Batch-Aware Drug Synergy Scores from Repeated Single-Drug Curves

A drug-combination screen asks whether two drugs together do more than their individual effects predict. The answer is a synergy score: the gap between the measured combination response and a "no-interaction" reference computed from the two single-drug responses. Because the score is a small difference between large, noisy quantities, it is unusually exposed to technical batch effects, meaning systematic, non-biological shifts caused by plate handling, reagent lots, readers, or the laboratory that ran the screen. Public repositories such as DrugComb merge screens from many laboratories, and every synergy predictor we know of treats the resulting scores as ground truth. We propose to fix the labels rather than the predictor. We will estimate each experimental block's technical distortion from something combination screens already contain: the same single-drug curve (drug A in cell line C) is re-measured in every block where A is paired with a different partner. These repeats act as built-in cross-batch replicates. They let us fit block-level nuisance parameters, remove them before computing synergy, and carry the remaining uncertainty through to the score.

The payoff is concrete. If technical distortion accounts for much of the disagreement between laboratories on the same drug pair and cell line, then corrected scores should agree more across studies, some published "synergies" should shrink toward additivity, and predictors trained on corrected labels should transfer better to a laboratory they never saw. None of this has been measured. In the literature we reviewed, batch-correction work stops at omics and images, and synergy work stops at the reference model.

Core Claim

Technical distortion does not simply add noise to synergy scores. It creates synergy or antagonism out of nothing. Let $v_A, v_B, v_{AB} \in [0,1]$ be the true fractions of cells surviving drug A, drug B, and the combination at a given dose pair. The Bliss excess, which assumes the drugs act independently, is $\Delta = v_A v_B - v_{AB}$, where positive means synergy. Now suppose a block's normalization to control wells is off by a multiplicative factor $s$, so every observed survival is $s\,v$. The observed excess is

$$ \Delta_{\text{obs}} = s^2 v_A v_B - s\,v_{AB} = s\,\Delta + s(s-1)\,v_A v_B . $$

A block that over-reads survival by 10% ($s=1.1$) therefore gains a spurious synergy of about $0.11\,v_A v_B$ even when the drugs do not interact. A block that under-reads survival gains spurious antagonism by the same mechanism. Additive offsets and cases where the single drugs and the combination sit on differently distorted plates produce similar artifacts. Loewe and ZIP scores inherit the problem through their fitted single-drug curves.

The claim is that a hierarchical model can identify $s$ and its relatives per block. The model fits monotone dose-response curves shared across blocks for each (drug, cell line), plus per-block location and scale terms shrunk toward zero with empirical Bayes. Correcting the scores this way, rather than rescaling finished scores per study, is what yields cross-study agreement.

Relation to Prior Work

Two lines of batch-effect work motivate the design. Classical results show that multi-source expression data keep clustering by source even after standard normalization (Warnat et al.). reComBat supplies the estimation method: it replaces ComBat's ordinary regression with ridge-penalized regression so that designs with many collinear batches, such as 114 batches in its microarray compendium, stay solvable. We borrow that shrinkage, but apply it to viability surfaces instead of gene vectors. reComBat and its relatives also require explicit batch labels, which is a known limitation. Our block structure is always available, because each dose-response matrix is a block, so the method does not depend on plate metadata that public repositories may not supply.

The replicate idea comes from image profiling. Set-DINO pairs cells from different batches that received the same genetic knockout. CWA-MSN aligns images of identical treatments across wells and plates. Both need replicated perturbations, which combination screens rarely have for any given combination well. Our observation is that combination screens do replicate their single-drug arms, heavily, so the cross-batch anchor exists after all, but on the single-drug side.

On the synergy side, predictors such as DeepSynergy, CongFu, DeepSignalingFlow, and the benchmarking library ChemicalX all consume precomputed Loewe, Bliss, ZIP, or HSA scores and change only the model. DeepSignalingFlow evaluates on NCI ALMANAC, O'Neil, DrugComb, and DrugCombDB separately without modeling differences between studies. We leave the predictors untouched and change their labels. The severe drop that DeepSynergy and CongFu show on unseen drugs and cell lines is consistent with models fitting study-specific structure, though it does not prove it.

Two cautions constrain the design. Deep Learning for BioImaging: What Are We Really Learning? shows that untrained networks nearly match pretrained ones on RxRx3-core (Recall@5% about 0.19 versus about 0.20). High benchmark scores can therefore reflect low-level confounders, so we evaluate correction by agreement across studies rather than by predictor accuracy alone. Normalized diagonal sampling assumes monotone single-drug responses, meaning no paradoxical regrowth at high dose, to prune antibiotic combination designs. We adopt the same assumption as a shape constraint on the shared curves, and we flag blocks where it fails.

Research Program

The model has three parts. First, for each (drug, cell line), a monotone single-drug survival curve is shared across every block in which that drug appears, within and across studies. Second, each block gets a location and scale distortion on the logit-survival scale. These distortions are drawn from a study-level distribution, which is in turn drawn from a global one, with ridge-type shrinkage in the spirit of reComBat. Third, a reference model (Bliss, Loewe, ZIP, or HSA) is applied to the corrected surfaces, and posterior samples give each synergy score an uncertainty interval.

The distortions are identifiable because the same curve for A in C must explain every block's single-drug row for A. How far a block departs from that consensus is the block's distortion, estimated from pairs that share no biology with the combination being scored.

We will build this as a preprocessing library that emits corrected score files compatible with the ChemicalX loaders, so that any existing predictor can be retrained on the new labels without code changes.

Three analyses follow. The first decomposes cross-study disagreement for (drug A, drug B, cell line) triplets measured by more than one study into three parts: technical distortion, residual block noise, and irreducible disagreement. This answers the open question of how much of the disagreement between synergy databases is technical. The second counts how many synergy calls at conventional thresholds lose their credible interval's exclusion of zero. The third retrains existing predictors on corrected labels and tests whether they transfer better across laboratories.

Baselines

The naive reference point is doing nothing: the Bliss, Loewe, ZIP, and HSA scores exactly as distributed in DrugComb and DrugCombDB. Every comparison reports the change relative to these raw labels on identical triplets.

The strongest cheap competitor is per-study standardization of finished synergy scores: z-scoring or quantile-matching each study's score distribution. It is the first thing a skeptic would try, and it removes any global shift between laboratories. It is run on exactly the same triplets with no extra information. Beating it means higher cross-study agreement on shared triplets. It also means better agreement between repeated measurements within a single study, which per-study standardization cannot touch.

The strongest existing batch-correction method is reComBat applied to the finished scores, with study as the batch variable and drug and cell-line identity as protected covariates. Its shrinkage is what our model borrows, so the comparison isolates whether correcting at the viability level, before the nonlinear reference model, matters. It gets the same penalty-tuning budget as our model, chosen on the same held-out study.

Two ablations isolate the key ingredient. In the first, each block's single-drug curves are fit independently, as standard per-matrix synergy tools do, so no borrowing across blocks is possible. This tests whether the repeated single-drug curves carry the benefit. The second keeps the correction but drops uncertainty propagation, which tests whether the intervals matter for deciding which calls are reliable.

For the downstream predictors we use DeepSynergy and CongFu, the strongest reported transductive and graph models in the file. On the O'Neil screen, DeepSynergy reports MSE 255, Pearson 0.73, and classification AUC 0.90 on random splits (DeepSynergy). CongFu reports HSA AUPRC gains of 0.036 to 0.061 over prior methods on DrugComb subsets (CongFu). No reported result was found on Lacuna for training on one study and testing on another (for example, O'Neil to NCI ALMANAC), so we will establish that raw-label number ourselves. Both predictors will also be tuned with RobustTuneC from Ellenbach et al., which selects hyperparameters against an external study. This separates gains from better labels from gains that robust tuning alone would give. All label variants get identical architectures, seeds, and tuning budgets.

Why not just…?

Why not just standardize scores per study? This is a real threat, and the experiments are designed to settle it. The equation above shows that the artifact depends on $v_A v_B$, so it varies across dose pairs and drug pairs within a block. A single per-study shift or scale cannot remove it, and it cannot remove within-study, block-to-block drift at all. The decisive check is agreement between repeated measurements inside a single study, such as repeated single-drug curves and any repeated combination blocks. If per-study standardization matches our model on cross-study agreement and on within-study repeats, the block-level machinery is unnecessary, and we will report that.

Why not just run ComBat or reComBat on the synergy scores? A synergy score is a nonlinear function of three measured responses. Correcting after that function has been applied leaves the distortion entangled with real interaction. COBRA makes the analogous point for expression data: location-and-scale correction of marginals leaves higher-order structure behind. The reComBat-on-scores baseline directly measures how much is lost by correcting too late.

Why not just add study identity as an input to the predictor and let it learn around the noise? This helps only for studies seen in training and does nothing for a new laboratory. It also leaves the labels, and every biological conclusion drawn from them, unchanged. ChemicalX found that simple fingerprint models match graph networks, which suggests extra capacity is not where the difficulty lies. We include a study-indicator variant of DeepSynergy in the cross-study test. We expect it to help within studies and not across them.

Why not just rely on robust hyperparameter tuning? Robust tuning reduces overfitting to one study, but it cannot correct a label that is biased in a fixed direction. We treat it as a complementary baseline and report it with and without corrected labels.

Proposed Experiments

The data are DrugComb, loaded through ChemicalX, together with its constituent O'Neil and NCI ALMANAC screens and DrugCombDB. We will enumerate the (drug, drug, cell line) triplets measured by at least two studies. Their number is not given in the file and will be reported.

The first decisive check is a simulation with known ground truth. We inject block-level scale and offset distortions of realistic magnitude, estimated from the observed spread among repeated single-drug curves, into surfaces generated from non-interacting drugs. We then verify that raw scores produce spurious synergy calls at the rate the equation predicts and that the corrected scores do not.

The second check is held-out single-drug repeats. We remove one block's single-drug row, predict it from the shared curve plus that block's estimated distortion, and compare the error against per-block fitting. This tests whether the distortions are real and transferable.

The third check is cross-study agreement. We compute Spearman correlation and the rate of opposite-sign calls on shared triplets for the raw, per-study-standardized, reComBat, and corrected scores, for all four reference models.

The fourth check is downstream transfer. DeepSynergy and CongFu are trained on O'Neil and tested on NCI ALMANAC, and vice versa, using MSE, Pearson correlation, and AUPRC, with each label variant evaluated against the same target study's corrected labels and, separately, its raw labels. This guards against the circular result of models simply agreeing with our own correction.

The project succeeds if corrected scores raise cross-study agreement beyond per-study standardization, reduce opposite-sign calls, and improve cross-study transfer. It fails informatively if the disagreement proves to be mostly irreducible biology, which would itself tell the field that differences between synergy databases are not a preprocessing problem.