Pinning Cell Death to Measured Survival in Single-Cell Drug-Response Models

When a cancer cell line is treated with a toxic drug and then sequenced one cell at a time, the treated sample differs from the untreated one for two reasons. Some cells changed their gene expression, and some cells died. Sequencing destroys every cell it measures, so no cell is ever seen both before and after treatment. As a result, a model fit only to the two snapshots cannot tell these explanations apart. A subpopulation that shrinks may have died, or it may have moved to a new expression state. Current models settle the question with a tuning knob: a penalty weight that sets how expensive it is to create or destroy cells compared with moving them. The biology plays no part in that choice. Yet the missing information is usually cheap to measure. How many cells survived at each dose is recorded by a viability assay, or can be approximated from how many cells are recovered per condition relative to the untreated control.

We propose to make measured survival a hard constraint rather than a tuned penalty. The model maps untreated cells to treated cells, and it gives every cell both a new expression state and a survival weight. It is trained so that the surviving mass it implies matches the measured survival for each drug and dose. The constraint is enforced by unsupervised primal-dual learning: one network makes the predictions, and a second "dual" network outputs a condition-specific Lagrange multiplier, a price on constraint violation that rises when the constraint is broken and falls when it is met. This replaces a single global tradeoff with a learned, per-condition one. The claim is that pinning the death-versus-transition split to measured survival gives drug-response predictions that attribute effects correctly and extrapolate across doses better, and that primal-dual training is the stable way to impose this across thousands of conditions in one model.

Core Claim

Two lines of work meet here, and neither closes the gap alone. Single-cell perturbation models based on optimal transport, the problem of moving one distribution onto another at minimal cost, all either conserve cell number or trade growth against movement with a fixed weight. CellOT learns a map from control to treated cells with input-convex networks inside a minimax game, and it strictly conserves mass. W1OT swaps that game for a Wasserstein-1 dual plus an adversarially trained step size. It scales to more than 1,000 genes but still conserves mass. WFR-MFM allows cells to proliferate and die through Wasserstein-Fisher-Rao transport, an unbalanced form of transport that permits mass to appear or vanish. It learns time-averaged velocity and growth fields by regression, but the amount of growth it infers is set by the metric's fixed balance between movement and mass change. These papers motivate the proposal. Unbalanced transport is necessary for cytotoxic drugs, but in current form the total mass is a free byproduct of a hyperparameter.

The work that most directly exposes the gap is PDE-constrained cell-cycle inference. It fits a continuity equation to static snapshots and shows that density pile-ups are ambiguous between arrest and local proliferation. It states that arrest cannot be separated from death without external death-rate data. We take that statement literally: the external data is the survival measurement, and it enters as a constraint. That paper works on a one-dimensional cell-cycle coordinate. We work in expression space with many genes.

The method comes from Self-Supervised Primal-Dual Learning (PDL). PDL trains a primal network together with a dual network that predicts instance-specific multipliers along an augmented-Lagrangian schedule, and it needs no pre-solved labels. On quadratic programs it reports optimality gaps of 0.176%–0.324% with near-zero violations, against 4%–6% for the penalty-and-correction method DC3. Lee et al. similarly found that primal-dual training met wireless power and interference constraints more reliably than penalty losses. Neither has been used for biological populations. In PDL an "instance" is a problem's input parameters. In our setting it is a drug, dose, and cell line, and the constraint value is that condition's measured survival.

Concretely, let $\mu_c$ be the control population for condition $c$ and $\nu_c$ the observed treated population. The primal network outputs a new state $T_\theta(x,c)$ and a survival weight $w_\theta(x,c)\in[0, g_{\max}]$ for each control cell $x$. Here $g_{\max}$ is a division cap: at most $2^{t/\tau_d}$ for assay length $t$ and doubling time $\tau_d$, and it is enforced by parameterization. Let $s_c$ be the measured survival fraction with a noise tolerance $\delta_c$. We solve

$$\min_\theta \; \mathbb{E}_c\Big[\, D\big(T_\theta \# (w_\theta \mu_c),\, \nu_c\big) + \epsilon \textstyle\int w_\theta(x,c)\,\|T_\theta(x,c)-x\|^2\, d\mu_c(x) \Big] \quad \text{s.t.} \quad \Big|\textstyle\int w_\theta(x,c)\, d\mu_c(x) - s_c\Big| \le \delta_c \;\; \forall c,$$

where $T_\theta \# (w_\theta \mu_c)$ is the weighted control population after being moved, $D$ is a Sinkhorn divergence between the normalized distributions, and $\epsilon$ weights the movement cost. The dual network $\lambda_\phi(c)$ predicts the multiplier for each condition's two-sided band. Inside the band the multiplier stays at zero, so noisy survival estimates are not overfit.

Research Program

The model is built in three stages. The first stage is a survival-conditioned transport model. It is a conditional map-plus-weight network over a 1,000-gene space, following W1OT's finding that latent compression is not required. It is conditioned on a drug embedding, the log dose, and the cell line. The second stage adds primal-dual training in PDL's form. The primal step lowers the Lagrangian. The dual network is regressed toward the augmented-Lagrangian multiplier update for each condition, and a global penalty coefficient grows only when violations stall. Plain alternating gradient steps can cycle in non-convex problems, and the augmented-Lagrangian damping is the standard remedy, which is why we inherit it rather than use raw ascent. The third stage is test-time refinement for unseen conditions. For a new drug or dose with only a viability readout, the primal network predicts directly. We then run a few dual steps against the measured $s_c$ and compare this with prediction without the readout, which tells us how much a cheap assay is worth.

Survival values come from per-condition recovered cell numbers relative to vehicle wells in sci-Plex 3. This is a noisy proxy, because capture efficiency varies. That noise is why the constraint is a band, and we will estimate $\delta_c$ from replicate spread. If recovered counts turn out too confounded, the synthetic and time-course experiments below still test the mechanism, and the sci-Plex result would be reported as inconclusive.

Baselines

The naive reference points are two. The first predicts the control population unchanged. The second predicts the nearest training dose of the same drug. Any method that cannot beat nearest-dose on held-out intermediate doses has learned nothing about dose.

The strongest existing unbalanced method is WFR-MFM, conditioned on perturbation embeddings as in its paper. It is the closest competitor because it already models proliferation and death. We will run it on identical gene sets, splits, and GPU budget, with its balance hyperparameter tuned by the same held-out search we give our model. The file reports its inference speed on MEF reprogramming (0.035 s versus more than 1,800 s for WFR-FM) but no accuracy number on sci-Plex 3. For CellOT and W1OT on sci-Plex 3, no reported result was found on Lacuna for held-out distribution error (MMD or Wasserstein): W1OT reports only a 25–45× speedup and 5-minute CPU training, and CellOT reports its sci-Plex results in plots. We will therefore rerun all three ourselves. Beating WFR-MFM means both lower held-out Sinkhorn divergence and implied survival within the measured band on held-out doses. Matching WFR-MFM on distribution fit while being correct on survival also counts, because survival is the quantity it cannot currently get right except by chance.

The mass-conserving methods CellOT and W1OT are included to measure how much unbalancedness matters at all on cytotoxic conditions. They are split by dose so that the difference should grow with toxicity.

Three ablations isolate the key ingredients, and all use the same architecture and tuning budget. The first replaces the constraint with a static penalty $\alpha|\int w - s_c|$ at the best global $\alpha$. The second keeps augmented-Lagrangian training but uses one shared multiplier for all conditions, which removes the dual network. The third is the survival-pinned bisection baseline described below, which serves as the main threat.

Why not just…?

Why not just tune the unbalanced-transport penalty per condition until mass matches survival? This is the real threat. With one scalar constraint per condition, bisection on the growth penalty can hit measured survival on training conditions. To make it amortized and fair, we train a single model conditioned on the penalty value and bisect per condition afterward. The dual network is in effect a learned version of this penalty-per-condition map. Its possible advantages are three: it predicts the price for unseen conditions without a readout, it avoids spending model capacity on a whole range of penalty values, and it extends to vector constraints such as survival at several time points, where a single scalar cannot satisfy all of them. If bisection matches us on sci-Plex 3 held-out doses and on the multi-time-point synthetic task, we will report that the value lies in survival pinning itself and that the dual network is unnecessary. That would still be a useful, simpler result.

Why not just rescale the predicted weights afterward so they sum to measured survival? Uniform rescaling leaves the normalized predicted distribution unchanged, and it leaves unchanged which cells the model treated as dying versus transitioning. The attribution was already fixed by the penalty during training, which is exactly the thing we claim is wrong. The synthetic experiment with known fates measures this directly. A rescaled baseline should have correct total mass but the same attribution error as the unpinned model.

Why not just use a static penalty on the survival mismatch? In optimization benchmarks, penalty-style methods left 4%–6% optimality gaps where PDL reached below 0.33% (PDL). A single $\alpha$ must serve conditions whose survival ranges from near one to near zero, and whose distribution-fit terms differ by orders of magnitude. The static-penalty ablation tests this. We expect per-condition violation to correlate with toxicity under a static penalty but not under the dual network. If it does not, the static penalty suffices and we will say so.

Why not just predict mean shifts, as scGen or conditional autoencoders do? CellOT already showed that mean-shift models miss bimodal responses. Death is a fundamentally distributional effect, because it removes sensitive subpopulations, so a mean-shift model cannot represent it at all. We include scGen only as a reference point.

Proposed Experiments

The first decisive check is synthetic, because only there are true fates known. We simulate branching expression manifolds in 50–1,000 dimensions where each subpopulation has a known death rate and a known transition. Survival is observed with realistic noise, and some settings include survival at multiple time points. The metric is attribution accuracy: the correlation between each cell's predicted weight and its true survival probability, together with the fraction of vanished mass correctly assigned to death. The pinned models should beat every unpinned model on this metric by a wide margin. The dual network should beat the shared-multiplier and bisection baselines when there are multiple time points.

The second check uses sci-Plex 3, holding out intermediate and highest doses per drug and cell line, with a separate split that holds out whole drugs. The metrics are Sinkhorn divergence and MMD against observed treated cells, absolute error between implied and measured survival, and per-condition band-violation rate. We also test enrichment of apoptosis and stress gene programs among the cells the model assigns low weight. That last test is an independent biological check, because those genes are never used in the constraint. The prediction that would falsify the proposal is that pinning improves survival error but worsens distribution fit on extrapolated doses. That would mean the constraint conflicts with the data rather than resolving an ambiguity.

The third check uses the MEF reprogramming and embryoid body time courses used by PACE and WFR-MFM. Here growth is not measured directly, so we use expected mass from proliferation signatures as a loose band. The question is whether pinning at multiple time points changes inferred fates in a way consistent with PACE's reported gains in alignment with RNA velocity (15.4%) and distributional error (23.7%). In this setting the constraint set is a vector per trajectory, which is where a learned dual should matter most.

The final check measures the value of the viability readout. For held-out drugs, we compare prediction without the readout against test-time dual refinement with it. The gap tells an experimentalist whether a plate-reader viability assay can substitute for part of a sequencing run.