Sister Cells as Clocks: Inferring Gene Networks from Hydrodynamically Trapped Daughter Pairs

Sequencing a cell's RNA destroys it. Every method that infers how genes regulate one another over time therefore works from disconnected snapshots of different cells, and has to guess which cell at one time point became which cell at the next. We propose to stop guessing for one kind of pair: two sister cells born from the same division. They start with nearly the same molecular state, and live imaging can tell us exactly how long ago that was. A passive microfluidic trap array, which uses only channel geometry and no valves or external fields, can hold a mother cell, record the moment it divides, and move the displaced daughter into the next empty trap. If both sisters are then sequenced at the same moment, the way their transcriptomes have drifted apart over a known interval constrains the interactions between genes directly. Because divisions happen at different times across the array, a single sequencing endpoint covers many elapsed intervals at once.

The core claim is that the growth of sister-to-sister differences as a function of measured time since division identifies the local regulatory interactions that static snapshots cannot. These interactions are captured by the Jacobian, which describes how each gene's rate of change responds to the others. We will test whether this physically measured coupling beats computationally reconstructed couplings, and whether it beats plain correlation, which currently wins the main perturbation benchmark.

Why this gap is real

Dynamic network inference from single cells is limited by what the file calls the destructive measurement barrier. Three computational lines try to work around it. Integrating Optimal Transport and Structural Inference Models for GRN Inference from Single-cell Data stitches snapshots into trajectories with Waddington optimal transport, which pairs cells across time points by minimizing expression change, and then runs a graph variational autoencoder on the stitched paths. It was validated only on 5- and 8-gene synthetic networks. A scalable gene network model of regulatory dynamics in single cells (FLeCS) fits prior-constrained ordinary differential equations to snapshot data at thousands of genes. Interpretable Neural ODEs for Gene Regulatory Network Discovery under Perturbations (PerturbODE) matches start and end distributions through a pathway-module bottleneck, and its authors report that it is unstable on true time series.

All three infer the cell-to-cell coupling from the same data they then fit. That is circular in a way no amount of model capacity fixes. They motivate our proposal, and we reuse PerturbODE's module bottleneck and FLeCS's prior-constrained production terms as model components.

On the device side, the microwell literature reviewed in Microwell array chip-based single-cell analysis already shows SCOPE-seq. That approach images cells in addressable wells over time and then sequences them with barcoded beads, which links a cell's history to its transcriptome. It does not model regulation, and it handles cells as independent units rather than as related pairs. The closest bridge between microfluidics and transcriptomics, Harnessing cell size to separate genetically and functionally distinct dental pulp-derived mesenchymal stromal cell subpopulations, sorts cells by size with Dean flow (a curved-channel secondary flow) and profiles the fractions in bulk. It is static and infers no network.

Two further papers constrain the design. geneRNIB: a living benchmark for gene regulatory network inference finds that Pearson correlation and GRNBoost2 beat multi-omic and foundation-model methods on perturbation metrics, so any gain we claim must survive those baselines. Division state reveals hidden genetic regulation during T cell activation and identifies immune disease–linked gene programmes predicts division state from the transcriptome alone, which raises the question of whether imaging adds anything. We treat that as a direct threat below.

Core Claim

Write $x$ for a cell's expression vector, linearized around its local state as $dx = J(x - \bar{x})\,dt + dW$. Here $J$ is the regulatory Jacobian, which is what we want, and $W$ is intrinsic noise with covariance rate $2D$. For a sister pair, the difference $\delta = x_a - x_b$ has no shared drift, so its covariance $\Sigma(\tau)$ at elapsed time $\tau$ since division obeys

$$\frac{d\Sigma}{d\tau} = J\Sigma + \Sigma J^{\top} + 2D, \qquad \Sigma(0) = \Sigma_0,$$

where $\Sigma_0$ is the partitioning noise at division. At observation, a technical-noise term that does not depend on $\tau$ is added.

A static snapshot sees only a steady-state covariance. That covariance mixes $J$ and $D$ inseparably. The time course from a near-zero start, together with the asymmetric part of $J$, is what carries directional information. The initial slope, $J\Sigma_0 + \Sigma_0 J^{\top} + 2D$, is already non-symmetric in informative ways, and the curvature over longer $\tau$ tells directed edges apart from shared noise.

Estimating a full gene-by-gene $J$ from a few thousand pairs is hopeless. We therefore fit $J = AB$ in a module space of roughly 20 to 50 dimensions, as in PerturbODE, with $B$ sparsified by a curated prior network of transcription-factor targets of the kind MetaTF assembles (Accurate Transcription Factor Activity Inference to Decipher Cell Identity from Single‐Cell Transcriptomic Data with MetaTF). Whether $J$ is identifiable from a finite set of $\tau$ values under realistic dropout is a hypothesis we test in simulation first. We do not assume it.

Research Program

The device builds on serial resistance-based traps (the trap-and-bypass loops described by Huang et al.). In these, an empty trap draws flow and an occupied trap shunts it to the bypass. The pocket is sized so that a mother cell of the suspension line K562 lodges alone. When it divides, the daughter that sticks out is carried by the bypass stream to the next empty trap, a few hundred micrometres downstream. This is exactly the self-regulating loading behaviour the trap is designed for. Brightfield time-lapse records each division and each daughter's arrival, which gives pairing and $\tau$ for every cell.

At the endpoint, traps are sealed with fluorinated oil to address the lysis cross-contamination problem the microwell review flags. Cells are then lysed in place with position-registered barcoded beads, as in SCOPE-seq. Flow rates are kept well below the inertial-focusing regime to limit shear. We also match an off-chip culture and sequence it, so that any mechanosensitive gene programme induced by the chip can be measured and masked.

The first work package is simulation. We generate sister pairs with SERGIO- or BoolODE-style simulators on the BEELINE mCAD and VSC networks used by the optimal-transport paper, and on larger random sparse networks. Division partitioning is binomial, dropout matches 10x-like capture, and pair counts range from 500 to 20,000. The simulation tells us whether the Lyapunov-fit estimator recovers directed edges, and how many pairs it needs, before any device is built.

The second package is the device itself. We measure pairing yield, which is the fraction of divisions where the daughter lands in an identifiable adjacent trap, and doublet contamination.

The third package is a K562 experiment. We run several chips with endpoints spread across 24 to 48 hours of culture, optionally including a drug pulse to widen the dynamic range, and collect a target of at least 3,000 imaged pairs.

Baselines

The naive reference is Pearson correlation and GRNBoost2 computed on the same chip cells with pairing and timing thrown away. geneRNIB identifies these as the methods to beat on perturbation metrics. The file gives no metric values for them, so there is no reported result on Lacuna for their scores on the K562 Replogle or Norman datasets. We will rerun them ourselves under geneRNIB's protocol.

The strongest existing dynamic competitors are FLeCS and WOT+NRI. FLeCS will be given the same cells, treated as snapshots binned by chip endpoint time, and the same prior network. WOT+NRI will be given the same bins, with optimal-transport couplings standing in for the measured pairs. For WOT+NRI's module-level variant we substitute our module bottleneck so that it does not run out of memory. PerturbODE will be run with the same module dimension. There is no reported result on Lacuna for FLeCS, PerturbODE, or WOT+NRI on geneRNIB's K562 metrics. FLeCS is only reported to match or outperform OT-CFM and scTour on A549 trajectories, with no numbers in the file. All methods get the same prior, the same genes (the top 2,000 variable genes plus all prior transcription factors), and the same hyperparameter search budget of 50 configurations, selected on held-out chips.

Two ablations isolate the key ingredient. In the first, pairing is shuffled: cells are randomly re-paired among cells with matching $\tau$, which keeps timing and marginal statistics but destroys shared ancestry. In the second, timing is removed: true pairs are kept but $\tau$ is replaced by its population mean. A third ablation replaces imaged division timing with transcriptome-predicted division state in the style of CellDivider.

The criterion for beating the baselines is a statistically significant improvement over the best baseline on geneRNIB's perturbation-prediction metrics, meaning its $R_1$, $R_2$ and Wasserstein scores for predicting knockdown responses in the external Replogle and Norman K562 data, and a clear drop when pairing is shuffled. If the method beats the baselines but the shuffle costs nothing, the gain is not coming from sisters.

Why not just…?

Why not just sequence many more time points and let optimal transport pair cells? More time points shrink the transport gaps, but the couplings are still inferred from the same expression values being modelled. They are also biased toward pairing cells that look similar, which is precisely the regime in which regulatory drift is invisible. The WOT+NRI paper validates only on toy networks, so there is no evidence it scales. Our shuffled-pairing ablation versus WOT+NRI at matched cell counts measures directly how much real ancestry is worth compared with inferred ancestry.

Why not just use correlation, since geneRNIB says it wins? This is a serious threat. Correlation already wins because perturbation metrics reward capturing co-regulated modules, and sister data may not add to that. The shuffled-pairing and removed-timing ablations, both scored on the same geneRNIB metrics, settle the question. If correlation on chip cells matches the full model, we will report that sister clocks add no information at this scale.

Why not just predict division state from the transcriptome instead of imaging it? The CellDivider approach recovers discrete division counts, not continuous time since the most recent division, and it cannot recover who is whose sister. The CellDivider-substitution ablation measures how much of the gain survives without imaging. If most of it survives, the device reduces to optional timing hardware, and we would recommend a purely computational variant.

Why not just use RNA-velocity or metabolic-labelling estimates of each cell's rate of change? These give per-cell derivatives without a chip. However, they rely on assumed splicing or labelling kinetics, and they say nothing about directed noise propagation between cells with a shared origin. The file contains no benchmarked result for them, so we will add a velocity-derived Jacobian, fitted on the same cells, as an extra baseline rather than argue the point.

Decisive Checks

The first go/no-go test is in simulation. On networks of 50 to 200 genes with realistic dropout, the Lyapunov-fit estimator must recover directed edges, measured by early precision at recall 0.1 against the known truth. It must do better than optimal-transport couplings and correlation with no more than 5,000 pairs, and the shuffled-pairing control must collapse to correlation-level performance. If that fails, the device is not worth building.

On hardware, pairing yield must exceed roughly 30 percent of observed divisions, and we must be able to measure chip-specific shear-response genes and mask them.

On real data, the full model must beat the best baseline on geneRNIB's K562 perturbation metrics, and it must lose that advantage under shuffled pairing. A final sanity check tests whether $\Sigma(\tau)$ actually grows in a structured, non-isotropic way across $\tau$ bins, well above technical noise estimated from sister pairs lysed within minutes of division. If it grows only isotropically, the regulatory signal lies below the measurement floor, and the claim is falsified for this cell type.