When stem cells are grown in large bioreactors for therapy, the cells talk to each other by releasing signaling proteins, called ligands, into the culture fluid. Other cells pick these up through matching receptors. This chatter can slowly push the culture toward unwanted states, such as aging-like inflammatory secretion or loss of the most therapeutic cells. Existing tools that infer this chatter from single-cell RNA sequencing (measuring which genes each individual cell is expressing) assume a tissue. In that picture, a ligand made by one cell type reaches a nearby receiver, and signal strength is roughly ligand expression times receptor expression for each sender–receiver pair. A stirred, perfused bioreactor behaves differently. The fluid is well mixed, so every secreted ligand joins a single shared pool. Fresh medium constantly flushes that pool out. Which cell sent a signal matters only through how much it adds to the pool.
We propose to model a culture as cells coupled through one shared, flushed medium compartment. Each cell's genes evolve by learnable differential equations. Each soluble ligand's concentration rises with secretion summed over the population and falls with perfusion washout. The flushing rate is an experimental knob, and that makes the model testable. If a model has correctly identified which ligand drives a drift, it should predict how faster or slower flushing, blocking antibodies, or transferred conditioned medium change the culture's cell-state composition, including at flushing rates it never saw.
Core Claim
The claim is that in well-mixed bioreactors, cell–cell communication is best represented as a population-to-pool-to-cell process, not as pairwise sender–receiver links, and that perfusion rate is a cheap, controllable intervention that makes the inferred communication axes falsifiable.
This differs from the closest work in specific ways. DIISCO learns time-varying interaction matrices between cell types from time-series single-cell data. It draws its signal from shifts in cell-type proportions, and it has no physical medium or washout term. In a single-line expansion where nominal composition barely moves, that signal is weak. FLeCS couples gene networks to learnable differential equations at transcriptome scale, but it models cells as independent systems with no extracellular channel. CellBRIDGE uses ligand–receptor potential only as a prior for aligning snapshots, not as a mechanism with dynamics. CellChat computes static mass-action pairwise scores. On the biomanufacturing side, Chen et al. showed that 3D automated microcarrier culture improves umbilical cord stem cell quality, but used bulk RNA sequencing. That approach cannot tell which subpopulation drives the effect.
The Model
Let $x_i(t)$ be the expression state of cell $i$ and $L_k(t)$ the medium concentration of soluble ligand $k$. Cells respond to the pool according to their own receptor expression $r_i$:
$$\frac{dx_i}{dt} = f_\theta\big(x_i,\; r_i \odot h(L(t))\big) - \gamma \odot x_i ,$$
Here $f_\theta$ is a sparse gene-network production function in the style of FLeCS, $h$ is a saturating receptor-occupancy function, and $\gamma$ holds per-gene decay rates. The pool is governed by
$$\frac{dL_k}{dt} = N(t)\,\mathbb{E}_i\big[\alpha_k\, s_k(x_i)\big] - (D + \delta_k)\,L_k - \beta_k\,\mathbb{E}_i[r_{ik}]\,L_k ,$$
In this equation, $N(t)$ is the cell count, $s_k$ maps a cell's ligand transcripts to secretion, $\alpha_k$ is a learned secretion scale, $D$ is the known perfusion dilution rate, $\delta_k$ is ligand degradation, and $\beta_k$ is receptor uptake.
Membrane-bound and contact-dependent ligands are not part of the pool. They enter through a separate pairwise term scaled by microcarrier occupancy density. Their number of free parameters is limited so that they cannot absorb the soluble signal.
Two features address the known weaknesses of mRNA-only inference. First, $D$ is set by the operator, so the washout term is known rather than fit. Second, a small multiplexed panel of supernatant protein measurements at each sampling point observes $L_k$ directly for a subset of ligands. These measurements anchor the scale of $\alpha_k$, which mRNA alone cannot identify.
Under constant perfusion, the pool reaches a quasi-steady state at roughly $L_k \approx N\bar{s}_k\alpha_k/(D+\delta_k+\beta_k\bar r_k)$, where $\bar{s}_k$ and $\bar r_k$ are population-average secretion and receptor expression. This gives a concrete prediction. Autocrine-driven state shifts should weaken in proportion to $1/D$ once $D$ dominates degradation and uptake. Shifts driven by cell-intrinsic programs should not depend on $D$.
Research Program
The work has three parts. The first is a simulator: a medium-compartment extension of a gene-network simulator in the spirit of SERGIO. It plants known autocrine axes, cell-intrinsic drift programs, and decoy co-expressed ligand–receptor pairs that have no causal effect, at several perfusion rates. This gives ground truth for axis recovery.
The second is a wet-lab dataset. Human umbilical cord stem cells will be expanded in an enclosed microcarrier bioreactor of the kind used by Chen et al. under one closed-batch and four perfusion settings. Each run gets six sampling points across expansion. Samples are pooled using cell-hashing to control cost, and every sampling point also gets a supernatant cytokine panel.
The third is an analysis of subpopulation fate. The readouts are the pro-vascular state that Duscher et al. showed vanishes invisibly to bulk assays, marked by Hif1a, Vegfa, and Cxcl12, and a senescence-associated inflammatory state. Duscher et al. found that aged populations indistinguishable by flow cytometry differed sharply in tube formation (3.1 versus 11.4 tubules per high-power field). This motivates using a functional tube-formation assay alongside transcriptomic fractions. Duscher et al. motivate the target; Chen et al. supply the culture platform; FLeCS supplies the differential-equation machinery we reuse; CellChatDB supplies the multi-subunit ligand–receptor dictionary. CellChat reports that modeling multi-subunit complexes cuts false positives by 10–14%, which constrains our design toward complex-aware ligand definitions.
Baselines
The naive reference is a persistence model. It predicts that the held-out condition's cell-state distribution at each time equals the pooled training distribution at the same time, ignoring perfusion. Any method that fails to beat this has learned nothing condition-specific.
The strongest static communication baseline is CellChat run at each time point. Its interaction strengths are regressed against perfusion rate, and the axes it ranks highest are taken as candidate drivers. It is the most widely used complex-aware method and represents current practice. It will see the same expression matrices and the same database. It will be beaten if our model's top-ranked axes better predict the outcome of blocking-antibody experiments and, in simulation, achieve higher axis-recovery area under the ROC and precision–recall curves.
The strongest dynamic baseline is DIISCO, given the same time points and the same database prior (OmniPath). It is the only cited method that learns time-varying interactions from single-cell time series. Its published evaluation reports qualitative superiority over CellChat and CellPhoneDB in temporal correlation, but no numerical coefficient. No reported result was found on Lacuna for DIISCO on any bioreactor or single-line culture.
The key ablation is FLeCS given perfusion rate as an exogenous input feature, with no ligand compartment. It is the same differential-equation backbone at matched parameter count and with a matched hyperparameter search budget. This isolates whether the mechanistic pool adds anything beyond conditioning on $D$. A second ablation removes the protein anchoring term, to measure how much direct observation of $L_k$ matters.
For reference points on synthetic recovery, SigXTalk reports an 18% increase in area under the ROC curve and 15% in area under the precision–recall curve over 12 gene-network methods on SERGIO simulations. miRTalk reports 0.93 and 0.98 on its own simulator. These are not directly comparable to our task, and we will re-run SigXTalk-style receptor-to-target scoring on our simulator rather than quote them.
For distribution prediction, the metric is Wasserstein distance between predicted and observed held-out cell-state distributions, the same metric CellBRIDGE used. No numerical value was found on Lacuna for CellBRIDGE. Beating a baseline means a lower Wasserstein distance at the held-out perfusion rate across at least three of four leave-one-rate-out folds, with bootstrap intervals over cells that do not overlap.
Why not just…?
Why not just run CellChat at each time point? Pairwise scores answer "who could signal to whom" under a tissue geometry that a mixed reactor does not have. They contain no washout, so they cannot predict what a new perfusion rate will do. The antibody-blocking experiment settles whether static rankings identify causal axes. If CellChat's top axis is blocked with the predicted effect as often as ours, our mechanism is not needed for discovery.
Why not just measure cytokines in the supernatant? Supernatant panels give pool concentrations but not which subpopulation secretes or responds. Under perfusion, many factors fall below detection. We use them, but as partial observations inside the model. The no-anchoring ablation and a protein-only regression baseline (supernatant levels predicting state fractions) show how far protein alone gets. If protein-only regression matches our held-out accuracy, the single-cell component is not earning its cost, and we would say so.
Why not just give FLeCS the perfusion rate as an input? This is the real threat. A flexible model conditioned on $D$ may interpolate between trained rates as well as a mechanistic one. Two tests separate them. The first is extrapolation to a rate outside the training range. The second is conditioned-medium transfer at fixed $D$, where the black-box input is unchanged but the pool is not. Only the pool model can represent that intervention. If the ablation matches us on both, the claim fails.
Proposed Experiments
In simulation, we vary planted axis strength, dropout rate, and perfusion range, and score axis recovery against CellChat, DIISCO, and SigXTalk-style scoring. The planted decoys test false-positive control.
In the bioreactor, we use leave-one-perfusion-rate-out prediction of state fractions and full distributions, plus one extrapolation run at twice the highest training rate.
The decisive checks are three prospective interventions chosen by the model before they are run. The first is a blocking antibody against the top predicted autocrine axis driving the inflammatory state. The second is transfer of conditioned medium from a high-density batch run into a low-density perfused run. The third is a run at a new perfusion rate. Each produces a pre-registered prediction of pro-vascular and inflammatory fractions and of tube formation.
The proposal succeeds if predictions fall within bootstrap error of observation for at least two of three interventions while the ablation does not. It fails if the inferred axes are no better than the CellChat rankings at predicting the antibody outcome.
As a portability check, the pool model with $D=0$ will be fit to the public CAR-T/lymphoma time course used by DIISCO. The goal is to confirm it recovers the reported ICAM1–IL2RA trend through the contact term, not the pool. Because that pair is membrane-bound, this also tests whether the separation into pool and contact channels behaves sensibly.
References
- DIISCO: A Bayesian framework for inferring dynamic intercellular interactions from time-series single-cell data
- A scalable gene network model of regulatory dynamics in single cells
- CellBRIDGE: Learning Cellular Trajectories via Interaction-Aware Alignment
- Inference and analysis of cell-cell communication using CellChat
- Automated and Enclosed Three-Dimensional Biofabrication System for Mesenchymal Stem Cell Culture to Enhance Diabetic Wound Healing
- Aging disrupts cell subpopulation dynamics and diminishes the function of mesenchymal stem cells
- Dissecting crosstalk induced by cell-cell communication using single-cell transcriptomic data
- Extracellular vesicle-derived miRNA-mediated cell–cell communication inference for single-cell transcriptomic data with miRTalk