When biologists measure a developing tissue, the measurement destroys the cells, so they see only unconnected snapshots of the population at a few time points. Models that reconstruct trajectories from these snapshots have to answer two questions at once. The first is how many distinct fates or lineages the population splits into. The second is whether a population that grows between snapshots got there by cells moving into it or by cells dividing in place. The literature treats these as separate problems. Unbalanced transport methods let cells divide and die, but they take the number of branches as given or count fates afterwards with ad hoc clustering. Methods that pick the number of clusters automatically come from video motion segmentation, where every tracked point is assumed to persist.
This note argues that the two questions cannot be answered separately, because the same data can be explained either way. A rare cell state that quadruples in place looks, to a model that conserves mass, like a new branch fed by migration from elsewhere. To a model that allows unlimited growth, a genuine new branch can look like a population that simply appeared. We propose to choose the number of lineages inside an unbalanced flow-matching objective. A per-lineage charge, borrowed from minimum-description-length (MDL) motion segmentation, competes directly against the cost of creating mass, so that "new branch" and "local proliferation" are priced in the same currency.
Core Claim
The claim is that jointly selecting lineage number and growth fields gives better fate counts and fewer invented migration paths than either of the two common approaches. The first is to fix mass conservation and then count clusters. The second is to fit growth and then count clusters post hoc. The advantage should appear exactly in a regime we can construct and measure: datasets where branching and local proliferation are confounded. We will build that regime and test whether simpler pipelines already resolve it.
Three ingredients are reused. From unbalanced transport we take the continuity equation with a source term, $\partial_t \rho + \nabla\cdot(\rho v) = g\rho$, where $\rho$ is cell density, $v$ a velocity field and $g$ a local birth-minus-death rate. We also take the simulation-free regression of $(v, g)$ along Wasserstein-Fisher-Rao paths ("traveling Gaussians" whose centre moves and whose mass scales) introduced by WFR-FM. From motion segmentation we take the label-cost energy of Event-Based Motion Segmentation With Spatio-Temporal Graph Cuts, in which a new motion model survives only if it reduces the data cost by more than a fixed charge $\beta$. From spatio-temporal fusion we take a narrower lesson than "use a two-stream network." Slowly varying structure (which lineage a region of tissue belongs to) should be represented separately from fast-varying quantities (each cell's velocity and growth). Here that means a piecewise-constant lineage label field, smoothed over spatial neighbours, sitting on top of continuous per-lineage fields.
How This Relates to Prior Work
Several papers motivate the gap. Branched Schrödinger Bridge Matching learns branch-specific growth rates and shows that balanced transport gets branching ratios wrong, but the user must specify the branch count and tree. Beyond Continuity (USB) combines diffusion with discrete birth-death jumps and recovers growth rates well, but it evaluates fates on fixed, externally clustered cell types. ChronoTILE has a strong multi-scale spatial encoder (cell, niche, tissue), but it assumes a stable population and fixes niche cardinality. That makes it evidence that more architecture alone does not address mass change. STORIES and moscot align spatial slices with fused Gromov-Wasserstein transport, a method that matches cells using both expression similarity and preserved spatial distances. Both rely on largely balanced couplings and pre-annotated lineages. We differ from all five by making the lineage count a variable of the transport objective rather than an input or a post-processing step.
Other papers supply the selection machinery, along with its known weaknesses. The graph-cut label-cost formulation of Event-Based Motion Segmentation With Spatio-Temporal Graph Cuts and the reconstruction-error criterion of 3D Rigid Motion Segmentation with Mixed and Unknown Number of Models both select model order automatically, but only over conserved point tracks. Both also leave the penalty weight as a sensitive hyperparameter, so a large part of this proposal is calibrating it rather than tuning it. Video segmentation by tracing discontinuities in a trajectory embedding avoids choosing $k$ by finding density breaks. Transport regularization deliberately smooths such breaks, which is why we prefer a cost-based criterion over discontinuity tracing.
The closest cross-direction work is ASOT. It uses unbalanced Gromov-Wasserstein transport for unsupervised action segmentation so that some actions can be brief or absent, but it still fixes the number of actions $K$. It is the natural second testbed, described below.
Research Program
The model keeps $K_{\max}$ candidate lineages. Each lineage $k$ has its own velocity network $v_k(x,t)$ and growth network $g_k(x,t)$. Each observed cell $i$ carries a soft lineage assignment $\pi_{ik}$, which for progenitor cells is its fate probability, as in BranchSBM's branch masses. The training objective is
$$ \mathcal{L} = \sum_{k}\sum_{i} \pi_{ik}\, m_t \Big( \lVert v_k(x_{i,t},t) - u_{i,t} \rVert^2 + \lambda\, \big(g_k(x_{i,t},t) - h_{i,t}\big)^2 \Big) + \gamma \sum_{i \sim j} \lVert \pi_i - \pi_j \rVert_1 + \beta\, \lvert \{k : \textstyle\sum_i \pi_{ik} > \epsilon\} \rvert $$
The symbols are as follows. $x_{i,t}$ is a point on the WFR conditional path for cell $i$ at time $t$. $u_{i,t}$ and $h_{i,t}$ are the target velocity and growth of that path, exactly as in WFR-FM, and $m_t$ is its mass weight. $\lambda$ sets the relative cost of creating mass versus moving it. $i \sim j$ ranges over spatial neighbours within a slice and over high-weight transport pairs across slices. $\gamma$ encourages spatially coherent lineage labels. $\beta$ is the charge per active lineage, with $\epsilon$ a small mass threshold.
The first term is where the confound lives. One lineage with large $g_k$ and many lineages with small $g_k$ compete on the same scale, and $\beta$ decides between them.
Optimization alternates between two steps. With the assignments fixed, the per-lineage fields are regressed simulation-free on minibatch unbalanced couplings. With the fields fixed, assignments are updated by expansion moves on the neighbour graph, using the label-cost graph-cut solver. Split moves are proposed when a lineage's residuals are bimodal, and merge moves when combining two lineages lowers $\mathcal{L}$.
The penalty $\beta$ is calibrated, not tuned. We fit the model to parametric-bootstrap replicates generated from the best single-lineage fit. On each replicate we record the loss reduction from the best spurious split, and set $\beta$ at the 95th percentile of that null distribution. This makes $\beta$ the price at which a split would arise by chance about 5% of the time, measured in units of the WFR cost. The same procedure is rerun when $\lambda$ changes, because the growth and branch prices are coupled.
Baselines
The naive reference point is balanced flow matching (WFR-FM with growth fixed at zero) followed by Leiden clustering of terminal cells, with resolution chosen by silhouette score. It is expected to invent branches wherever a subpopulation expands. It also yields a spurious-migration measurement against which every other method is scored.
The strongest existing pipeline for fate counting is USB from Beyond Continuity, followed by the same post-hoc clustering. USB is the strongest competitor because it already models birth and death well. On the Simulation Gene dataset it reports a Pearson correlation of 0.9739 between predicted and true growth rates. Any gain we claim must therefore come from fate counting and lineage assignment, not from growth accuracy. We will run USB and WFR-FM with their released settings and give them the same GPU-hour budget as our method. The clustering step gets the same number of hyperparameter trials as our $\beta$ calibration. Beating this baseline means higher exact fate-count accuracy and higher adjusted Rand index on lineage assignment on the confounded synthetic families, at no worse held-out-time Wasserstein-1 error and relative mass error.
The second strong baseline is BranchSBM from Branched Schrödinger Bridge Matching, swept over branch counts 1 through $K_{\max}$, with the count selected by interpolation error on a held-out time point. It answers the obvious question of whether a model-selection sweep already does the job. It is run at $K_{\max}$ times a single fit's budget, and we also report a budget-matched version.
We found no reported fate-count or branch-number accuracy on Lacuna for BranchSBM, USB or WFR-FM. We also found no reported scalar Wasserstein errors for WFR-FM or ChronoTILE on the axolotl data. Those numbers will be produced by our reruns. For a sense of what automatic cardinality accuracy looks like in another domain, NCRE correctly estimates cluster number on 91.61% of Hopkins155 sequences (3D Rigid Motion Segmentation). This is not a comparable target.
Two ablations isolate the key ingredient. The first removes growth from our objective ($g_k \equiv 0$), giving label-cost selection under mass conservation; it should over-count fates in the proliferation-only family. The second keeps growth but decouples selection, fitting a single lineage and then running the label-cost graph cut on the learned $(v, g)$ fields. This tests whether joint fitting is what matters. A third ablation sets $\gamma = 0$ to measure the contribution of the spatial label smoothing.
Why not just…?
Why not just fit WFR-FM or USB and cluster the trajectories afterwards? This is the most serious threat. A good growth model might already absorb local proliferation, leaving the clustering of terminal states easy. The decoupled ablation and the USB-plus-clustering baseline test this directly. If they match the joint method on the confounded families, we will report that joint selection is unnecessary. The contribution then shrinks to the benchmark and the calibrated $\beta$. Our expectation that they will not match rests on one mechanism. With a single field, the learned $g$ near a branch point blurs two fates with different growth rates into one average. Post-hoc clustering then sees the blurred field, not the two rates. This is the failure BranchSBM documents for balanced transport, one level up.
Why not just sweep the number of branches in BranchSBM and pick the best validation score? This is also a real threat. If interpolation error on a held-out time point reliably peaks at the true count, a sweep is simpler and needs no graph cut. We suspect held-out error decreases almost monotonically with more branches, because extra branches add fitting capacity. That is the over-segmentation trade-off documented in Video segmentation by tracing discontinuities, where a fixed-$k$ baseline had an over-segmentation error of 1.15 versus 0.29 for boundary-based merging. The sweep baseline settles this both at matched compute and at $K_{\max}$ times compute. If the sweep wins at equal cost, we adopt it and drop our optimizer.
Why not just use a bigger spatio-temporal encoder, such as ChronoTILE's multi-scale transformer or a SlowFast-style two-stream network? More capacity does not remove a non-identifiability in the objective. ChronoTILE has a strong hierarchical encoder and still assumes a stable population. We nevertheless include ChronoTILE as a reference on the axolotl data, to check whether its encoder plus a growth head closes the gap.
Proposed Experiments
The decisive experiment uses synthetic data with known division and apoptosis rates, built in the style of the Simulation Gene and Dyngen benchmarks used by USB and WFR-FM. We generate four families with matched final snapshots: true branching with equal growth; a single lineage in which a rare subpopulation proliferates in place; branching with asymmetric growth; and a branch that vanishes through apoptosis. A sweep over proliferation strength tunes how strong the confound is.
We measure five things: exact fate-count accuracy; adjusted Rand index of lineage assignment; Wasserstein-1 error and relative mass error at withheld time points; growth-rate correlation; and a spurious-migration score. That score is the fraction of transported mass moving between modes that are not connected in the ground-truth lineage graph. The central prediction is a crossover. Balanced pipelines should over-count fates in the proliferation family. Unconstrained-growth pipelines should under-count in the asymmetric-branching family. The joint method should stay correct across the proliferation sweep. We will also plot fate-count accuracy against $\beta$ to show whether the bootstrap-calibrated value sits on a plateau, or whether the method is as sensitive to the penalty as graph-cut segmentation is known to be.
On real data, we use EMT and embryoid-body time courses to check that discovered fates do not degrade interpolation. We then use the axolotl brain-regeneration Stereo-seq series (2 to 60 days post-injury, with 10 and 30 days withheld, as in ChronoTILE and STORIES). The test there is whether the method recovers the injury-proximity-dependent fate split of reactive progenitor cells reported by STORIES. Recovery should come without pre-annotated lineages, and without assigning a separate "fate" to regions that merely proliferate after injury.
A secondary transfer test checks that the mechanism is not specific to biology. We add the per-cluster label cost to ASOT on YouTube Instructions and 50-Salads, letting it choose the number of actions rather than fixing $K$. In that setting "a brief segment" and "an absent class" play the roles of proliferation and branching. ASOT reports 52.9% mean-of-frames accuracy on YouTube Instructions and an F1 of 37.4 on 50-Salads with $K$ given. The test is whether estimating $K$ costs little relative to those figures while recovering the annotated action count.
References
- WFR-FM: Simulation-Free Dynamic Unbalanced Optimal Transport
- Event-based Motion Segmentation with Spatio-Temporal Graph Cuts
- Branched Schrödinger Bridge Matching
- Beyond Continuity: Simulation-free Reconstruction of Discrete Branching Dynamics from Single-cell Snapshots
- Multi-Scale Flow Matching for Continuous-time Generative Modeling of Spatiotemporal Tissue Dynamics from Spatial Transcriptomics
- STORIES: learning cell fate landscapes from spatial transcriptomics using optimal transport
- Mapping cells through time and space with moscot
- 3D Rigid Motion Segmentation with Mixed and Unknown Number of Models
- Video segmentation by tracing discontinuities in a trajectory embedding
- Temporally Consistent Unbalanced Optimal Transport for Unsupervised Action Segmentation