same hyperparameter search budget as the full method. The first is identical inference with fixed thresholds ($\eta = 0$), which is quantized greedy pursuit. The second is the same pursuit with a static penalty on each type's total claimed fraction, tuned on the same validation split. This second ablation reproduces the average desensitizing effect of adaptation without its temporal history.
Why not just…?
Why not just use BayesPrism, which already adapts the tumor profile per sample? This is a real threat. If per-sample profile re-estimation removes most of the misspecification, little leftover tumor signal remains to be misallocated, and refractoriness has nothing to correct. We settle this by also running our method with BayesPrism-style profile re-estimation and comparing the two at matched settings. If BayesPrism's curve dominates ours in the high-purity bins, we report that the mechanism adds nothing beyond profile adaptation.
Why not just penalize large components in a convex fit? If adaptation only discourages dominance on average, a tuned penalty or entropy term does the same with a guarantee of reaching the optimum. The static-penalty ablation tests exactly this. Our claim survives only if history dependence helps. For example, a dominant type that pauses and then resumes should allocate differently from one that is uniformly shrunk, because minority units fire on specific evidence while the dominant unit is suppressed. If the penalized fit matches us, we will say so and recast the contribution as a cheap count-conserving regularizer.
Why not just train a neural network such as Scaden or an LSTM on simulated mixtures? Learned models can, in principle, learn to resist the purity trap. The benchmark nonetheless lists Scaden among the methods that degrade with purity. A trained model also inherits whatever purity distribution it was simulated on. Our inference is training-free. We include Scaden on the same mixtures and test both methods on purity ranges held out from Scaden's training simulation.
Why not just use a standard spiking network with fixed thresholds? A fixed threshold removes the desensitization we are testing. Fixed-threshold spiking networks reach 63.3% on Sequential MNIST against 96.4% for adaptive-threshold networks (Long short-term memory and Learning-to-learn in networks of spiking neurons). That evidence comes from a different task, however, so the $\eta = 0$ ablation, not that figure, decides the question here.
Proposed Experiments
The primary benchmark reconstructs the 38,000 pseudo-bulk mixtures from the Wu et al. breast cancer atlas (100,064 cells, 26 patients) at purities from 5% to 95%, following Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures. We add one change: references are built from patients held out of the mixtures, so misspecification is realistic. We also plant rare types at 0.5–2% abundance.
Metrics are the following: per-type error; false-positive and false-negative rates for detecting a type's presence, the benchmark's own measure; the normal-versus-malignant epithelial confusion rate; and, for each method, the slope of error against purity. A flat slope is the purity-invariance target suggested by the mushroom-body result. External checks use the Bassez and Pal single-cell cohorts to build fresh pseudo-bulks, and the Kassandra sorted-population simulated profiles used in Cross-modal transfer learning for mapping bulk transcriptomes at cellular level. On the 1,038 TCGA bulk profiles, where there is no ground truth, we only check that estimated proportions agree with other methods.
Four results decide the question. First, the full method's curve of minority recall against false calls must dominate both ablations in the 70–95% purity bins, with confidence intervals obtained by resampling patients. Second, that advantage must grow with injected misspecification and vanish when the reference is exact. Third, the method must not lose to BayesPrism with matched profile re-estimation. Fourth, it must not increase false positives for planted absent types. If the first or second result fails, the mechanism has no value for deconvolution, and we will publish it as a negative result.
References
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Long short-term memory and Learning-to-learn in networks of spiking neurons
- Incorporating the Refractory Period into Spiking Neural Networks through Spike-Triggered Threshold Dynamics
- Cellular-Resolution Population Imaging Reveals Robust Sparse Coding in the Drosophila Mushroom Body