Inside a living tissue, cells sit packed together and constantly secrete signaling proteins. Sometimes a cell releases a molecule that binds receptors on that same cell — autocrine signaling — and when the secreted ligand and the transcription factors it activates begin to reinforce each other's production, the pair forms a biological switch. Once enough ligand accumulates locally, a subpopulation of cells flips into a self-sustaining "ON" state that no longer needs the original trigger. This kind of circuit underlies scar-forming fibrosis in the kidney, cancer stemness in head and neck tumors, and acquired drug resistance. Two things keep us from studying it properly. First, a culture dish is a physically different place from a tissue: the large fluid volume and frequent medium changes dilute secreted ligands below the level needed to trip the switch, so the very state we want to study is chemically suppressed in most laboratory screens. Second, standard assays grind up millions of cells and measure the average, which cannot tell us whether the circuit is weakly active everywhere or fully locked in a small, dangerous minority — nor when that minority committed.
The proposal here is to build and test a generative model that treats the autocrine loop as what it is: a two-state switch embedded in an explicit physical environment. The model takes longitudinal bulk or single-cell expression data and infers, per cell population and per time point, the fraction of cells in the loop-ON state and the time at which they crossed the point of no return. The same model carries a ligand mass-balance term that converts secretion rates into local ligand concentrations under either culture-dish or tissue conditions, which is what lets it translate in vitro measurements into in vivo predictions. The motivating biology is solid: the LIF–LIFR–STAT3–SOX2 autocrine loop that sustains cancer stemness in head and neck squamous cell carcinoma is mapped end-to-end, from super-enhancer to secreted ligand, in Super‐Enhancer Driven LIF/LIFR‐STAT3‐SOX2 Regulatory Feedback Loop Promotes Cancer Stemness in Head and Neck Squamous Cell Carcinoma, and the EP4–prostaglandin E2 loop that makes kidney scarring self-sustaining after the original TGF-β stimulus is gone is established in A Maladaptive Role for EP4 Receptors in Mouse Mesangial Cells. Both studies validate the loop in vivo but measure it only with bulk assays and static endpoints, leaving the single-cell kinetics of loop locking — the thing this proposal targets — explicitly open.
Core Claim
A model with three ingredients — a bistable (two-attractor) state variable per cell, an explicit ligand mass-balance that accounts for how fast the environment removes secreted ligand, and temporal coupling across time points — can do two things no current method can: recover the commitment time of a rare loop-locked subpopulation from longitudinal averaged data, and predict which in vitro observations will fail to translate in vivo because the loop that drives the in vivo phenotype never had a chance to turn on in the dish.
The physics is compact enough to write down. Let $L$ be the local concentration of the autocrine ligand, $\sigma$ the secretion rate per loop-ON cell, $N$ the number of cells in the local microenvironment, $f_{\mathrm{ON}}$ the fraction in the ON state, and $k_{\mathrm{loss}}$ the rate at which ligand disappears from that environment (diffusion into a large medium volume, medium replacement, or blood-flow clearance in tissue):
$$L = \frac{\sigma\, f_{\mathrm{ON}}\, N}{k_{\mathrm{loss}}}, \qquad \text{the loop self-sustains iff } L > L_{\star}.$$
The central empirical claim is that $k_{\mathrm{loss}}$ differs by orders of magnitude between a microtiter well and a dense interstitium, so identical cells with identical secretion rates sit below $L_{\star}$ in vitro and above it in vivo. Because the switch is bistable, the $L$ needed to flip a cell on is higher than the $L$ needed to keep it on — hysteresis — which means the drug dose required to dismantle an established loop exceeds the dose that prevents it, and any model without a two-state variable cannot represent this at all.
Research Program
The closest existing work defines what must be built differently. Latent Causal Diffusions for Single-Cell Perturbation Modeling models single-cell gene expression as an equilibrium diffusion process — a generator that denoises toward a stationary steady state — and its authors concede that their linearized causal readout "may oversimplify extremely complex, multi-stable biological switches." Our model is deliberately non-stationary and multi-stable: the drift toward each attractor depends on $L$, so it can represent a commitment that cannot be undone by removing the stimulus. Cross-modal transfer learning for mapping bulk transcriptomes at cellular level (POPPY) deconvolves bulk patient profiles against static single-cell reference atlases and improved cohort integration by up to 73% by silhouette score, but it maps each sample to states that must already exist in the reference; a mid-transition cell on its way to loop lock-on has no anchor there. Count Bridges enable Modeling and Deconvolving Transcriptomic Data resolves bulk and spatial spots into single-cell count distributions with Poisson bridges and beat CIBERSORTx, MuSiC, and STDeconvolve qualitatively, but each sample is still an independent snapshot. Finally, Cell-Level Virtual Screening showed that drug-induced network rewiring retrieves clinical indications far better than raw expression, and that population averaging inflates network prediction error by 12–47% — evidence that both the network view and the heterogeneity view matter, but it computes a static post-perturbation rewiring on bulk L1000 profiles, with no time axis and no extracellular ligand.
What we build, then, is a temporally coupled extension of count-based deconvolution in which the per-population expression distribution is generated by a latent trajectory with a switch: cells traverse a continuous path from the OFF attractor toward the ON attractor, and the crossing time is a learned latent variable shared across a time series rather than inferred per sample. The intermediate states come from the model's own dynamics, not from a reference atlas, which is precisely what POPPY's cross-attention cannot supply. Onto this we bolt the mass-balance term above, with $\sigma$, $k_{\mathrm{loss}}$, and $L_{\star}$ fit from a small paired calibration experiment (secretion measured by enzyme-linked immunosorbent assay under two culture volumes), and used to predict the ligand concentration — and therefore the loop state — under tissue conditions. The temporal-profiling literature motivates the time course design: Zhihua Liu et al. showed that acute and chronic phases of disease are separated only by longitudinal sampling, and Ruikang Li et al. found that regenerative programs unfold as a sequence of chromatin-opening events across nine time points — sharp transitions that asynchronous averaging flattens unless the model explicitly represents cell-to-cell timing spread.
Baselines
The do-nothing reference is direct transfer of raw in vitro molecular features to in vivo questions, and its ceiling is known: on the drug–disease retrieval benchmark built by Cell-Level Virtual Screening, molecular fingerprints achieved a Hits@1 of 0.0357 against a random baseline of 0.0179, while network rewiring reached 0.1250 — raw molecular features barely beat chance at retrieving clinical indications. We will run this same transfer with our loop-state features and count any retrieval performance above the fingerprint level, with the gap to the 0.1250 network baseline reported alongside, as the minimal bar.
The strongest competitor for single-cell perturbation prediction is chemCPA, Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution, which on sci-Plex3 at 10 µM achieved a median $r^2$ of 0.82 across all genes and 0.79 on differentially expressed genes (versus 0.64 for its non-pretrained version); scPPDM, scPPDM: A Diffusion Model for Single-Cell Drug-Response Prediction, beat the nearest baseline by +36.11% Spearman correlation on unseen drugs in the Tahoe-100M benchmark. We will retrain both on identical folds of our time-course data under the same compute and tuning budget. Beating them means matching their 24-hour endpoint accuracy while additionally predicting delayed, ligand-mediated secondary responses at 48–96 hours, where cell-autonomous models have no mechanism to produce signal. The strongest competitor for in vitro-to-patient translation is DeepSADR, DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction, which achieved an AUC of 0.957 on clinical Sorafenib response versus 0.811 for the next best baseline; our patient-stratification experiments will be judged against DeepSADR-style static transfer run on the same cohorts, and beating it means higher AUC on patients whose tumors are inferred to be loop-ON, not on the cohort average. For deconvolution, the competitor is Count Bridges run independently at each time point; no reported result on Lacuna gives its numeric cell-type-proportion recovery, so we will rerun it ourselves under matched inputs, with Wasserstein-2 distance and absolute commitment-time error as the metrics. Three ablations isolate the key ingredients: removing bistability (single-attractor drift), removing the dilution correction (using in vitro ligand levels as if they applied in vivo), and removing temporal coupling (independent per-time-point inference).
Why not just…?
Why not just run existing deconvolution at every time point? If time points are dense and the ON state already exists in the reference atlas, this may genuinely suffice, and our ablation will say so honestly. But the situation the proposal targets is exactly the one where it fails: POPPY's cross-attention can only attribute bulk profiles to states present in its atlas, and a cell mid-transition into loop lock-on is not in any atlas. Independent snapshots also cannot identify a commitment time when cells progress asynchronously — the commitment appears as a smeared trend rather than a crossing. Our synthetic benchmark, where the true crossing time is known by construction, settles which regime you are in.
Why not just fix the culture conditions instead of modeling them? Growing cells at low volume without medium changes, or in three-dimensional matrices, is a real alternative, and Aldehyde dehydrogenase activity is a cancer stem cell marker of tongue squamous cell carcinoma shows microenvironment architecture governs stemness, so it would help. It does not scale to the 379 drugs and 50 cell lines of Tahoe-100M, and each ligand has its own clearance rate, so a dish tuned to preserve one loop suppresses another. We treat low-dilution culture as calibration data for $\sigma$, $k_{\mathrm{loss}}$, and $L_{\star}$ — a complement, not a substitute, because the model still has to extrapolate from any dish to patient tissue.
Why not just fine-tune scPPDM or chemCPA on in vivo patient data, DeepSADR-style? Fine-tuning is a real threat on static endpoints, and we run it as a baseline rather than argue around it. But it cannot manufacture state variables that were never observed: those models train on 24-hour endpoints in isolated-cell screens, in which autocrine ligands are diluted below threshold, so the loop-ON state is largely absent from their training distribution. DeepSADR's own ablation is cautionary — removing its adaptive readout dropped average AUC from 0.856 to 0.662, showing how brittle static readouts are to distribution shift — and the equilibrium assumption behind latent causal diffusions is stated by its authors to oversimplify multi-stable switches. The decisive experiment is delayed-response prediction and the hysteresis dose-shift, where a fine-tuned static model has no signal source.
Why not just use bulk biomarkers — mutational burden or a stemness signature score? A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer already beat tumor mutational burden head-to-head for immunotherapy response (median AUROC 0.77, with $P = 0.02$ on MSK_2020), and the tamoxifen-resistance study Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties showed treatment failure is driven by expansion of a rare stem-like minority that bulk scores dilute. A signature measures how stem-like the average is; it cannot say which cells flipped, when, or whether the flip will happen in vivo at all.
Proposed Experiments
Four experiments, each anchored to a benchmark in the file. First, a ground-truth deconvolution benchmark: we simulate longitudinal bulk profiles from single-cell references (the Onek1k and 3CA data used by Count Bridges and POPPY) in which a known 5% subpopulation flips on at a known hour, and measure absolute commitment-time error and Wasserstein-2 distance against Count Bridges and POPPY rerun per time point. Second, calibration and translation on the LIF system: we culture the head and neck lines from the LIF–STAT3–SOX2 study (Cal27, Cal33, Fadu, and others) at two medium volumes, measure secreted LIF over a 0–96 hour course, fit the mass-balance parameters, and test whether the inferred loop-ON fraction in TCGA-HNSC tumors stratifies patient outcome beyond what static expression transfer achieves. Third, the 5/6 nephrectomy model of the EP4 study: we ask whether the loop-ON mesangial fraction deconvolved from early post-injury samples predicts the eight-week scarring endpoints that paper reports (blood urea nitrogen of 15.78 mmol/L in wild-type versus 12.45 in EP4-heterozygous mice; creatinine of 28.78 versus 17.52 µmol/L) better than static early composition — the model's claim is that early commitment, not early abundance, predicts scarring. Fourth, perturbation-benchmark hygiene and the hysteresis test: on sci-Plex3 and the cytokine perturbations in scPerturb: Harmonized Single-Cell Perturbation Data — powered per that paper's finding of at least 300 cells per perturbation and 1,000 UMI counts per cell — we verify we lose nothing against chemCPA at 24 hours and win at 48–96 hours, and we measure the dose required to prevent loop establishment versus the dose to dismantle an established loop. The model predicts the ratio of those two doses in advance; no reported result on Lacuna exists for this ratio, which is exactly why it is the decisive check.
References
- Super‐Enhancer Driven LIF/LIFR‐STAT3‐SOX2 Regulatory Feedback Loop Promotes Cancer Stemness in Head and Neck Squamous Cell Carcinoma
- A Maladaptive Role for EP4 Receptors in Mouse Mesangial Cells
- Latent Causal Diffusions for Single-Cell Perturbation Modeling.
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cell-Level Virtual Screening
- Potential biomarkers of acute myocardial infarction based on weighted gene co-expression network analysis
- A scATAC-seq atlas of stasis zone in rat skin burn injury wound process
- Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution
- scPPDM: A Diffusion Model for Single-Cell Drug–Response Prediction
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- Aldehyde dehydrogenase activity is a cancer stem cell marker of tongue squamous cell carcinoma
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- scPerturb: Harmonized Single-Cell Perturbation Data