Graded or Fractional? Single-Cell Dose-Response Curves That Separate How Many Cells Respond From How Strongly

A classic dose-response curve reports one number per concentration, such as average viability or average expression of a marker gene. When that average sits halfway between "untreated" and "fully treated," it can mean two very different things. Either every cell has moved halfway (a graded response), or half the cells have switched fully while the rest are untouched (a fractional response). These cases call for different drug decisions. A fractional response points to a resistant or insensitive subpopulation, and raising the dose may only kill bystander cells. A graded response means dose is a real dial. Single-cell RNA sequencing (scRNA-seq), which measures the expression of thousands of genes in each of many individual cells, can in principle tell the two apart. Yet the main machine-learning models for single-cell drug response are built so that they cannot.

We propose a dose-response model for single-cell transcriptomes that gives each cell its own sensitivity threshold. The model predicts two curves per drug and cell population instead of one: the fraction of cells that have responded at each dose, and how strongly a responding cell has changed. It also learns which baseline cell states carry low or high thresholds. The scientific output is a subpopulation-resolved therapeutic window: the dose range in which the intended cells respond and others do not. The modeling claim is that this split can be identified from ordinary snapshot data, even though sequencing destroys each cell and no single cell is ever observed at two doses.

Core Claim

Current perturbation models handle dose by sliding every cell the same amount along a drug direction. scGen (Generative modeling and latent space arithmetics) treats a perturbation as a single difference vector added to a cell's position in an autoencoder's compressed representation, and it has no dose parameter. The Compositional Perturbation Autoencoder, CPA (Learning interpretable cellular responses to complex perturbations), adds a learned nonlinear dose scaling to that vector. The scaling is a function of dose only, so at a given dose every cell receives the same shift and differs from its neighbors only through its starting state. These models therefore encode the graded hypothesis as an architectural assumption. The research file names this problem from the pharmacology side, as population averaging that hides whether an intermediate response is uniform or fractional. It also names it from the modeling side, as Gap 5: models predict mean shifts and miss how the shape of the cell distribution changes with dose. DYNAMO (Gene-Embedding Perturbation Operators) explicitly states that it works only on population averages.

Our model keeps the useful structure of CPA and changes where dose enters. The latent state of cell $i$ at concentration $c$ is

$$z_i(c) = b_i + s(c;\theta_i, h)\,\Delta_d, \qquad s(c;\theta_i,h) = \frac{c^h}{c^h + e^{h\theta_i}},$$

where $b_i$ is the cell's basal (untreated) latent state, $\Delta_d$ is the full-response direction for drug $d$, $\theta_i$ is the cell's log-concentration threshold, and $h$ is a steepness parameter shared within a drug and cell-line pair. The threshold comes from a learned distribution conditioned on basal state, $\theta_i \sim \mathcal{N}(\mu_d(b_i), \sigma_d^2)$. A small spread $\sigma_d$ with a shallow $h$ gives a graded response, because all cells sit at the same intermediate value of $s$. A wide spread with a steep $h$ gives a fractional response, because cells pile up near $s=0$ and $s=1$. The population fraction responding at dose $c$ is then the threshold distribution's cumulative probability at $\log c$. This is the per-cell analogue of a Hill curve, with $\mu_d(b)$ showing which basal states are sensitive.

Identifiability rests on one observation. The magnitude curve and the fraction curve leave different fingerprints on the distribution of cells projected onto $\Delta_d$ at intermediate doses: a single shifted mode under the graded case, two modes with dose-dependent weights under the fractional case. This holds without tracking individual cells, so the destructive-measurement limit of scRNA-seq does not block it. The two cases are confounded only when per-cell noise along $\Delta_d$ is comparable to the response size, and the experiments measure exactly when that happens.

Research Program

The model is built on a CPA-style encoder and decoder. An adversarial term, as in CPA, keeps the basal state $b_i$ free of drug information. The training objective combines a distributional term on held-in doses with the weighted mean-squared error of Needles in the Haystack. The distributional term is an energy distance between predicted and observed cell clouds at each dose. The weighted loss counts genes by how differentially expressed they are, which keeps a few low-dose response genes from being swamped by roughly twenty thousand unchanged ones. This is particularly important for dose curves, where the lowest concentrations may move only a handful of high-affinity target genes.

Sequencing only sees cells that survive, so a fractional response at a cytotoxic dose can look graded once the responders have died. We therefore add a survival term. Cells whose $s$ crosses a learned toxicity level are removed with some probability, and the model is also fit to recovered cell counts per dose. This links the transcriptomic curves to the viability concerns raised in Common and cell-type specific responses to anti-cancer drugs. That work showed that standard potency summaries are confounded by growth rate. It also found conditions with large transcriptional change but no growth arrest, which is exactly where knowing the responding fraction would help.

The work has three stages. First, a synthetic benchmark generates single-cell counts with known threshold distributions, steepness values and death rates. Second, the model is fit to the multi-dose sci-Plex chemical screens: sci-Plex 2 as used by CPA, and the multi-dose, multi-cell-line SCIPLEX screen used by scDEBART. Third, the fitted curves are used to compute selectivity windows across cell lines.

As a secondary step we test generalization to new compounds. The drug direction $\Delta_d$ is predicted from molecular fingerprints, using an encoder trained on the single-dose Tahoe-100M atlas (Tahoe-100M). The threshold distribution and steepness still need multi-dose data. The decomposition assigns jobs to the data: a large single-dose atlas can teach which direction a drug pushes cells, while dose-resolved screens teach where, and in which cells, that push turns on.

Baselines

The naive reference point is the standard pharmacology pipeline applied to single-cell data. Cells are pooled into a pseudobulk profile per dose, a Hill curve is fit per differentially expressed gene, and held-out doses are predicted from those curves, with linear interpolation between measured doses as the cruder version. It is run on the same training doses and the same differentially expressed genes. By construction it cannot represent distribution shape. It counts as beaten if our model matches it on mean-profile accuracy and clearly improves distributional accuracy at held-out intermediate doses. Alongside it we report the dataset-mean and technical-duplicate references from Needles in the Haystack, so that every score sits between a calibrated zero and the measurement ceiling.

The strongest existing method is CPA, which is the only model in the file that combines continuous dose with single-cell outputs and was tested on held-out doses. On sci-Plex 2 out-of-distribution dose splits, CPA reports $R^2$ of 0.92 for Nutlin and 0.94 for BMS, against 0.85 and 0.89 for linear interpolation (CPA). We will retrain CPA on our exact splits with the same latent size and an equal hyperparameter search budget. We will also run a version of CPA trained with weighted mean-squared error, so that any gain cannot be credited to the loss function alone. For distributional metrics on held-out doses, we found no reported result on Lacuna for CPA or any other model, so these numbers must be produced here. Beating CPA means two things. Our model must stay within noise of CPA's mean-profile $R^2$ and weighted delta $R^2$. It must also significantly reduce energy distance at held-out intermediate doses for drug and cell-line pairs where the data show bimodality along the drug direction.

The third baseline is a generic dose-conditioned generative model. We adapt the dual-time flow matching of Joint Flow Matching from image latents to the same transcriptome latent, with concentration as a continuous input. Flow matching learns a generator by fitting a velocity field that carries noise to data. This baseline can in principle produce any distribution shape. It tests whether our explicit threshold structure adds anything beyond flexibility.

Two ablations isolate the key ingredient. The first fixes all thresholds to one value per drug, removing per-cell sensitivity and collapsing the model to a CPA-like graded model with a Hill-shaped dose scaling. The second keeps the spread but makes thresholds independent of basal state, which removes the subpopulation attribution.

Why not just…?

Why not just fit Hill curves to pseudobulk, or cluster treated cells and count responders per dose? Pseudobulk fits are the "do nothing" baseline above, and by definition they cannot tell graded from fractional. Cluster-and-count is the serious threat, because on strongly bimodal drugs at measured doses it may recover the fraction curve almost as well as our model. It has three weaknesses. It needs a responder cluster defined by hand at high dose. It misassigns cells at intermediate doses where responses are partial. It cannot predict unmeasured doses or attribute sensitivity to basal state. On the synthetic benchmark we will run cluster-and-count with oracle-tuned clustering and compare fraction-curve error across steepness values. If it matches our model on real data even at held-out doses, we will report that the per-cell threshold machinery is only needed in the intermediate-steepness regime and narrow the claim to that regime.

Why not just use a flexible conditional generator such as flow matching on transcriptomes? Such a model could fit bimodal marginals, but it outputs samples, not a fraction curve, a magnitude curve or a sensitive-state map. Its interpolation between measured doses is also unconstrained. Joint Flow Matching reports smooth, monotone latent paths with dose (Joint Flow Matching), but only for images and with one model per drug. Its dose-estimation accuracy, 68% within one neighboring concentration, suggests that dose information in cell clouds is real but noisy. The held-out-dose comparison settles which approach interpolates distribution shape better. If the flow model wins, we would keep our decomposition as a readout fitted on top of flow samples rather than as the generator.

Why not just train CPA with a weighted loss? The weighting of Needles in the Haystack raised variance recovery for GEARS from 14.51% to 40.56% on Replogle22. It fixes which genes count, but not CPA's assumption that every cell shifts by the same dose-dependent amount. The CPA-plus-weighted-loss baseline and the fixed-threshold ablation together isolate whether per-cell thresholds, rather than the loss, produce the gain.

Why not just profile doses with cheap bulk methods like L1000? L1000 profiles roughly a thousand landmark transcripts in bulk lysates. It gave multi-dose coverage across about 8,000 conditions (Common and cell-type specific responses), but bulk lysates erase the distinction this proposal measures. If our analysis finds that almost all intermediate responses in sci-Plex are graded, that would be a useful negative result. It would mean bulk dose curves lose little, and the field could save single-cell budgets for other questions.

Proposed Experiments

The first decisive check is synthetic recovery. We simulate counts with realistic sparsity, a known threshold mean and spread, steepness and a death rate. We then measure the error of the recovered fraction and magnitude curves, and map the region of noise-to-response ratio where graded and fractional responses become confounded. The model should recover truth outside that region and report wide uncertainty inside it. Recovery also needs to hold when the baseline methods fail, otherwise real-data claims are unsupported.

The second check is prediction at held-out doses on sci-Plex 2 and the multi-dose SCIPLEX screen. Each intermediate dose is removed in turn, and we report weighted delta $R^2$ on means, energy distance on full cell clouds, and agreement of predicted and observed bimodality along the drug direction. The technical-duplicate ceiling is reported alongside each. The key comparison is our model against CPA and the fixed-threshold ablation, restricted to pairs flagged as fractional in the full data. That restriction keeps the effect from being averaged away over the many drugs that barely act.

The third check is a census. Across all multi-dose drug and cell-line pairs, we count how many intermediate responses are fractional versus graded, and test whether fitted threshold spreads agree between biological replicates. Without that agreement the decomposition is noise, not biology.

The fourth check concerns subpopulation attribution and selectivity. Within each cell line, we test whether basal states predicted to be insensitive are enriched among cells that remain unshifted at high dose in held-out replicate data. Across cell lines screened together, we compute predicted selectivity windows from the fraction curves and compare them with windows measured directly from the held-out doses.

As a stretch goal, fingerprint-predicted drug directions trained on Tahoe-100M are used for compounds held out of the multi-dose training set. Only the threshold distribution and steepness are fit, from two of the measured doses, and the remaining doses are scored.