Every industrial fermentation that makes a recombinant protein is a black box in one specific sense: the measurements used to control it — optical density, dissolved oxygen, bulk metabolite titers — average over millions of cells and hide the very thing that determines whether the run succeeds. A culture can lose productivity not because growth slows but because a minority of cells stops producing: plasmid-loss segregants, secretory-pathway-exhausted cells, or clones drowning in misfolded-protein aggregates. Meanwhile, a mature computational literature exists for exactly this "heterogeneity masking" problem — deconvolving bulk, averaged molecular measurements back into constituent cell states — but it has been developed and validated almost entirely on human tumors and developmental tissues, never on a bioreactor.
This proposal applies and extends bulk transcriptomic deconvolution to recombinant protein production. The core technical obstacle is that bioreactor cultures violate the two assumptions every current deconvolution tool makes: they contain extreme perturbed states absent from any reference atlas (heavy overexpression, proteostatic collapse), and their cells diverge wildly in size and total RNA content across induction phases. The plan is to build an induction-stress reference atlas, then deconvolve routine bulk RNA-seq from fermentation time-courses with a method that corrects for RNA-content scaling and conserves integer counts exactly — and to show that the recovered sub-population trajectories predict fermentation outcomes that optical-density models cannot.
Core Claim
The claim is that the fraction of cells in high-burden, actively producing states versus low-burden "escaped" or dying states can be recovered from cheap, routine bulk RNA-seq of bioreactor samples — provided the deconvolution handles three things no off-the-shelf tool handles together: reference states that include overexpression stress, RNA-content differences between induced and uninduced cells, and the discrete integer geometry of low-count transcripts that is exactly where rare producer sub-populations live. If true, this turns a per-run diagnostic that is currently invisible into a routine measurement costing the price of a bulk RNA-seq library.
Research Program
Three components, each grounded in a limitation documented in the source literature.
First, the reference atlas problem. Deconvolution accuracy collapses when the reference lacks the states present in the sample: POPPY (an ontology-aware contrastive encoder aligning bulk transcriptomes to single-cell foundation-model latents, validated on 330 melanoma bulk samples) and scProjection (cell-type-specific variational autoencoders trained on reference atlases, resolving 87% of cell states on titrated mixtures where CIBERSORTx resolved 61%) both force data onto unperturbed manifolds when perturbed states are absent. Recombinant hosts under strong induction are precisely such perturbed states — TMPRSS2 overexpression in E. coli drives the protease into inclusion bodies unless rescued by the artificial XXA tag (a 16-fold activity gain over His-tag), and the FAS2FURIOUS pipeline exists precisely because difficult secreted targets place extreme stress on host secretory machinery. We will construct a small single-cell reference atlas of a production host (Pichia pastoris or CHO, chosen for biological relevance and droplet compatibility) sampled across induction intensity and time, spanning high-burden producers, low-burden escapees, and dying cells. This is a one-time cost of tens of thousands of dollars, amortized over every subsequent bulk-only run — the asymmetry that motivates the whole approach.
Second, the method. We extend the strongest existing deconvolution machinery along two axes that fermentation data demands. Cell-size and total-RNA correction follows CSsingle, which uses ERCC spike-in controls and iteratively reweighted least squares to correct the systematic bias that arises when cells differ in RNA content — a bias it demonstrated across ten conventional deconvolution tools on HEK/Jurkat titrations. In fermentations this bias is not an edge case: induced cells accumulating inclusion bodies change volume and RNA content dramatically relative to uninduced cells. Integer conservation follows Count Bridges, a generative bridge over integer lattices with Poisson birth–death dynamics that guarantees deconvolved unit-level counts sum exactly to observed bulk counts and outperformed CIBERSORTx and MuSiC in proportion recovery. Rare high-producer sub-clones in a bioreactor are exactly the low-frequency, low-count signal that continuous methods distort. Our contribution is not a new generative architecture but the composition: spike-in-calibrated RNA scaling feeding an integer-conserving bridge, applied to time-series (not single-sample) data with a smoothness prior across fermentation time — sub-population fractions should not jitter run-to-run within one batch.
Third, the coupling to process outcomes. The deconvolved trajectories are linked to yield through a simple predictor: does the inferred producer-fraction trajectory early in induction forecast final titer better than optical density does? The comparison target is the LSTM process model of A Deep Learning Approach to Optimize Recombinant Protein Production in Escherichia coli Fermentations, which maps process parameters (pH, temperature, dissolved oxygen, feed) to OD600 time-series, reporting root mean squared error of 4.00–4.10 and relative error on final yield of 7.21%–7.52% — but whose own authors report breakdown on atypical batches such as batch 12, exactly the runs where clonal dynamics are suspected to shift. Our claim is not that deconvolution replaces the process model, but that the inferred sub-population fractions supply the missing explanatory variable that rescues those atypical batches.
Baselines
The naive reference point is doing nothing beyond current practice: monitor OD600 and standard process parameters, which is what essentially all production runs do and what both the LSTM fermentation model and the biokinetic-prior work of Leveraging Biokinetic Knowledge Priors for Data-Sparse Bioprocess Modeling are built on. That work pre-trains neural networks on synthetic growth curves from classical biokinetic ODEs, lifting $R^2$ from roughly 0.168 for an unregularized MLP to about 0.515 with pre-training and 0.554 with its built-in ODE architecture — the strongest published number for bacterial cultivation prediction on its 11-dataset suite. The do-nothing baseline is what our method must beat on explanatory grounds: on typical batches it will match us on yield prediction; the test is whether it remains blind on the deliberately perturbed batches.
The strongest existing method for the deconvolution half is scProjection, which on the CellBench titrated-mixture benchmark recovered 87% of high-resolution cell states versus 61% for CIBERSORTx. We will run scProjection unmodified on our fermentation bulk samples against our extended method, using the same reference atlas, the same preprocessing, and the same downstream yield-prediction readout, so the only difference is the RNA-content correction and the integer-conserving bridge. Beating it means two things concretely: higher recovery of known spike-in sub-population fractions at low frequencies (below ~10%), and no systematic proportion drift across induction time — the specific failure mode CSsingle demonstrated for size-uncorrected tools. We will also run CIBERSORTx and MuSiC as received baselines, and POPPY as the cross-modal alternative, though it is trained on human clinical data and will likely need re-training, which we will budget for equally.
The key ablation isolates our two additions one at a time: the full method versus the same bridge without ERCC spike-in RNA-content correction, versus continuous relaxation of the integer constraint. Prediction: removing size correction produces proportion estimates that drift systematically with induction intensity (the CSsingle failure mode, transplanted to fermentation), and removing the integer constraint degrades rare-sub-population recovery specifically, while leaving majority-fraction estimates intact — which would show that our two ingredients matter exactly where we claim they matter. A morphological baseline using the Twin-VAE approach (~5.7% mean relative error on CHO-K1 single-cell counts in microfluidic cultivation, versus 29% for Watershed and 25% for EfficientNet) counts cells but, as its authors note, reads out no molecular state; we include it to test whether cell counts alone already predict yield collapse — if they do, our molecular machinery is unnecessary, and we would want to know that early.
Why not just…?
Why not just run single-cell RNA sequencing directly on bioreactor samples? This is the strongest objection, and we do not dismiss it — indeed our reference atlas is single-cell data, taken once. The argument against running it routinely is cost and physics: POPPY notes single-cell sequencing is too expensive and logistically complex for routine workflows, and industrial monitoring samples dozens of timepoints across many runs; more importantly, high-density cultures of hosts with rigid walls (yeast, bacteria) or shear sensitivity handle poorly in droplet microfluidics, and the data-processing burden is exactly what the CHO-K1 microfluidic study flags. The honest empirical settlement is a paired experiment: one run where every timepoint gets both bulk and single-cell sequencing. If direct single-cell measurement at a few timepoints matches our deconvolved fractions within noise, the bulk pipeline is validated as a cheap interpolator; if the two diverge persistently, the deconvolution is biased and the proposal's premise fails.
Why not just use the existing LSTM/ODE process models, which already predict yield well? They predict yield on typical batches — 7.21%–7.52% relative error on final yield for the LSTM model — but their own paper reports breakdown on atypical ones, and the biokinetic-prior work shows cross-species transfer failure (E. coli models failing on Listeria), indicating the macro-level curves miss underlying mechanism. Our decisive check is deliberately adversarial: we will construct batches designed to have identical OD trajectories but divergent sub-population composition (for example, mixed induction timing), where a perfect OD model cannot distinguish them but the deconvolution should. If deconvolved fractions add no predictive value over OD in those runs, the masking problem is empirically unimportant in this regime, and we would report that.
Why not just use CIBERSORTx or MuSiC, since they are standard and free? Because both are documented to fail on exactly our two failure modes: CIBERSORTx resolved 61% of states on CellBench versus 87% for scProjection; CSsingle showed across ten conventional tools that equal-RNA-per-cell assumptions misestimate proportions whenever cell sizes differ, which induced-versus-uninduced cells do; and Count Bridges showed continuous methods distort low-count transcripts and violate exact summation to the bulk, which is where rare producer clones live. We run them anyway as baselines, so the reader can see whether the gap materializes in fermentation data rather than trusting tumor-benchmark numbers.
Why not just use microfluidic imaging to track heterogeneity? Because morphology is not function. The Twin-VAE work counts cells well (~5.7% error) but cannot distinguish a slowly dividing cell that is productively secreting from one dying of inclusion-body toxicity — the distinction that determines whether a run is worth continuing. Imaging also does not scale to thousands of parallel industrial runs the way a bulk library does. We include a morphology baseline, and if counts alone turn out to predict yield collapse as well as deconvolved fractions, we will concede the point and the imaging route wins on cost.
Proposed Experiments
The anchor experiment is a fermentation time-course with known ground truth. We will run a production host expressing a tagged reporter construct at titrated induction levels, sampling bulk RNA-seq with ERCC spike-ins at six timepoints spanning pre-induction to harvest, with end-point single-cell sequencing and fluorescence-activated cell sorting on parallel aliquots providing ground-truth sub-population fractions. Sub-population fractions are then deliberately skewed by mixing pre-induced and uninduced cultures at known ratios spanning 1% to 50%, reproducing the titrated-mixture logic of the CellBench benchmark on which scProjection's 87% and CIBERSORTx's 61% were measured — so the numbers are directly comparable. Primary metrics are sub-population fraction recovery error against the known mixing ratios, integer-conservation residual (deconvolved counts summing to bulk), and drift of estimated fractions across induction phase. Secondary metrics are final-titer prediction error from early-timepoint deconvolved fractions versus from OD alone, on both typical and adversarially matched-OD batches. Success is defined in advance: fraction recovery within 10% absolute at mixtures down to 5%, and a statistically significant improvement in titer prediction on atypical batches over the OD-based LSTM baseline's reported 7.21%–7.52% relative error on final yield. No reported result found on Lacuna for deconvolution applied to any bioreactor transcriptome, so the paired bulk-versus-single-cell run is the decisive check: if the deconvolution reproduces directly measured fractions, the masked heterogeneity of industrial fermentation becomes a routinely measurable quantity, and the gap between the two literatures closes.
References
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Projecting clumped transcriptomes onto single cell atlases to achieve single cell resolution
- Improving Soluble Expression of SARS-CoV-2 Spike Priming Protease TMPRSS2 with an Artificial Fusing Protein
- FAS2FURIOUS: Moderate-Throughput Secreted Expression of Difficult Recombinant Proteins in Drosophila S2 Cells
- CSsingle: A Unified Tool for Robust Decomposition of Bulk and Spatial Transcriptomic Data Across Diverse Single-Cell References
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- A Deep Learning Approach to Optimize Recombinant Protein Production in Escherichia coli Fermentations
- Leveraging Biokinetic Knowledge Priors for Data-Scarce Bioprocess Modeling
- Towards an automatic analysis of CHO-K1 suspension growth in microfluidic single-cell cultivation