When a blood progenitor cell is tagged with a heritable DNA barcode and its descendants are followed, sister cells tend to choose the same fate far more often than their measured gene expression predicts. Lineage tracing on transcriptional landscapes links state to fate during differentiation reported that a network predicting fate from a cell's own transcriptome was 26% less accurate than predicting it from a sister's fate. The standard reading is that cells carry hidden heritable state, such as chromatin, protein reserves, or division history, that single-cell RNA sequencing cannot see. This proposal tests a competing explanation drawn from distributed control. Sisters also share an environment, and in particular a delayed one. Cytokine signals from mature cells reach progenitors only after secretion, diffusion and receptor turnover, so sisters act on the same stale report of the population state. In networks, many agents acting on the same stale report produce synchronized over-corrections, or "herd effects."
The core claim is testable. Part of the sister-concordance gap is shared delayed environmental exposure rather than intrinsic memory. If that holds, fate choice depends on population composition some delay τ earlier, and culture composition should overshoot or oscillate once τ is large relative to the division time. We would separate the two explanations with a split-sister design and a controlled feedback delay, and we would fit a model in which fate depends on a clone-specific latent state plus the environment at $t-\tau$.
Why this is the right intersection
Delay-instability work supplies the mechanism and the knob to turn. Integrating local static and dynamic information for routing traffic shows that routers steering by neighbors' stale queue reports all rush to the same "empty" path and then all abandon it, producing load oscillations. Dynamic Network Service Selection in Intelligent Reflecting Surface-Enabled Wireless Systems proves that delayed replicator dynamics have a critical latency above which allocation oscillates. Load Balancing in Compute Clusters With Delayed Feedback shows that tracking a belief over the delayed state restores stability where memoryless heuristics fail. These papers motivate a specific hypothesis: progenitor fate choice is a routing decision made on stale population information.
Hematopoiesis modeling already establishes that the feedback exists but does not reach the clonal level. Multistage feedback driven compartmental dynamics of hematopoiesis makes upstream self-renewal and differentiation depend on downstream mature-cell densities and finds oscillations and alternative equilibria. It is deterministic, works on aggregate counts, and leaves delay implicit in transit through compartments. It therefore cannot say whether feedback explains sister concordance.
Lineage-tracing methods cover the complementary ground and treat differentiation as autonomous flow. CoSpar (Learning dynamics by computational integration of single cell genomic and lineage information) assumes transcriptionally similar cells have similar fates. That assumption fails if identical states sampled at different phases of a feedback cycle diverge. Super-OT (Optimal Transport using GANs for Lineage Tracing) learns a deterministic map between time slices. Deep Lineage (Deep Lineage: Single-Cell Lineage Tracing and Fate Inference Using Deep Learning) runs recurrent networks over discrete sampling days. None of the three has a population-to-cell feedback term, delayed or not.
This proposal differs from the obvious combination, which would fit a neural delay differential equation (Neural Delay Differential Equations) to existing barcoded snapshots. We reject that combination below, because three time points cannot identify a delay. We instead make the delay an experimental variable.
Core Claim and Model
Let $p_i$ be the probability that cell $i$ commits to fate A rather than fate B, for example neutrophil versus monocyte. Let $x_i$ be the cell's transcriptome embedding, $z_c$ a latent vector shared by all members of clone $c$ (the intrinsic heritable memory), and $E_w(t)$ the composition of culture well $w$, such as the fractions of mature lineages. The model is
$$\mathrm{logit}\, p_i = g(x_i, z_c) + h\big(E_w(t-\tau)\big),$$
where $g$ and $h$ are small neural networks and $\tau$ is a learned feedback delay. Composition evolves as a delay equation,
$$\frac{dE_w}{dt} = F\big(E_w(t), E_w(t-\tau)\big),$$
with $F$ built by aggregating the cell-level fate probabilities and division rates. The first term alone recovers "hidden heritable variable" models. The second term alone recovers compartmental feedback models. The claim is that both are needed, that the fitted τ is substantially positive, and that the $h$ term absorbs a measurable share of sister concordance.
Research Program
The experiment is the load-bearing part. Mouse hematopoietic progenitors are barcoded LARRY-style and allowed to divide once or twice. Sisters are then deliberately split between wells that receive different feedback histories. The delay knob is conditioned medium. A donor culture enriched for mature myeloid cells is sampled continuously, and each recipient well is fed medium harvested τ hours earlier, with τ set to 0, 12, 24, or 48 hours. A fresh-medium well serves as a no-feedback control, and a heat-inactivated-medium well controls for nutrient depletion. This construction mirrors the latency sweeps of the network papers.
Composition is tracked densely and cheaply by flow cytometry on lineage markers, for example every six hours. This dense series is what makes τ identifiable. Barcoded single-cell RNA sequencing is performed at three to four time points to read clone identity and fate.
The analysis decomposes sister-fate concordance three ways. Same-well sisters share both intrinsic memory and environment. Split-well sisters with matched τ share intrinsic memory and an equivalent environment. Split-well sisters with different τ share intrinsic memory only. The concordance drop from the second group to the third estimates the environmental share directly, without the model.
The model is then fit on a subset of τ conditions. It is judged on whether it predicts composition trajectories and clonal fate biases in held-out τ conditions.
A secondary arm tests the herding prediction using clonal division synchrony. Clonal memory of cell division in humans diverges between healthy haematopoiesis and acute myeloid leukaemia found that healthy human progenitor sisters and cousins divide with correlation R between 0.74 and 0.9, and that this synchrony is lost in leukemia. In our model, synchronized clones emit mature output in bursts. Bursts should enlarge the overshoot under long τ, just as synchronized senders worsen incast in Pulser. We would compare long-τ overshoot between wells seeded with healthy progenitors and wells seeded with leukemic progenitors. We would also treat healthy progenitors with the division-desynchronizing condition and leukemic progenitors with JQ1, which that paper reports partially restores synchrony.
Baselines
The naive reference point is forward prediction from each cell's transcriptome alone. For it, the file reports only a relative figure: 26% less accurate than lateral sister prediction (Lineage tracing on transcriptional landscapes).
The strongest existing method is Deep Lineage, which reports 77% early fate-bias accuracy on the LARRY hematopoiesis data. The same paper reports 65% for CoSpar and 65% for Waddington-OT, and 52% for Deep Lineage with clonal information removed (Deep Lineage). Super-OT reports 71.88% on monocyte versus neutrophil prediction, against a supervised bound of about 75.3% (Optimal Transport using GANs for Lineage Tracing). These numbers come from the published LARRY data. On our new data, every method is retrained on identical train/test splits held out by clone and by τ condition, with the same hyperparameter search budget. Deep Lineage and CoSpar also receive well identity and current well composition as extra inputs, so that any advantage we show is not merely an information advantage.
The model is also compared against a Mon Père-style compartmental model (Multistage feedback driven compartmental dynamics) fitted to the flow-cytometry counts, on composition trajectories.
The ablations isolate the delay:
- the model with τ fixed at 0, so the environment is current;
- the model with $h$ removed, so fate depends on intrinsic memory only;
- the model with $z_c$ removed, so fate depends on environment only.
Beating the baselines means lower held-out-τ composition error than the compartmental model and the τ=0 ablation, and higher held-out fate-bias accuracy than Deep Lineage given the same inputs. The key extrapolation test is training on τ values of 0 and 24 hours and testing at 48 hours.
Why not just…?
Why not just fit a neural delay equation to the existing LARRY snapshots? The data have days 2, 4 and 6 only. With three snapshots, a delay term is not separable from a slower rate or from an extra hidden dimension, and any fitted τ would be an artifact. Neural Delay Differential Equations was validated on densely sampled Mackey-Glass series, which is precisely what lineage snapshots lack. This is why the delay becomes a designed intervention with dense, cheap readout.
Why not just add current well composition as a covariate, with no delay? This is the most serious threat, and it is exactly the τ=0 ablation. If the τ=0 model matches the delayed model on held-out τ conditions and no overshoot appears at long τ, the delay claim fails. The weaker claim that shared environment explains part of sister concordance would still stand if the split-sister contrast shows it. The theory in Dynamic Network Service Selection predicts that a zero-delay model cannot produce overshoot that grows with τ, which gives a clean distinguishing signature.
Why not just give Deep Lineage the well identity? On seen wells, this would likely match our fate accuracy, and we expect that result. What well identity cannot do is predict a new τ condition, because it names the environment without modeling it. The held-out-τ test settles whether mechanism adds anything beyond memorization.
Why not just accept the hidden-heritable-variable explanation? Existing evidence such as the 26% gap and Deep Lineage's 77%-to-52% ablation cannot distinguish intrinsic memory from shared exposure, because sisters in those designs share both. The split-well concordance contrast is the first measurement that does.
Decisive Checks
The proposal succeeds or fails on three observations. First, concordance must be lower for split-well sisters under different τ than under matched τ. If there is no drop, the environmental share is negligible and the hidden-variable view is strengthened, which is itself a useful result. Second, composition overshoot must grow with τ and be predicted out of sample by the delayed model but not by the τ=0 ablation. Third, as a weaker, secondary check, overshoot at long τ should be larger in synchrony-preserving cultures than in desynchronized ones. That would link the clonal metronome to delay-induced herding. A null on the third check would not undermine the first two.
References
- Lineage tracing on transcriptional landscapes links state to fate during differentiation
- Integrating local static and dynamic information for routing traffic
- Dynamic Network Service Selection in Intelligent Reflecting Surface-Enabled Wireless Systems: Game Theory Approaches
- Load Balancing in Compute Clusters With Delayed Feedback
- Multistage feedback-driven compartmental dynamics of hematopoiesis
- Learning dynamics by computational integration of single cell genomic and lineage information
- Optimal Transport using GANs for Lineage Tracing
- Deep Lineage: Single-Cell Lineage Tracing and Fate Inference Using Deep Learning
- Neural Delay Differential Equations
- Clonal memory of cell division in humans diverges between healthy haematopoiesis and acute myeloid leukaemia
- Pulser: Fast Congestion Response Using Explicit Incast Notifications for Datacenter Networks