When microbiologists heat a fermentation from 37°C to 53°C, the bulk instruments record a dramatic rewrite of the cell's chemistry: in Bacillus subtilis BJ3-2, 1,970 genes, 356 proteins, and more than 130 metabolites change, with ABC transporters pumping in the amino acid precursors for a savory "soy sauce-like" flavor profile (Multi-omics analyses of the mechanism for the formation of soy sauce-like and soybean flavor in Bacillus subtilis BJ3-2). But that measurement averages millions of cells, and the average is ambiguous in a way that matters industrially: it cannot tell us whether every cell turned up the same pathways a little, or whether most cells went dormant while a small, heat-adapted minority did all the work. The same blind spot recurs in the fungal pachymic acid system, where a sharp thermal optimum at 28°C switches on a four-gene Cytochrome P450 cluster and a 7.38-fold jump in product when WcCYP52 is overexpressed (An omics-based characterization of Wolfiporia cocos reveals three CYP450 members involved in the biosynthetic pathway of pachymic acid), and in anaerobic digesters, where temperature decides whether hydrogenotrophic methanogens or acetogens win the competition for H₂ and CO₂ (Bo Fu et al.). Bulk thermal omics — cheap, routine, and already standard in every bioprocess lab — systematically hides the answer, the same averaging artifact that bulk sequencing has long been known to create in tissues (David Adams et al.).
This proposal builds a computational method, which I will call ThermoBridge, that deconvolves existing bulk thermal multi-omics into inferred single-cell-like profiles, using the physics of the temperature shift itself as the constraint that makes the unmixing well-posed. The key observation is that temperature is not just a label on each sample; it is a quantitative, mechanistic parameter. Enzyme capacities follow Arrhenius kinetics, and thermal regulatory programs switch on at known thresholds. Those facts sharply restrict which single-cell metabolic states can plausibly coexist at a given temperature, and that restriction is exactly what current deconvolution tools lack.
Core Claim
Every thermal shift study in the literature interprets bulk differential expression as a uniform, population-wide response. I claim this interpretation is unfalsifiable with current data and current methods, and that a generative deconvolution model whose latent single-cell states are constrained to be metabolically feasible at the measured temperature can recover the hidden population structure — including the fraction of cells in a hyper-active subpopulation versus a graded whole-population shift — directly from bulk data, without needing a single-cell reference atlas collected at that temperature.
The gap is precise. On the deconvolution side, the strongest generative method, Count Bridges, models bulk RNA-seq as a sum of discrete integer count profiles and recovers them with Poisson birth-death bridges and an EM loop (Count Bridges enable Modeling and Deconvolving Transcriptomic Data); but it is purely statistical and cannot tell a biologically feasible single cell from an infeasible one. On the metabolic side, GEFMAP maps single-cell RNA-seq onto reaction fluxes by combining a graph neural network over the metabolic network with a network constrained to the null space of the stoichiometric matrix (Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning); but it assumes single-cell data already exist and treats the metabolic network as static, ignoring the temperature-dependent kinetics that thermal regulation is all about. DestVI shows deconvolution can carry a continuous latent state variable rather than discrete cell types (Multi-resolution deconvolution of spatial transcriptomics data reveals continuous patterns of inflammation), but only in mammalian immunology and only with a complete reference atlas. No existing method combines all three: discrete-count generative unmixing, stoichiometric and Arrhenius feasibility, and temperature as a continuous conditioning variable. That combination is what ThermoBridge is.
Research Program
ThermoBridge has three coupled components. First, a Count-Bridges-style generative core models each latent cell as an integer gene-count vector $x_k$, with the bulk observation constrained to equal $\sum_k x_k$ exactly, preserving the discrete, ordinal character of transcript counts. Second, a metabolic feasibility layer maps each latent cell's enzyme-coding gene counts onto a flux vector $v$ that must satisfy temperature-parameterized mass balance and kinetic limits:
$$S v = 0, \qquad 0 \le v_i \le A_i \, e^{-E_{a,i}/RT}$$
where $S$ is the stoichiometric matrix of the organism's reconstructed network, $A_i$ and $E_{a,i}$ are the pre-exponential factor and activation energy of reaction $i$'s rate-limiting enzyme, $R$ is the gas constant, and $T$ is the measured absolute temperature. Reaction bounds therefore tighten or loosen with temperature exactly as physical kinetics dictates, and thermal regulatory switches (such as the WcCYP cluster's 28°C optimum) enter as prior beliefs on which objective vector $c(T)$ the cell is pursuing, in the spirit of GEFMAP's maximum-weighted-clique objective. Third, a continuous temperature latent variable, following DestVI's design, replaces discrete cell-type bins, so a cell can sit anywhere on the gradient between mesophilic and thermophilic programs. Training runs an EM loop in which the E-step proposes latent single-cell profiles across a temperature series of bulk samples and the M-step updates generative parameters under the feasibility constraint; infeasible proposals are projected out, which is what lets the model invent thermally induced states that no reference atlas contains — precisely where reference-based tools fail through misattribution bias (Bergenstråhle et al., 2020).
The deliverables are the method, an open semi-simulated benchmark for thermal deconvolution, and reanalyses of two published bulk datasets: the B. subtilis BJ3-2 temperature series (37/45/53°C) and the W. cocos series (22.5/28/33.5°C). The headline scientific output is a posterior over population architecture: for each thermal regime, the probability that the observed flux redirection is graded-uniform versus subpopulation-driven, and the estimated fraction and expression profile of that subpopulation. For industry this matters directly — a process whose yield depends on a 10% elite subpopulation is optimized very differently from one where every cell contributes.
Baselines
The do-nothing reference point is the standard practice in the thermal regulation literature itself: interpret bulk differential expression as a population-wide response, as the B. subtilis study does when it attributes the 1,970 DEGs between 37°C and 53°C to activated cysteine/methionine and ABC transporter programs without any population-level qualification ( Multi-omics analyses of the mechanism for the formation of soy sauce-like and soybean flavor in Bacillus subtilis BJ3-2). Any deconvolution method that does not decisively beat this interpretation on held-out data has no reason to exist, so the first check is whether the uniform model is even identifiable against alternatives: on semi-simulated data where the ground truth is a planted subpopulation, the do-nothing baseline should systematically misestimate per-cell pathway activity, and quantifying that error sets the bar.
The strongest existing deconvolution competitor is Count Bridges ( Count Bridges enable Modeling and Deconvolving Transcriptomic Data), which reported state-of-the-art bulk and spatial deconvolution, outperforming CIBERSORTx, MuSiC, and STDeconvolve; no reported result found on Lacuna for its numerical deconvolution accuracy on the MERFISH mouse brain or Onek1k benchmarks, so I will re-run it on my benchmark rather than compare against a published figure. It will be run on identical pseudo-bulk inputs, with an equal tuning budget (same random-search configuration count and compute) and its single-cell reference either present or deliberately absent, mirroring the real thermal setting. Count Bridges is the strongest competitor because it shares the discrete-count generative machinery and therefore isolates my added ingredient: the temperature-conditioned metabolic constraint. Beating it means recovering planted subpopulation fractions and novel thermally induced state profiles with higher Pearson correlation and near-zero misattribution into reference categories, especially when the reference was built at a different temperature. A second strong baseline is SCDC, the best reference-based ensemble deconvolution with a reported number: Pearson $r = 0.991$ against ground truth on an in vitro three-cell-line mixture mixed at a 6:3:1 ratio (SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References); matching that ceiling on in-distribution mixtures while exceeding SCDC in thermally shifted, out-of-reference conditions is the target profile. For the flux layer, GEFMAP is the baseline: it reports Pearson correlation $0.928 \pm 0.016$ on its 892-graph synthetic benchmark and $0.841 \pm 0.014$ on the augmented E. coli benchmark of more than 3,000 graphs with 95 reactions ( Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning); ThermoBridge's flux predictions from deconvolved profiles should reach comparable correlation on the same style of benchmark, demonstrating that unmixing does not destroy metabolic signal. Finally, an ablation removes each key ingredient in turn — the temperature conditioning, the stoichiometric/Arrhenius constraint, and the discrete-count core (replaced with a continuous VAE in the style of inVAE, which disentangles biological from technical variation in single-cell data but assumes bijective continuous decoders (Conditionally Invariant Representation Learning for Disentangling Cellular Heterogeneity)) — to attribute any gain to the specific mechanism rather than to added capacity.
Why not just…?
Why not just run single-cell RNA-seq at each temperature? This is the most serious alternative, and I will not argue it away. Direct single-cell profiling of thermally shifted cultures would settle the uniform-versus-subpopulation question with no modeling assumptions. Three things blunt it in practice. Cost and logistics make it hopeless for the temperature grids, time series, and routine bioreactor monitoring that industrial thermal screening actually uses, which is why the entire thermal regulation literature runs on bulk. Protocol bias is worse here than in mammalian work: SCDC documents how droplet protocols systematically drop large, fragile cells (adipocytes in mammary gland references), skewing every downstream estimate (art_36dc2e3865324d78b4f1954e73b82b5e); thick-walled Gram-positive bacteria and fungal mycelia require harsh enzymatic digestion that preferentially destroys exactly the heat-stressed, fragile states we care about. Concretely, my validation plan includes a small targeted single-cell or flow-cytometry arm on B. subtilis at 37°C and 53°C precisely to test ThermoBridge's inferred architecture, so the alternative is built in as ground truth rather than dodged.
Why not just apply CIBERSORTx or MuSiC with an existing reference atlas? Because thermal shifts violate the closed-world assumption these tools depend on. When the bulk mixture contains a state absent from the reference — here, the heat-induced CYP450 or ABC transporter programs that only exist above certain temperatures — supervised deconvolution cannot discard those transcripts and instead forces them into the nearest reference category (Bergenstråhle et al., 2020; Yang, 2021). A reference built at 22.5°C in W. cocos has no vocabulary for the 28°C CYP state. SCDC itself showed how reference choice flips conclusions: on 77 pancreatic islet bulk samples, standard methods missed the significant HbA1c–beta-cell association ($p = 0.0019$) that SCDC's ensemble recovered (art_36dc2e3865324d78b4f1954e73b82b5e). The cross-thermal extrapolation experiment below makes this failure measurable rather than rhetorical.
Why not just do deeper bulk multi-omics — add proteomics and metabolomics — and skip deconvolution? The B. subtilis BJ3-2 study did exactly this, pairing transcriptomics with nanoLC-MS/MS proteomics and GC×GC-TOF-MS metabolomics ( Multi-omics analyses of the mechanism for the formation of soy sauce-like and soybean flavor in Bacillus subtilis BJ3-2), and it still cannot answer the population question, because all three modalities average over the same cells. The cancer literature gives a calibrated warning about how much this can cost: aggregate bulk biomarkers like tumor mutational burden were significantly outperformed by models that used subclonal architecture for predicting immunotherapy response ($P = 0.02$ on MSK_2020, $P = 6.0 \times 10^{-3}$ on MSK_2018) (A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer). Depth does not buy resolution; more modality channels on the same averaged pool just average more things. ThermoBridge does use the multi-omic structure — metabolite pools constrain feasible fluxes — but as constraints on unmixing, not as a substitute for it.
Why not just run standard Flux Balance Analysis with biomass maximization? FBA on the bulk assumes one representative cell optimizing growth, which is the uniform-response hypothesis baked in by assumption. GEFMAP already showed that biomass optimization obscures the diverse, condition-specific metabolic states cells adopt under environmental change, while transcriptomically weighted objectives recovered them at scale ( Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning). In thermal regimes, growth is often not the active objective at all — secondary metabolite production is — and a single LP solution cannot represent a mixture of dormant and hyper-active cells sharing one bulk metabolome.
Proposed Experiments
The first experiment builds the benchmark. Starting from GEFMAP's augmented E. coli framework (more than 3,000 flux graphs, 95 reactions, 72 metabolites), I will generate synthetic single-cell states whose enzyme-expression profiles and flux objectives vary with a temperature parameter under the Arrhenius-bounded stoichiometric model, then aggregate them into pseudo-bulk mixtures at subpopulation fractions spanning 5% to 95%, mirroring the purity sweep of the breast cancer deconvolution benchmark built from 38,000 simulated mixtures across 26 patients (Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures); that study found nearly all tools mis-predict normal epithelial as cancer epithelial cells at high purity due to shared lineage programs, and I expect the analogous failure in fermentation, where one strain dominates biomass, so this sweep directly stress-tests the lineage-confusion failure mode. Metrics are Pearson correlation of recovered versus true profiles and proportions, Wasserstein-2 distance and Energy Score as used by Count Bridges, and a new misattribution rate for transcripts belonging to thermally induced states with no reference category.
The second experiment is the cross-thermal extrapolation test: train each deconvolution method with a reference built at one temperature (say, the mesophilic end) and evaluate on pseudo-bulk from a hotter regime containing a novel state. The decisive prediction is that reference-based methods (SCDC, Count Bridges in reference mode) show high misattribution while ThermoBridge, which can only propose states that are stoichiometrically and kinetically feasible at the evaluated temperature, invents the novel state. Count Bridges' own training regime — projecting generated counts to sum exactly to the bulk — is retained so the comparison isolates the feasibility constraint and nothing else.
The third experiment is the reanalysis of the two real datasets. On the B. subtilis temperature series, ThermoBridge will report a posterior over uniform versus subpopulation structure for the ABC transporter and sulfur-metabolism programs; on the W. cocos series, it will ask whether the 28°C CYP450 activation is graded or carried by a subpopulation. These produce falsifiable predictions: if the model infers a bimodal architecture at 53°C, a targeted flow-cytometry or reporter-strain validation in B. subtilis (cheap, fast, ground truth) should confirm it. A negative result — the posterior collapsing onto the uniform model — would itself be the first quantified evidence that thermal rewiring in these systems is genuinely population-wide, converting an assumption in the literature into a measurement.
References
- Multi-omics analyses of the mechanism for the formation of soy sauce-like and soybean flavor in Bacillus subtilis BJ3-2
- An omics-based characterization of Wolfiporia cocos reveals three CYP450 members involved in the biosynthetic pathway of pachymic acid
- Competition Between Chemolithotrophic Acetogenesis and Hydrogenotrophic Methanogenesis for Exogenous H2/CO2 in Anaerobically Digested Sludge: Impact of Temperature
- As we come to the end of 2011, several members of the Genome Biology Editorial Board give their views on the state of play in genomics
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning
- Multi-resolution deconvolution of spatial transcriptomics data reveals continuous patterns of inflammation
- Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography
- Multi-omics analyses of the mechanism for the formation of soy sauce-like and soybean flavor in Bacillus subtilis BJ3-2
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
- Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning
- Conditionally Invariant Representation Learning for Disentangling Cellular Heterogeneity
- art_36dc2e3865324d78b4f1954e73b82b5e
- AdRoit is an accurate and robust method to infer complex transcriptome composition
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures