Tracking Cells That Change Identity: Joint Association and State-Switching Estimation for Lineage Plasticity

Cancer cells can escape a targeted drug in two ways. They can switch what kind of cell they are, for example from an epithelial "basal" state to a mesenchymal one. Or a subpopulation that was already resistant can simply outgrow the rest. Telling these apart requires following individual living cells over days and recording both who descends from whom and what state each cell is in. Automated cell trackers seem built for this, but they carry an assumption that works against the question. A tracker decides which detection in frame $t+1$ is the same cell as a detection in frame $t$, a step called data association, and it rewards links between detections that look alike. A cell that changes its marker expression, flattens, or stalls in a drug-induced dormant state looks unlike its past self. The tracker is then most likely to break the track at exactly the moment the biology is happening. A broken track reads as one cell disappearing and another appearing, so tracking errors turn plasticity into apparent selection.

This note proposes an association method that treats phenotype change as an expected, modeled event rather than a mismatch. Each track carries a hidden discrete state, such as basal, mesenchymal or luminal. The cost of linking two detections combines motion and division evidence with the probability of the observed phenotype change under a state-switching model. The switching rates, division rates and death rates are estimated jointly with the tracks. Because association is uncertain precisely where cells switch, the rate estimates are averaged over the $m$ most probable association hypotheses rather than a single best track set. The output is what the biology needs: per-cell transition and fate rates with honest error bars, and a direct decomposition of an end-point resistant population into cells that switched state and cells that were already in that state.

Core Claim

Our central claim is that tracking errors in drug-treated cultures are not random. They cluster around state transitions. Standard trackers therefore bias transition-rate estimates downward and inflate apparent birth of new, "unlinked" resistant cells. A tracker that scores phenotype change with a learned switching model can remove most of that bias without losing tracking accuracy on ordinary benchmarks, especially if it also keeps long-range links across dormancy gaps and propagates association uncertainty.

The claim is testable, and it could fail. If identity switches turn out to be spread uniformly over time, a good off-the-shelf tracker followed by counting state changes along tracks would be enough, and we would report that.

Relation to Prior Work

Four lines of work bear on this.

The method comes mainly from global association on graphs. Moral Lineage Tracing solves segmentation and tracking together as an integer program over cell fragments. It enforces a "morality" constraint: cells may divide one-into-two, but lineages never merge. Efficient Algorithms for Moral Lineage Tracing makes this practical with greedy and local-search heuristics, reaching optimality on HeLa in about 15 minutes instead of over 43 hours. We keep the morality constraint as the genealogical backbone and reuse the fast heuristics. Neither paper represents a phenotypic state, so neither can distinguish "this cell changed" from "this is a different cell."

Lifted Disjoint Paths adds long-range "lifted" edges that score whether two distant detections lie on the same path. It raised MOT17 IDF1 from 52.3 to 65.6. We borrow lifted edges to bridge drug-induced arrest and low-contrast periods. We change their cost: in the pedestrian setting a long-range appearance difference is penalized as evidence of different identities, whereas here it is scored against the switching model.

Joint Probabilistic Data Association Revisited shows that the $m$ best association hypotheses capture nearly all posterior mass. We use that idea to make downstream rate estimates uncertainty-aware rather than conditioned on a single track set.

Motivation, and the main competing inference strategy, come from Modeling differentiation-state transitions linked to therapeutic escape in triple-negative breast cancer. That paper fits a five-state switched linear model to fixed-cell snapshots of HCC1143 cells under trametinib and BEZ235. It shows that the drugs reshape transition rates. The authors themselves note that snapshots cannot separate individual reprogramming from differential survival. We adopt their state space and their drug-modulated rates as the generative ground truth for our simulations. The difference is that we estimate rates from linked single-cell histories.

Two further papers constrain the design. Hybrid BAG-seq shows that genetic clones map many-to-many onto expression states. This is why we separate a tree-structured genealogy layer from an unconstrained, possibly reversible phenotype layer. Cell lineage inference from mitochondrial mutations reconstructs static trees from endpoint sequencing. It is complementary to our approach, but it has no temporal resolution.

Deep trackers such as CELLECT link cells over a two-frame window in a learned embedding space. They are fast, but they are exactly the kind of appearance-driven local linker whose failures we expect to concentrate at transitions.

Research Program

Observation model. The imaging model assumes two kinds of channel. A nuclear channel is used for detection and does not depend on state. One or more reporter channels carry phenotype, for example the K14, VIM and K19 markers used by Chapman et al. Each detection $i$ yields a noisy reporter vector, and a small classifier or mixture model maps that vector to a state posterior $q_i(s)$.

Linking cost. Let $\Theta$ denote the per-drug switching matrix together with state-dependent division and death rates. A base or lifted edge from detection $i$ to detection $j$ spans $\Delta$ frames. Its cost is

$$c_{ij} = c^{\text{motion}}_{ij} - \log \sum_{s,s'} q_i(s)\, P_\Theta(s' \mid s, \Delta)\, q_j(s'),$$

where $c^{\text{motion}}_{ij}$ scores displacement and nuclear shape, and $P_\Theta(s' \mid s,\Delta)$ is the $\Delta$-step transition probability. This cost uses phenotype as identity evidence when switching is rare. It stops penalizing a change when the model says change is likely, for instance under a drug that raises exit rates from the basal state.

Division edges receive a separate cost under $\Theta$. Track termination costs include state-dependent death rates, so that selection enters the model explicitly. Association is solved with moral-lineage constraints plus lifted edges, using the greedy and local-search heuristics of Rempfler et al. on a fixed segmentation.

Estimation. We alternate two steps. First, with $\Theta$ fixed, we solve for the $m$ best lineage forests, using a binary partitioning of the solution space in the style of Rezatofighi et al. Second, with those forests fixed, we re-estimate $\Theta$ by running forward–backward inference on each track's hidden state, weighting each forest by its normalized score. Confidence intervals come from bootstrapping over tracks within each hypothesis and mixing across hypotheses.

Analysis output. For every cell in the resistant state at the end of an experiment, the model reports the posterior probability that its lineage entered that state during treatment, which indicates plasticity. It also reports the probability that the lineage was already in that state at treatment onset, which indicates selection.

Baselines

The naive reference is frame-to-frame Hungarian matching on nuclear position with phenotype-similarity appearance costs. After tracking, state changes are counted along the tracks. This is what many labs would run, and it sets the floor.

The strongest existing association method for lineages is moral lineage tracing with the fast heuristics of Efficient Algorithms for Moral Lineage Tracing. We run it on the same detections and segmentation, with the same appearance features and the same wall-clock budget. On Flywing, the original Moral Lineage Tracing reports TRA 0.9813. That is the level of tracking accuracy we must not fall below on standard data.

As a strong learned linker we run a CELLECT-style contrastive embedding tracker trained on the same imaging. CELLECT reports 46% long-trajectory accuracy on C. elegans.

As an upper reference for association quality we use the CTMC result for DeepSORT given ground-truth detections: MOTA 99.32 and IDF1 94.23 in CTMC. That paper also reports how badly automated cell tools do in practice: Viterbi averages TRA 0.39 and DeepCell 0.32.

For rate estimation, the strongest baseline is the snapshot switched-linear model of Chapman et al. We fit it to the same simulated populations, subsampled as fixed-cell counts at matched time points. The file quotes no numerical accuracy for it. We found no reported result on Lacuna for transition-rate recovery from tracked lineages by any method.

Three ablations isolate the ingredients. The first is phenotype-blind tracking, which drops the reporter term. The second is phenotype-as-appearance, which uses a similarity penalty instead of the switching likelihood. The third is single best forest versus the $m$-best mixture. We also toggle lifted edges on and off.

Beating the baselines means two things at once. Transition-rate bias must be at least halved relative to the best tracker-then-count pipeline, with nominal-coverage confidence intervals. TRA on unperturbed Cell Tracking Challenge sequences must stay within 0.01 of moral lineage tracing.

Why not just…?

Why not just use the best available tracker and count state changes along its tracks? This is the real threat. If identity errors are rare or unrelated to transitions, it will work. The first experiment below measures directly whether identity switches concentrate within a window around true transitions. If the ratio of local error to background error is near one and the rate bias is negligible, we will conclude that tracking plus counting suffices and report that as the result.

Why not just fit compartment models to fixed snapshots, as Chapman et al. did? Snapshot counts constrain only net flows. A population that grows in state B can arise from A-to-B switching or from faster division and slower death of cells that were already in B, and the counts alone cannot always tell these apart. Chapman et al. needed separate proliferation and death stains and still raise this ambiguity themselves. We construct simulated regimes, parameterized by their drug-specific rates, in which the two explanations produce matched snapshot counts. We then test whether each method recovers the true decomposition.

Why not just read clones retrospectively from mitochondrial mutations or co-assayed DNA and RNA? Endpoint trees from mt-SCITE reach 96% purity in vitro. They cannot say when a lineage switched or whether it switched back. Hybrid BAG-seq shows that clone and state are decoupled, so knowing the clone does not reveal the history of states. We see these tools as endpoint validation for tracked lineages, not as substitutes.

Why not just remove the phenotype channel from tracking entirely? This is the phenotype-blind ablation. It avoids biasing against switches, but it discards identity evidence during the long stable stretches, and it has no model of arrest gaps. We expect it to lose IDF1 in dense or low-contrast regimes. If it matches the full model, the switching likelihood is unnecessary, and the contribution reduces to lifted edges plus uncertainty propagation.

Proposed Experiments

We have no public live-imaging dataset with ground-truth single-cell phenotype switches. We therefore build a semi-synthetic benchmark that keeps real cell motion and division. Ground-truth lineages come from the Cell Tracking Challenge, using its gold-truth-annotated 2D sequences such as HeLa, and from CTMC. Onto these we write synthetic reporter channels. The reporters are generated by Chapman-style five-state dynamics, with rates set to mimic control, trametinib and BEZ235 conditions. We add three realistic confounds: contrast loss and motility arrest correlated with entering a drug-tolerant state, measurement noise calibrated so that fixed thresholds misclassify states, and death events driven by state-dependent rates.

The decisive checks follow from the claims.

The first check concerns where errors fall. We measure identity-switch density within $\pm k$ frames of true transitions, relative to background, for every baseline. This establishes whether the problem exists.

The second check concerns rate estimation. We measure bias and 95%-interval coverage of the recovered transition, division and death rates as detection noise and arrest duration increase.

The third check targets the plasticity-versus-selection decomposition. We use the matched-snapshot regimes and compute the error in the fraction of end-state cells attributed to switching.

The fourth check guards against degradation on standard data. We report TRA, SEG and IDF1 on the unmodified CTC and CTMC sequences, along with runtime relative to the moral-lineage heuristics.

As a stress test, we sweep the switching rate from zero upward. At zero switching the full model should match phenotype-as-appearance tracking. The gap between the two should widen as switching becomes frequent.

The main claim will be judged by a single figure: rate bias against arrest duration, drawn for each method. If the full model's curve stays flat while the tracker-then-count curves climb, the claim holds.