acceptodds
Withdrawn from ICLR 2027

Separating Response Ranking and Classification for Bulk-to-Single-Cell Drug Response Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Single-cell drug response prediction helps characterize heterogeneous responses to treatment, providing a finer-grained view of drug sensitivity than population-level measurements. Existing methods transfer drug response supervision from bulk data to single-cell data through domain adaptation. However, transferring drug response information requires both preserving the relative sensitivity of cells and determining an appropriate classification boundary without target response labels. We propose scRA-DRP, a framework that separately adapts response ranking and a classification reference. Starting from a shared domain-adversarial initialization, the ranking branch aligns target features with local source regions constructed using drug response labels. The classification branch uses weighted domain adaptation to adapt the classification rule to the target population and provide a reference number of sensitive cells. This count determines how many top-ranked cells are assigned to the sensitive class, preserving the response ordering. Experiments across 18 drug–single-cell dataset pairs show higher mean performance in both response ranking and classification compared with the evaluated baselines. The ablation studies further support the complementary roles of regional alignment and the separately adapted classification reference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell drug response prediction, Domain adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_vuSCOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper addresses unsupervised bulk-to-single-cell drug response prediction: training on labelled GDSC cell lines and predicting on unlabelled scRNA-seq data. It argues that ranking cells by sensitivity and deciding where to cut that ranking are separate problems. scRA-DRP starts from a shared domain-adversarial model and adapts two copies independently: The ranking branch pulls target features into class-specific regions around selected source anchors, using a bidirectional hinge loss The classification branch uses class-weighted adversarial alignment, plus target-specific BN statistics at inference Both branches use SAM on the source loss At inference, the number of cells the classification branch scores ≥ 0.5 sets k, and the top-k cells by ranking score are labelled sensitive

Critical strengths
  1. The results are transparent and details The authors provide: per-dataset means +- SD (Table 5) a common-subset comparison with DREEP (Table 6) cell-level corrections and new errors (Fig. 3) an explicit note that EpiSen pairs reuse the same cells (App. A.3)
Critical weaknesses
  1. The central claim is not supported. The paper claims the classification branch adapts the sensitive-cell count to the target population, but the evidence doesn’t show this. 1.1 k₀ (Eq. 15) counts cells scored >= 0.5 by a classifier trained on class-balanced batches (Section 3). This builds in a 50 percent prior. The branch’s own adaptation term has little effect, even at β = 0 (Fig. 2c). 1.2 11 of the 18 targets are exactly 50 50 (Table 4). On these, a fixed 50 percent rule already gives the correct count, so they cannot show adaptation. 1.3 On the only strongly imbalanced target (Gefitinib/GSE112274, 37 of 507 sensitive), AUROC is almost 1 but accuracy is 0.73 (Table 5). This implies a count nearly five times too high. 1.4 Model selection is not described, and the defaults match the best test-set values (Section 4.4, Fig. 2).

  2. Limited novelty, with weak evidence for individual components.

2.1 The regional loss (Eqs. 6–8) is a hinge-based form of support alignment, which is not cited: Adversarial Support Alignment (ICLR 2022) and CASUAL (AAAI 2025) . 2.2 The count is “classify and count”, which is known to be biased when class proportions shift. Standard corrections are not discussed or compared, including adjusted counting, EM prior re-estimation (Saerens et al., 2002) and BBSE (Lipton et al., 2018). 2.3 The class weights (Eq. 11) and target BN statistics (Eq. 14) are never ablated separately.

Questions for authors

How were the hyperparameters (M, κ, α, β, ρ, number of epochs) chosen? Were target labels used at any stage, including for Fig. 2 and Table 7? What is the predicted count k versus the true count for each dataset? For Gefitinib/GSE112274 (Table 5), is the accuracy gap caused by an overestimated k?

How does the ranking branch perform with each of these counts:k = n_t/2, the true k, a prior-corrected count (e.g., adjusted classify-and-count or EM) Do the main results hold when all hyperparameters are selected on source validation data only? What do the class weights (Eq. 11) and the target-specific BN statistics (Eq. 14) each contribute when ablated separately?

5 comments
AuthorsOpenReview ↗

W1.1: Class priors, weighted domain adaptation, and target counts

Thank you for your review. After submission, we identified preprocessing inconsistencies between the Docetaxel/Erlotinib datasets and the other GEO pairs. We have unified the preprocessing, rerun the relevant experiments, and updated the affected results. The updated results preserve the main conclusions. Revised text and newly added standard deviations are highlighted in blue in the revised manuscript.

The target sensitive-cell count is determined by the classification predictions for target cells. In Eq. 15, k0=∑j1(cj≥0.5)k_0=\sum_j\mathbf{1}(c_j\geq0.5) counts cells whose classification scores reach the threshold. Here, 0.5 is a per-cell classification threshold, not an assumed target sensitive-cell proportion.

The 50% class prior mentioned in the review arises from class-balanced sampling in Source-Supervised Learning, which gives the two classes equal emphasis in the source drug response classification loss. Weighted Domain Adaptation weights the source-side domain loss by πs(y)/qs(y)\pi_s(y)/q_s(y) to correct, at the population-distribution level, the change induced by balanced sampling. Alignment therefore uses the original empirical source distribution as its reference; training uses weights normalized within each batch. Class-balanced classification supervision and domain alignment against the original source distribution are handled separately. Target expression participates in domain adaptation without class-balanced sampling. The weighting corrects the source-side alignment reference, while target sensitive-cell counts are determined by the classification branch's target predictions.

The parameter β\beta controls the weighted domain-adversarial term in the classification branch’s second stage. At β=0\beta=0, the branch retains the adaptation learned during shared pretraining and uses target-domain batch-normalization statistics at inference. Across 18 pairs and three random seeds with the updated preprocessing, the default β=0.2\beta=0.2 improves mean ACC and Weighted-F1 by 0.57 and 0.66 percentage points, respectively, compared with β=0\beta=0, as shown below. These gains show that the second-stage weighted domain-adversarial term further improves classification, while their modest magnitude may reflect the alignment already established during shared pretraining. AUROC and AUPR remain unchanged because the ranking branch is fixed.

Metricβ=0\beta=0Default β=0.2\beta=0.2
ACC0.76140.7671
Weighted-F10.76300.7696
AuthorsOpenReview ↗

W1.2–W1.4, Q1, and Q2: Target class composition and evaluation settings

W1.2: Applicability to balanced and imbalanced target datasets

As explained in W1.1, our method does not assume that 50% of target cells are sensitive. The count is determined by per-cell classification outputs, k0=∑j1(cj≥0.5)k_0=\sum_j\mathbf{1}(c_j\geq0.5), where 0.5 is the classification threshold. Although 11 targets in Table 4 have a true sensitive proportion of 50%, predictions are made without target response labels or a preset 50% counting rule.

We compare scRA-DRP with SCAD and DAGFormer separately on the 11 balanced targets (50% sensitive cells) and the seven imbalanced targets. SCAD and DAGFormer are the strongest baselines for overall ranking and classification performance, respectively, in the main experiment. The tables report mean results across datasets within each group and three random seeds. scRA-DRP achieves higher mean AUROC, AUPR, ACC, and Weighted-F1 than both baselines in both groups, showing that its performance advantage extends to targets with unequal sensitive/resistant proportions.

11 balanced targets

MethodAUROCAUPRWeighted-F1ACC
scRA-DRP0.87370.88120.80480.8070
SCAD0.78920.78090.50580.5793
DAGFormer0.75580.82290.73630.7367

Seven imbalanced targets

MethodAUROCAUPRWeighted-F1ACC
scRA-DRP0.78100.73670.71420.7043
SCAD0.73700.68490.61380.6481
DAGFormer0.66760.54340.66190.6513

As detailed in our response to Reviewer 2, W5, the original evaluations of SCAD, DAGFormer, scDEAL, scAdaDrug, and FourierDrug cover 6-10 single-cell tasks, 5-9 drugs, and 4-7 cell lines. Our evaluation covers 18 drug-dataset tasks, 14 drugs, and 10 cell lines, extending all three dimensions. GEO includes nine tasks covering seven drugs, eight cell lines, and five cancer types, with true sensitive proportions of approximately 7.3%-60.0%. EpiSen evaluates nine drugs in single-cell populations from JHU006 and SCC47. Adding EpiSen increases total drug coverage from seven drugs in GEO to 14 drugs overall, providing additional evidence across drug conditions. These results support the applicability of scRA-DRP to target datasets with both balanced and imbalanced sensitive/resistant proportions across different drugs and cellular backgrounds.

W1.3: Ranking and classification on Gefitinib/GSE112274

Gefitinib/GSE112274 represents an extreme class-imbalance case in our evaluation. Among the 18 targets, it is the only one with a true sensitive-cell proportion below 10%: 7.30%, compared with 30.15%-60.00% for the other 17 targets.

On this target, scRA-DRP achieves mean AUROC of 0.9949 and AUPR of 0.9528, both higher than all comparison methods, indicating strong response-ranking performance. Its final classification performance is poor on this target. Across the other 17 targets, mean Weighted-F1 and ACC are 0.7683 and 0.7694, respectively, supporting its broader applicability across the evaluated target datasets.

Improving sensitive-cell count estimation and classification calibration under extreme class imbalance is an important direction for our future work.

W1.4, Q1, and Q2: Hyperparameters and source-validation selection

Section 4.1 specifies the default values of the main hyperparameters and training epochs. Section 4.4, Figure 2, and the appendix subsection "Additional Hyperparameter Sensitivity" report the corresponding sensitivity experiments. scRA-DRP uses unsupervised domain adaptation: target response labels are not used as supervision for model training or domain adaptation.

To test whether the main results hold when selection relies exclusively on source validation data, we reselected MM, κ\kappa, α\alpha, β\beta, ρ\rho, and the numbers of training epochs for both stages. The selection criterion was the mean of source-validation AUROC and balanced accuracy. We froze the configurations before evaluating the target sets.

The table compares scRA-DRP with SCAD and DAGFormer over 18 targets and three random seeds.

MethodAUROCAUPRWeighted-F1ACC
scRA-DRP0.78620.77360.68530.6906
SCAD0.74760.72180.56720.6232
DAGFormer0.67430.67290.67190.6681

Under the configurations shown, scRA-DRP achieves higher mean AUROC, AUPR, Weighted-F1, and ACC than both comparison methods. Relative to DAGFormer, the improvements are 11.19, 10.07, 1.33, and 2.25 percentage points, respectively.

AuthorsOpenReview ↗

W2.1: Regional alignment versus ASA and CASUAL

Thank you for pointing out Adversarial Support Alignment (ASA) and CASUAL.

Mechanisms. ASA aligns marginal source and target supports through bidirectional nearest-neighbor distances in the learned discriminator-output space, without explicitly constructing class-specific regions. CASUAL combines features with predicted class probabilities and uses entropy weighting to incorporate prediction confidence for conditional support alignment.

Our method constructs local regions directly in normalized feature space. It selects multiple source anchors within each response class and sets each region's radius from the distance to the nearest opposite-class sample, tying region size to source class separation. The target-to-source term encourages target features to enter source regions and becomes zero inside a region, allowing within-region variation. The reverse term encourages coverage of selected source regions by the current target batch. Targets select regions by geometric distance without restricting matches by predicted class. Our contribution is source-label-guided finite-radius region construction and bidirectional alignment, rather than the hinge form itself.

We evaluated ASA and CASUAL both as complete methods and as replacements for the regional alignment term in scRA-DRP, using 18 drug–dataset pairs and three seeds. Full-method comparisons retain the original training objectives; replacements keep the remaining scRA-DRP components and training settings fixed. Values are mean ± sample standard deviation across three seed-level averages, with equal weights for dataset pairs.

GEO

SettingAUROCAUPRW-F1ACC
scRA-DRP0.7791± 0.02880.7522± 0.04520.7240± 0.01140.7167± 0.0121
ASA (full)0.6410± 0.01520.5906± 0.02100.4634± 0.01640.5810± 0.0106
CASUAL (full)0.6845± 0.05730.6765± 0.02590.6028± 0.02870.6293± 0.0310
ASA (replacement)0.7306± 0.01650.6812± 0.01650.6664± 0.01870.6597± 0.0191
CASUAL (replacement)0.6765± 0.04440.6291± 0.04920.6392± 0.03020.6324± 0.0313

EpiSen

SettingAUROCAUPRW-F1ACC
scRA-DRP0.8962± 0.02420.8979± 0.02480.8152± 0.01020.8174± 0.0099
ASA (full)0.7333± 0.02560.7173± 0.02080.4469± 0.00600.5512± 0.0042
CASUAL (full)0.7080± 0.01650.6868± 0.02460.5204± 0.02970.5763± 0.0209
ASA (replacement)0.8178± 0.02350.8139± 0.02860.7468± 0.01470.7494± 0.0146
CASUAL (replacement)0.7882± 0.02510.7834± 0.01760.7162± 0.02060.7189± 0.0206

Overall

SettingAUROCAUPRW-F1ACC
scRA-DRP0.8377± 0.00590.8250± 0.01070.7696± 0.00570.7671± 0.0060
ASA (full)0.6872± 0.01080.6539± 0.01420.4552± 0.01070.5661± 0.0065
CASUAL (full)0.6962± 0.02320.6816± 0.00670.5616± 0.00670.6028± 0.0131
ASA (replacement)0.7742± 0.01910.7475± 0.01760.7066± 0.01370.7045± 0.0141
CASUAL (replacement)0.7323± 0.03460.7062± 0.03330.6777± 0.02520.6757± 0.0257

scRA-DRP achieves higher mean scores on all four metrics than both complete methods and both replacements in GEO and EpiSen. Compared with the ASA and CASUAL replacements, our regional alignment improves overall AUPR by 7.75 and 11.88 percentage points, respectively. These improvements support the contribution of our regional alignment to response ranking and sensitive-cell selection.

Reviewer_vf4LOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes scRA-DRP, which divides the process of transferring drug response information from bulk RNA-seq to single cells into two branches: the Ranking branch uses local region alignment to rank the relative drug sensitivity of cells, and the Classification branch uses weighted domain adaptation to estimate the number of sensitive cells. Finally, classification is carried out based on the ranking and the number.

Critical strengths

The problem decomposition is reasonable. The sorting and classification thresholds do indeed involve different statistical requirements. Especially in the cross-domain migration scenario from bulk to single-cell, even if the model can correctly arrange the relative sensitivity of cells, it does not mean that the source domain threshold is applicable to the target cell population. The paper clearly separates the two sub-tasks and has certain methodological value.

The comparative experiments are comprehensive. The authors compared methods such as SCAD, scDEAL, DAGFormer, and scAdaDrug, and added the feature representations of single-cell basic models for comparison.

The ablation study also indicates that region alignment helps to improve the sorting performance.

Critical weaknesses

The estimation of the number of classification branches still lacks reliable calibration basis. The paper determines k by counting the number of cells whose classification scores in the target domain exceed 0.5, which is unreasonable.

Unclear task motivation. The necessity of separately modeling response ranking and binary classification is not sufficiently justified. Since both tasks aim to characterize drug sensitivity, it remains unclear why a unified continuous-response prediction framework would not be more appropriate.

Lack of regression-based comparisons. Drug response prediction is commonly formulated as a regression task using continuous sensitivity measurements such as IC50. The paper does not investigate whether regression models could simultaneously support response ranking and binary classification through thresholding. This comparison is important for establishing the necessity of the proposed two-branch framework.

Outdated drug response data. The study relies on an older GDSC-derived dataset and binary annotations from Iorio et al. (2016). The authors do not explain why more recent resources, such as GDSC2 or combined GDSC1/GDSC2 data, were not considered. These resources could support a more comprehensive evaluation of the proposed method.

Limited evaluation diversity. Although 18 drug–dataset pairs are evaluated, several experiments reuse the same single-cell populations. This limits the independence of the evaluation and raises concerns about generalizability across different cell populations and drug response distributions.

Questions for authors

I don't quite understand the value of this task. On one hand, it requires sorting/ranking, and on the other hand, it involves binary classification. So why not do a regression task instead of a classification task?

GDSC was the dataset for the early DRP tasks. Now, most of them use GDSCv2. Why not use a larger-scale dataset? As far as I know, it can definitely be used as a classification task and is widely used.

3 comments
AuthorsOpenReview ↗

W1: Classification threshold and sensitive-cell count estimation

Thank you for your suggestion. After submission, we identified preprocessing inconsistencies between the Docetaxel/Erlotinib datasets and the other GEO pairs. We have unified the preprocessing, rerun the relevant experiments, and updated the affected results. The updated results preserve the main conclusions. Revised text and newly added standard deviations are highlighted in blue.

The classification branch outputs a sigmoid-based sensitive-class probability cic_i for each cell, rather than a sensitive-cell count. We estimate the sensitive-cell count as k0=∑i1(ci≥0.5)k_0=\sum_i\mathbf{1}(c_i\geq0.5), and the ranking branch determines which cells form the sensitive set. A probability threshold of 0.5 is standard for binary classification and is also used in existing single-cell drug response studies and our baseline implementations, including SSDA4Drug, SCAD, scAdaDrug, FourierDrug, and DAGFormer. We have made this description more explicit in the “Ranking-Guided Decisions” section.

To evaluate this rule, we compared five classification thresholds while keeping the ranking fixed, using all 18 drug-dataset pairs and three random seeds after updating the preprocessing. The threshold of 0.5 achieves the highest mean accuracy (ACC) and Weighted-F1.

Table 1. Performance at different classification thresholds

ThresholdACCWeighted-F1
0.400.6646 ± 0.01520.6279 ± 0.0220
0.450.7126 ± 0.01790.6966 ± 0.0228
0.500.7671 ± 0.00600.7696 ± 0.0057
0.550.7346 ± 0.01970.7310 ± 0.0244
0.600.6707 ± 0.01150.6451 ± 0.0191

Using the same data and model outputs, we also compared source error-rate correction, black box shift estimation (BBSE), expectation-maximization (EM) prior re-estimation, and bias-corrected temperature scaling (BCTS). These correction procedures use source-validation information and model predictions on unlabeled target cells; target labels are used only for evaluation. BBSE-hard and error-rate correction are equivalent in this setting and are reported together. None of these alternatives exceeds the original 0.5 counting rule in overall mean classification performance (Table 2), supporting its use in the current evaluation setting.

Table 2. Calibration and adaptation alternatives

RuleACCWeighted-F1
Original 0.5 counting0.7671 ± 0.00600.7696 ± 0.0057
Source error-rate correction / BBSE-hard0.6873 ± 0.00590.6298 ± 0.0141
BBSE-soft0.6629 ± 0.02120.6031 ± 0.0270
EM (source label-proportion prior)0.5628 ± 0.00510.4765 ± 0.0051
EM (mean source-prediction prior)0.6047 ± 0.00800.5333 ± 0.0094
BCTS + 0.5 counting0.5377 ± 0.00080.3886 ± 0.0015
BCTS + EM proportion-based counting0.5543 ± 0.01040.4329 ± 0.0129

Both tables report mean ± standard deviation across three seed-level averages, with equal weight assigned to each drug–dataset pair.

AuthorsOpenReview ↗

W2, W3, and Q1: Task motivation and continuous-response regression

Our goal in bulk-to-single-cell transfer is to obtain both reliable response rankings and accurate sensitive/resistant predictions. These are complementary and equally important: ranking describes relative cellular sensitivity, while classification also requires a boundary between sensitive and resistant cells.

We agree that cell-line drug response prediction commonly uses continuous measurements such as IC50 for regression. In contrast, existing methods for single-cell drug response prediction generally formulate the task as sensitive/resistant classification, as in scDEAL, SCAD, and DAGFormer, among others. Single-cell drug response benchmarks commonly provide binary sensitivity/resistance labels rather than experimentally measured per-cell IC50 values. We therefore cannot directly measure target-domain IC50 regression error, but can assess how well regression predictions distinguish sensitive from resistant cells. This assessment uses the available binary response labels, with AUC/AUPR computed from continuous prediction scores and ACC/Weighted-F1 from binary decisions.

To test a unified continuous-response approach, we used a base network of the same size as each scRA-DRP branch, removed the final sigmoid, and trained it on source log(IC50) using mean squared error and domain-adversarial adaptation. We compared two prediction rules:

  • Standalone regression: rank cells by negative predicted log(IC50), apply min-max normalization within each drug-target dataset and seed, and classify scores ≥0.5 as sensitive. The same regression output thus provides both ranking and classification.
  • Regression with a shared classification count: keep the regression ranking fixed and reuse the sensitive-cell count kk estimated by the scRA-DRP classification branch at threshold 0.5. The top kk cells in the regression ranking are classified as sensitive. This combined control introduces classification-branch information at inference and compares the two rankings under the same count.

The comparison covers 18 pairs and three seeds, using the data in the revised main table, matched source samples and splits, and identical target cells.

MethodAUCAUPRACCWeighted-F1
Standalone regression + min-max + 0.50.6733 ± 0.01020.6650 ± 0.01320.5938 ± 0.00700.5583 ± 0.0157
Regression ranking + shared classification count0.6733 ± 0.01020.6650 ± 0.01320.6351 ± 0.00290.6380 ± 0.0028
scRA-DRP0.8377 ± 0.00590.8250 ± 0.01070.7671 ± 0.00600.7696 ± 0.0057

scRA-DRP exceeds standalone regression on all four overall metrics. Compared with standalone regression, combining the regression ranking with the sensitive-cell count from scRA-DRP’s classification branch improves ACC/Weighted-F1 by 4.12/7.98 percentage points. The full method uses the same count with its own ranking scores, further improving ACC/Weighted-F1 by 13.20/13.16 percentage points over this combined control.

The revised Introduction clarifies the scarcity and typically binary labels of single-cell drug response data and emphasizes the motivation for our task. Reliable bulk-to-single-cell prediction requires both preserving relative drug sensitivity and determining an appropriate sensitive/resistant boundary in the target population.

AuthorsOpenReview ↗

W4, W5, and Q2: Drug response data, source consistency, and evaluation diversity

W4 and Q2: Drug response data and GDSC1/GDSC2 consistency

Thank you for your suggestion. We use the original GDSC data and the drug-specific binary labels released by Iorio et al. to retain the sensitive/resistant definitions of the established SCAD benchmark. This provides consistent, traceable, and reproducible response labels for method comparison.

We examined the GDSC2 release GDSC2_fitted_dose_response_27Oct23.xlsx. It lacks five drugs used in our study, affecting six evaluation pairs. Replacing the current source data entirely with GDSC2 would reduce coverage from 18 to 12 drug-single-cell dataset pairs. We therefore retain the original data as the main benchmark to preserve all current evaluation tasks.

Combining GDSC1 and GDSC2 also requires accounting for differences in assay technology, dose design, and normalization controls, as documented in the official GDSC assay description. For the nine shared drugs relevant to our study, the two resources contain 4,845 common cell-line/drug records with expression data. Their per-drug Spearman rank correlations range from 0.336 to 0.675. Applying the same Iorio binarization rule, with thresholds estimated separately from each source training set, yields 661 discordant labels, or 13.64% of the shared records.

Our task evaluates both cellular sensitivity ranking and classification under a consistent supervision definition. Combining these resources would introduce differences in both source response ordering and binary labels. We therefore do not use their combination as the source for the current tasks.

W5: Diversity of drugs, cell lines, and response distributions

Thank you for highlighting evaluation diversity. Public drug response datasets with per-cell sensitive/resistant labels remain limited. Evaluations in existing studies also include tasks that reuse the same cell lines or single-cell populations. For example, scDEAL and scAdaDrug include two and three tasks, respectively, using the same cell line. In the main single-cell evaluations of SCAD, DAGFormer, and FourierDrug, seven tasks reuse single-cell populations from two cell lines.

The table compares coverage under a similar per-cell classification evaluation setting. DREEP (Pellecchia et al., 2023) primarily evaluates predictions aggregated from single cells to the cell-line level, so its evaluation uses a different counting basis and is not included here.

MethodSingle-cell evaluation tasksDrugsCell lines
scDEAL (Chen et al., 2022)656
scAdaDrug (Liu et al., 2025)1087
SCAD (Zheng et al., 2023)1095
DAGFormer (Yan et al., 2025)1095
FourierDrug (Song et al., 2026)994
scRA-DRP181410

Combining GEO and EpiSen, scRA-DRP evaluates 18 drug-dataset tasks covering 14 drugs and 10 cell lines, exceeding the drug and cell-line coverage of the evaluations listed above. The two collections contribute complementary coverage:

  • GEO: nine tasks covering seven drugs, eight cell lines, and five cancer types. Sensitive-cell proportions range from approximately 7.3% to 60.0%, providing variation in cellular background and response distribution.
  • EpiSen: nine tasks covering nine drugs and two cell lines. Four and five drug tasks reuse the single-cell populations of JHU006 and SCC47, respectively, extending drug coverage within these populations.

Combining evaluation across cell lines in GEO with evaluation across drugs in EpiSen provides broad coverage within the currently available public data.

Reviewer_SADLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors introduce a framework for predicting single-cell drug response from bulk data. It separately adapts response ranking and classification. The framework is evaluated on 18 drug--single-cell dataset pairs, and compared against 8 baselines. The authors perform ablations to understand the contributions of the framework.

Critical strengths

The authors compare against 9 baselines on 18 drug--single-cell dataset pairs, and perform ablation analyses to understand the contributions of components in the framework.

Critical weaknesses

There is no standard deviation on multiple seeded runs to understand the variability of performance and significance of performance gains. Performance gains in Table 1 seem limited.

Ablation analyses (Tables 2-3) show that several components lead to minimal/no performance gains. Again, there is no standard deviation to understand whether improvements in performance are significant.

It is unclear what the novelty and impact of this work is.

Questions for authors

Please clarify the above weaknesses.

1 comment
AuthorsOpenReview ↗

Response on performance, ablations, and contribution

Thank you for your suggestions. We address the three concerns below.

1. Standard deviations and performance gains. Table 5 in the original submission already reports per-dataset means and standard deviations. In the revision, we have also added standard deviations across three random seeds for all methods in Table 1, highlighted in blue. For each seed, we first average scores equally across drug-dataset pairs, then compute the standard deviation across these seed-level averages.

Compared with the strongest baseline for each metric, scRA-DRP improves overall area under the ROC curve (AUC), area under the precision-recall curve (AUPR), Weighted-F1, and accuracy (ACC) by 6.88, 8.15, 6.22, and 6.36 percentage points, respectively. Its corresponding standard deviations are 0.0059, 0.0107, 0.0057, and 0.0060. scRA-DRP also achieves the highest mean scores on all four metrics in both GEO and EpiSen.

After submission, we identified preprocessing inconsistencies between Docetaxel/Erlotinib and the other GEO pairs. We have unified the preprocessing, rerun the relevant experiments, and updated the affected results. The numbers above use these updated results, which preserve the main conclusions.

2. Contributions of the ablated components. The core designs improve mean performance. Compared with the ranking branch alone, the classification branch alone, and single-model inference, the full method improves Weighted-F1 by 16.31, 4.00, and 3.19 percentage points, respectively. Removing regional alignment reduces AUC and AUPR by 8.52 and 11.10 percentage points. We emphasize that in scRA-DRP, the ranking branch affects both ranking metrics (AUROC and AUPR) and classification metrics (W-F1 and accuracy), while the classification branch adjusts the sensitive-cell count, affecting classification metrics without changing AUROC or AUPR. These results support separate adaptation of the two branches and the effectiveness of regional alignment.

We have added standard deviations across three random seeds for all settings in Table 2. The full method has lower cross-seed standard deviations in Weighted-F1 and ACC than all four branch-combination variants. In the regional-alignment ablation, it achieves the highest mean scores across all four metrics while maintaining low cross-seed variability.

Sharpness-aware minimization (SAM) mainly improves stability: while maintaining similar mean performance, it reduces the cross-seed standard deviation of Weighted-F1 from 0.0161 to 0.0057.

3. Novelty and research value. Our central contribution is to adapt response ranking and a classification reference separately, then combine them through a sensitive-cell count constraint. This allows the relative sensitivity of cells and the classification boundary for the target population to be optimized separately. We also propose bidirectional regional alignment guided by source response labels, which transfers discriminative structure while allowing feature variation within each region. The branch-combination and regional-alignment ablations support these designs for improving ranking and classification without target response labels.

We compare against nine baselines across 18 drug-dataset tasks, 14 drugs, and 10 cell lines. This coverage exceeds the evaluations in the original papers of five related methods among our baselines, which cover 6-10 tasks, 5-9 drugs, and 4-7 cell lines. GEO evaluates different cellular backgrounds and response distributions, while EpiSen further expands drug coverage. This experimental coverage and the mechanism-focused ablations jointly support the value of the method.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.