acceptodds
Under review as a conference paper at ICLR 2027

BASIL: Batch-Aware Single-Cell Perturbation Response Modeling with Conditional Distribution Alignment

Abstract

Predicting transcriptional responses to unseen genetic perturbations can reduce the experimental search space in functional genomics. However, single-cell perturbation screens are complicated by experimental batch variation and by destructive sequencing, which provides no one-to-one correspondence between control and perturbed cells. We introduce BASIL, a batch-aware framework centered on batch-conditioned dual-resolution learning. For each perturbation–batch condition, same-batch control cells define an explicit pseudo-bulk response target, while the same deterministic state-conditional response operator transforms individual control cells and is regularized against the corresponding cell-unpaired perturbed population. Conditional maximum mean discrepancy (MMD) provides this population-level regularization without assuming cell-level pairs. To support out-of-distribution prediction, BASIL combines GenePT features with multi-relational biological priors to construct perturbation representations that remain available for unseen genes and combinations. Across three genome-scale CRISPRi datasets and the Norman CRISPRa benchmark, BASIL achieves strong performance against representative perturbation-prediction baselines, including challenging unseen-condition and gene-holdout combinatorial settings. Ablations further support the roles of condition-level pseudo-bulk supervision, batch-aware referencing, GenePT-based perturbation features, and conditional population regularization. Overall, BASIL provides a practical direct predictor for robust genetic perturbation response modeling under batch variation and cell-unpaired observations. Code is available at https://anonymous.4open.science/r/BASIL-main-624C.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell perturbation prediction, virtual cell, batch-aware modeling, distribution alignment

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Ho7dOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents BASIL, a batch-aware framework for predicting transcriptional responses to unseen genetic perturbations using single-cell data. The authors identify two important challenges in perturbation-response modeling: experimental batch effects, which can confound biological responses, and the absence of paired control–perturbed cells due to destructive sequencing. To address these challenges, BASIL introduces a dual-resolution learning framework that combines batch-matched pseudo-bulk supervision with condition-specific distribution alignment. To support generalization to unseen genes, the framework further combines GenePT embeddings with a multi-relational biological graph encoder incorporating transcription-factor regulation, Gene Ontology similarity, and protein interactions. The evaluation covers three genome-scale CRISPRi datasets and the Norman combinatorial CRISPRa benchmark, including unseen-condition, one-unseen-gene, and two-unseen-gene settings. The paper reports strong performance relative to several statistical, graph-based, foundation-model, and distribution-aware baselines.

Critical strengths

S1. The paper correctly distinguishes matching experimental conditions from establishing one-to-one cellular correspondences. Its combination of batch-specific reference states and population-level regularization is well motivated by the structure of the available data.

S2. The framework provides a coherent connection between mean-response supervision and distributional modeling. This design avoids training completely separate mean-prediction and distribution-generation models. The pseudo-bulk objective provides a relatively stable target for perturbation-effect estimation, while conditional MMD offers an additional constraint on the predicted single-cell population.

S3. The evaluation includes meaningful out-of-distribution generalization settings. Unlike evaluations that merely hold out cells from previously observed perturbation conditions, BASIL evaluates unseen perturbation identities.

Critical weaknesses

W1. The methodological novelty is somewhat incremental relative to existing perturbation-prediction approaches. The proposed framework largely integrates several established ideas. E.g., 1) Batch-matched control referencing and biological knowledge graphs have been investigated in prior methods, including TxPert. 2) Pseudo-bulk supervision and direct prediction of population-level perturbation effects are established approaches. 3) MMD-based distribution alignment has previously been employed in single-cell perturbation modeling, including scDFM. 4) GenePT and relation-aware graph propagation provide an established mechanism for representing unseen genes.

The main contribution therefore appears to be the specific integration of these components through a shared response operator rather than a fundamentally new modeling principle. This integration is reasonable, but the paper should more clearly identify what cannot be achieved by simpler combinations of existing techniques. In particular, the relationship to TxPert and scDFM should be sharpened. An important question is whether the improvement primarily comes from stronger gene representations, improved batch referencing, or the proposed dual-resolution objective.

W2. The empirical contribution of conditional MMD is surprisingly small, and its intended distributional benefit is not directly evaluated. This is my primary technical concern. Conditional MMD is central to the paper's title and methodological motivation. However, Table 3 shows that removing conditional MMD decreases Pearson delta only from 0.688 to 0.684 on RPE1 and from 0.671 to 0.668 on HepG2. These gains are modest, particularly relative to the much larger contributions from GenePT features and pseudo-bulk supervision.

W3. The biological perturbation encoder is not sufficiently isolated experimentally. BASIL combines GenePT embeddings with three biological relation types: transcription-factor regulation, Gene Ontology similarity, and STRING protein interactions. The paper provides an ablation replacing GenePT with nonsemantic features, but this changes a major source of perturbation information rather than independently assessing the graph encoder. Important missing comparisons include GenePT without graph propagation, graph propagation with individual relation types removed, and simpler perturbation encoders with matched model capacity.

These experiments would clarify whether the biological graph contributes meaningful information beyond that already encoded by GenePT. This issue is particularly important because the strongest results occur under unseen-gene generalization, where pretrained biological representations are likely to play a central role.

W4. The proposed combinatorial representation is relatively simple and does not explicitly model genetic interactions. For multiple perturbations, BASIL averages their graph-encoded gene embeddings. Although this produces a permutation-invariant representation, it does not explicitly represent interactions between the constituent perturbations before fusion with the basal cellular state. This raises a question about the interpretation of the strong two-unseen-gene results.

W5. The experimental evaluation does not fully establish robustness to batch variation or generalization across biological contexts. Although batch awareness is a primary motivation, the ablation removing batch matching produces relatively modest performance changes: from 0.688 to 0.679 on RPE1 and from 0.671 to 0.654 on HepG2. These results indicate that matching can help, but they do not establish robustness under substantial batch shifts. A more direct experiment would evaluate performance across batches with different technical characteristics, including held-out batches or deliberately challenging batch-transfer settings.

Questions for authors

Q1. Since conditional MMD is a central component, can the authors report quantitative evaluation of held-out predicted cell distributions, rather than relying exclusively on mean-effect Pearson correlation?

Q2. Are the small improvements from conditional MMD statistically significant across the four random seeds? Please report standard deviations or confidence intervals, preferably for all principal comparisons.

Q3. How much improvement comes from multi-relational graph propagation beyond the pretrained GenePT representations? Please provide GenePT-only and relation-specific graph ablations.

Q4. How does BASIL perform on strongly non-additive gene combinations? Can the authors compare against a simple additive response predictor and report results stratified by genetic-interaction strength?

4 comments
AuthorsOpenReview ↗

W1-W2 Comment

W1: BASIL’s methodological contribution and the sources of its gains need clearer distinction from established approaches. 1: Shared response formulation and relation to prior work. BASIL’s contribution is a shared, state-dependent response operator jointly trained at population-mean and single-cell resolutions. For each perturbation–batch condition (p,b)(p,b), mean supervision targets the observed response yˉp,b−xˉb\bar{y}_{p,b}-\bar{x}_b, while conditional MMD constrains the population generated from individual control cells:

\mathcal{L}=\mathcal{L}_{\mathrm{mean}}+\lambda(t)\mathcal{L}_{\mathrm{MMD}}.$$ Both objectives update the **same response function**, without assuming correspondence between individual control and perturbed cells. Mean supervision anchors the population response, while conditional MMD constrains cell-level outputs beyond their mean. TxPert combines biological graphs and basal-state representations primarily through expression-MSE training; BASIL couples explicit mean-response supervision with conditional population alignment. scDFM combines conditional flow matching and MMD to learn a time-dependent velocity field; BASIL directly learns an expression-response mapping at both resolutions. We will make these distinctions explicit in the revision. **2: Controlled evidence and component attribution.** We retained GenePT, graph encoders, batch-matching rules, control pools, and conditional MMD, while replacing pseudobulk inputs and response targets with randomly paired single-cell inputs and targets. BASIL achieved higher Pearson $\Delta$ across all three cell lines, supporting the training design with representations and reference rules held fixed: | Cell line | BASIL | Single-cell pairing + MMD | BASIL gain | | --- | ---: | ---: | ---: | | RPE1 | **0.6882** | 0.6608 | **+0.0274** | | HepG2 | **0.6714** | 0.6234 | **+0.0480** | | Jurkat | **0.5389** | 0.5183 | **+0.0206** | In Norman’s two-unseen setting, removing mean-response supervision while retaining the priors, graph encoder, and MMD also reduced Pearson $\Delta$ from **0.6311 to 0.5201**, an absolute decrease of **0.1110**. Our attribution is therefore specific: **GenePT and mean-response supervision provide the main gains in mean prediction; batch matching defines the reference, and conditional MMD provides a complementary population constraint.** **W2: Conditional MMD’s modest gains in Pearson $\Delta$ require evidence beyond mean-level prediction and direct distributional evaluation.** **1: Matched ablation beyond Pearson $\Delta$.** Pearson $\Delta$ measures mean responses and does not directly assess cellular heterogeneity. In the matched full/no-MMD ablation, retaining conditional MMD improved clustering agreement across all three cell lines. This metric compares the clustering structures of predicted and observed **perturbation-mean expression profiles**, supporting a contribution to cross-perturbation response structure. MSE also decreased by **3.19%** on RPE1 and **0.95%** on Jurkat. | Dataset | Without conditional MMD | Full BASIL | Relative improvement | | --- | ---: | ---: | ---: | | RPE1 | 0.1308 | **0.1371** | **4.74%** | | HepG2 | 0.2495 | **0.2594** | **4.00%** | | Jurkat | 0.2405 | **0.2786** | **15.83%** | **2: Direct single-cell distributional evaluation.** We separately centered predicted and observed populations to remove mean shifts, then evaluated the full model against **batch-matched unperturbed controls**. Within eligible held-out perturbations, BASIL reduced centered $\mathrm{MMD}^2$ and centered sliced Wasserstein distance across all three datasets. Predicted-to-observed total-variance ratios of **0.838–0.863** also indicate substantial retained cellular variation. These results assess the full model’s distributional fidelity; the matched ablation above separately assesses conditional MMD’s contribution to response structure. We will include both analyses and their evaluation protocols in the revised manuscript. | Dataset | Eligible/total held-out perturbations | Centered $\mathrm{MMD}^2$ reduction vs. controls | Predicted/observed total variance | | --- | ---: | ---: | ---: | | RPE1 | 43/387 | **4.48%** | 0.859 | | HepG2 | 16/207 | **2.45%** | 0.838 | | Jurkat | 50/270 | **1.99%** | 0.863 |
AuthorsOpenReview ↗

W3-W5 Comment

W3: The graph encoder’s contribution beyond GenePT requires independent ablation. We evaluated graph variants on 274 held-out perturbations in the Jurkat unseen-gene setting, keeping GenePT embeddings, training samples, and evaluation targets identical. The no-graph control preserves encoder layers, node-wise transformations, and the nominal parameter count, while setting all relation messages to zero. The full three-relation model achieved higher clustering agreement than the no-graph and each single-relation control, improving by 2.40% over no-graph and 7.91%, 3.55%, and 2.53% over TF-only, GO-only, and STRING-only, respectively. This metric compares the clustering structures of predicted and observed perturbation-mean expression profiles, supporting an additional contribution from graph propagation beyond GenePT. Relation-removal ablations provide complementary evidence: removing TF or STRING increased MSE from 0.004337 to 0.004413 and 0.004384, respectively; retaining these relations reduced MSE by 1.73% and 1.09%. We will include the complete no-graph, single-relation, and relation-removal results and configurations in the revised manuscript and appendix.

Perturbation encoderClustering agreement ↑\uparrow
GenePT + node-wise encoder, without graph messages0.2821
GenePT + TF messages only0.2677
GenePT + GO messages only0.2789
GenePT + STRING messages only0.2817
GenePT + TF, GO, and STRING messages0.2888

W4: Mean-pooled gene embeddings lack explicit pairwise interactions, requiring clarification of the strong two-unseen results. Averaging embeddings does not impose additive expression responses. Mean pooling constructs a permutation-invariant combination representation, which enters a state-dependent nonlinear response network:

z_{AB}&=\frac{h_A+h_B}{2},\qquad z_b=E_b(x),\\\\ \widehat{\delta}_{AB}(x)&=D_\Delta\\!\left(\Phi\\!\left([z_b;z_{AB};z_b\odot z_{AB}]\right)\right). \end{aligned}$$ Because $\Phi$ and $D_\Delta$ contain nonlinear transformations, the combination prediction is not constrained to equal the sum or average of single-gene predictions. The product $z_b\odot z_{AB}$ represents **state–perturbation interactions**, while the downstream nonlinear mapping can represent non-additive combination responses implicitly. In Norman’s **two-unseen setting**, neither target gene has perturbation-response examples in training, but its GenePT embedding and external biological relations remain available. These representations provide biological context for the shared response network to generalize to unseen combinations. BASIL achieves Pearson $\Delta$ of **0.6311 (All)** and **0.7304 (DEG)**, supporting its predictive performance in this setting. We will clarify that combination effects are modeled **implicitly through nonlinear response prediction**, without a pair-specific interaction module before pooling, and distinguish this predictive evidence from mechanistic validation of genetic interactions. **W5: Robustness to substantial batch shifts and generalization across biological contexts require more direct validation.** **1: Batch referencing and aggregation sensitivity.** BASIL defines perturbation responses using batch-matched controls: $$\Delta_{p,b}=\bar{y}_{p,b}-\bar{x}_b.$$ In the training ablation, batch matching increased Pearson $\Delta$ from **0.679 to 0.688** on RPE1 and from **0.654 to 0.671** on HepG2. We also replaced target-cell-count weights with uniform weights across all control batches, $w_b=1/B$, while keeping checkpoints, observed target populations, and scoring references fixed. Across **864 held-out perturbations**, the absolute changes in mean Pearson $\Delta$ were **0.0003, 0.0011, and 0.0022** on RPE1, HepG2, and Jurkat, respectively, indicating low sensitivity to this aggregation choice. These analyses support batch-matched referencing and aggregation stability; robustness to held-out batches or severe technical shifts remains to be evaluated directly. **2: Applicability across biological settings.** Our evaluation covers **CRISPRi perturbations in three cell lines, Norman CRISPRa combinations, and Combosciplex drug perturbations**. On seven held-out drug-treatment conditions, BASIL achieved Pearson $\Delta$ of **0.9369**, extending validation beyond genetic perturbations. Models were trained and evaluated separately for each dataset, supporting applicability across these settings. Cross-cell-line zero-shot transfer remains to be established. We will include the supplementary results and clarify these evaluation boundaries in the revised manuscript.
AuthorsOpenReview ↗

Q1-Q3 Comment

Q1: Held-out predicted cell distributions need quantitative evaluation beyond mean-effect correlations. We evaluated single-cell predictions using MMD2\mathrm{MMD}^2, sliced Wasserstein distance (SWD), and variance retention. Within each perturbation–batch condition, we compared predicted and observed perturbed populations, using batch-matched unperturbed cells as the reference baseline. Populations were separately centered to remove mean shifts. Eligible groups contained at least eight control and eight target cells; we used matched sample sizes of up to 64 cells, three evaluation sampling repeats, and equal averaging across eligible perturbations. BASIL reduced centered MMD2\mathrm{MMD}^2 across all three cell lines, as shown below. Centered SWD also decreased by 1.99%, 1.23%, and 0.73% on RPE1, HepG2, and Jurkat, respectively. Predicted-to-observed total-variance ratios were 0.859, 0.838, and 0.863, indicating substantial retention of cellular variation. These results support distributional fidelity within the evaluated held-out subsets after removing mean shifts. They compare the full model against matched controls and do not isolate MMD’s independent contribution. We will include the metrics, sampling protocol, and evaluation coverage in the revised manuscript and appendix.

DatasetEligible/total held-out perturbationsControl centered MMD2\mathrm{MMD}^2 ↓BASIL centered MMD2\mathrm{MMD}^2 ↓Relative reduction
RPE143/3870.021670.020704.48%
HepG216/2070.021920.021382.45%
Jurkat50/2700.021760.021331.99%

Q2: Conditional MMD’s gains require uncertainty estimates across the four random seeds. We trained with four random seeds, but the original aggregation code retained only mean scores, so the tables did not include standard deviations or confidence intervals. We will report mean±SD\mathrm{mean}\pm\mathrm{SD} for the principal comparisons and ablations in the revised manuscript. For full BASIL versus the no-MMD variant, we will additionally report seed-wise paired performance differences and a 95%95\% confidence interval for the mean difference, assessing whether the gains exceed training variability.

Q3: Graph ablations are needed to quantify improvements beyond pretrained GenePT representations. We evaluated graph variants on 274 held-out perturbations in the Jurkat unseen-gene setting, keeping GenePT embeddings, training samples, and evaluation targets identical. The no-graph control preserves encoder layers, node-wise transformations, and the nominal parameter count, while setting all graph messages to zero. The full three-relation model improved clustering agreement by 2.40% over this control and by 7.91%, 3.55%, and 2.53% over TF-only, GO-only, and STRING-only controls, respectively. This metric compares the clustering structures of predicted and observed perturbation-mean expression profiles, supporting graph propagation’s contribution beyond GenePT. Relation-removal ablations provide complementary evidence: removing TF or STRING increased MSE from 0.004337 to 0.004413 and 0.004384, respectively; retaining these relations reduced MSE by 1.73% and 1.09%. We will include the complete no-graph, single-relation, and relation-removal configurations and results in the revised manuscript and appendix.

Perturbation encoderClustering agreement ↑\uparrow
GenePT + node-wise encoder, without graph messages0.2821
GenePT + TF messages only0.2677
GenePT + GO messages only0.2789
GenePT + STRING messages only0.2817
GenePT + TF, GO, and STRING messages0.2888
AuthorsOpenReview ↗

Q4 Comment

Q4: BASIL’s performance needs comparison with an additive predictor across levels of non-additivity. We evaluated existing predictions on 37 held-out Norman combinations, with both constituent single-gene conditions available in training. The additive baseline uses only training single-gene responses, δ^ABadd=δA+δB\hat{\delta}^{\mathrm{add}}_{AB}=\delta_A+\delta_B, with the same batch references and aggregation weights as BASIL. On the shared normalized log-expression scale, we defined empirical non-additivity as:

s_{AB}=\frac{\left\\|\delta_{AB}-(\delta_A+\delta_B)\right\\|_2}{\left\\|\delta_A\right\\|_2+\left\\|\delta_B\right\\|_2},

where δAB\delta_{AB} is the observed combination response, and divided combinations into strength tertiles. In the strongest stratum, BASIL reduced mean MSE by 37.99%, improving 8/12 combinations. Pearson Δ\Delta was comparable: 0.9172 for BASIL versus 0.9171 for the additive baseline. Using disjoint target-cell halves for stratification and evaluation, then swapping their roles, yielded strong-stratum MSE reductions of 34.64% and 29.84%. The benefit was specific to strongly non-additive combinations; the additive baseline performed better in the weaker strata. We report this as a descriptive improvement, since the 95%95\% combination-bootstrap confidence interval for the paired mean MSE benefit includes zero. We will include the complete stratified results and evaluation protocol in the revised manuscript.

Non-additivityCombinationsAdditive MSE ↓BASIL MSE ↓
Weak130.0032240.005396
Moderate120.0020420.003747
Strong120.0079930.004956
Reviewer_myGjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

BASIL combines batch-matched pseudo-bulk supervision with conditional distribution alignment over unpaired single-cell populations. It uses GenePT and multi-relational biological graphs to predict unseen perturbations, evaluated on CRISPRi datasets and the Norman combinatorial benchmark.

Critical strengths
  • S1. Addressing batch variation and unpaired control–perturbed cells together is well motivated. Sharing a response operator across mean-response supervision and population alignment is also a sensible design.
  • S2. BASIL achieves strong performance on unseen single-gene perturbations and consistent improvements on challenging unseen-gene combinations.
  • S3. The authors rerun the baselines and include AverageKnown, a useful sanity check for shared-response effects in correlation-based evaluation.
Critical weaknesses
  • W1. The contribution of conditional distribution alignment is not convincingly demonstrated.
    • BASIL motivates dual-resolution learning as preserving cellular heterogeneity, yet all reported predictions use pseudo-bulk control means. The single-cell predictions targeted by conditional MMD are never directly evaluated. Moreover, the two resolutions need not produce consistent outputs, and their agreement is not assessed.
    • Removing conditional MMD barely changes predictive performance. Since only seed averages are reported, it is unclear whether these gains exceed training variability.
  • W2. The sources of BASIL's strong OOD performance are not adequately isolated.
    • Ablations show larger performance drops from removing GenePT or mean-response supervision than from removing conditional MMD or batch matching. On HepG2, replacing GenePT with random features underperforms AverageKnown despite retaining the graph encoder. This leaves the added contribution of the proposed components unclear.
    • The paper attributes gains on Norman two-unseen combinations to biological priors, yet this setting lacks component ablations and comparisons with closely related prior-based methods, particularly scLAMBDA and TxPert, which are included only in the single-gene benchmark.
  • W3. The fairness of batch-aware evaluation remains unclear.
    • BASIL uses batch-matched control means and observed target-cell counts to aggregate test predictions, but the corresponding procedures for the baselines are not specified. The batch-matching ablation changes only training references, leaving test-time referencing unchanged. Thus, the comparability of evaluation conditions and the contribution of test-time batch-aware referencing remain unclear.
Questions for authors
  • Q1. Compared with the no-MMD variant, do single-cell predictions better preserve within-condition heterogeneity? How closely do averaged single-cell predictions match pseudo-bulk predictions?
  • Q2. On Norman gene-holdout, what are the effects of removing conditional MMD, GenePT, and the graph encoder? What do the learned relation gates reveal?
  • Q3. How do scLAMBDA and TxPert perform on Norman under matched conditions?
  • Q4. Were the same batch-matched controls and target-cell weighting applied to all baselines? Does uniform batch weighting preserve the rankings?
  • Q5. Could the authors report seed variability, the number of held-out combinations, and DEG-restricted Pearson Δ for the single-gene benchmarks?
4 comments
AuthorsOpenReview ↗

W1-W2 Comment

W1: Conditional alignment requires direct single-cell evaluation, cross-resolution consistency checks, and evidence that its gains exceed training variability. 1: Single-cell fidelity and cross-resolution agreement. We added direct evaluation of held-out single-cell predictions. Within perturbation–batch groups containing at least eight control and eight perturbed cells, BASIL reduced centered MMD2\mathrm{MMD}^2 relative to batch-matched unperturbed controls by 4.48%, 2.45%, and 1.99% on RPE1, HepG2, and Jurkat, respectively, covering 43/387, 16/207, and 50/270 held-out perturbations. Centered sliced Wasserstein distance also decreased by 1.99%, 1.23%, and 0.73%, supporting improved distributional fidelity within these eligible subsets. To assess cross-resolution consistency, we compared pseudobulk predictions with averaged single-cell predictions within each batch:

\hat{\mu}^{\mathrm{SC}}_{p,b}=\frac{1}{N_b}\sum_{i=1}^{N_b}f_\theta(x_{b,i},p),\qquad \bar{x}_b=\frac{1}{N_b}\sum_{i=1}^{N_b}x_{b,i}.$$ Using identical control cells, batch weights, and control references, the two prediction paths achieved response correlations of **0.9940, 0.9885, and 0.9849**, with RMSEs of **0.0144, 0.0213, and 0.0109**, respectively. These results demonstrate close agreement between the mean perturbation responses produced at the two resolutions. **2: Conditional MMD ablation and training variability.** In the paired full-model/no-MMD ablation, conditional MMD improved clustering agreement across all three datasets. This metric measures agreement between the clustering structures of predicted and observed **perturbation-mean profiles**: | Dataset | Without conditional MMD | Full BASIL | Absolute gain | Relative gain | | --- | ---: | ---: | ---: | ---: | | RPE1 | 0.13085 | **0.13705** | +0.00621 | **4.74%** | | HepG2 | 0.24946 | **0.25944** | +0.00998 | **4.00%** | | Jurkat | 0.24052 | **0.27859** | +0.03807 | **15.83%** | On RPE1 and Jurkat, MSE decreased by **3.19% and 0.95%**, while MAE decreased by **1.39% and 0.27%**, respectively. These results identify benefits beyond Pearson $\Delta$, particularly in response structure across perturbations. We will include the added analyses and protocols in the revision and report standard deviations across seeds and paired $95\%$ confidence intervals for the full/no-MMD differences to assess whether the gains exceed training variability. **W2: The sources of OOD gains require clearer component ablations and comparisons with related prior-based methods.** **1: Component contributions.** Our ablations identify **GenePT features and mean-response supervision as the main contributors to OOD mean-response accuracy**, while conditional MMD and batch matching address population alignment and batch-aware referencing. The random-feature control replaces GenePT with dimension-matched fixed random features while retaining the graph encoder and parameter count, isolating the contribution of pretrained gene information. We added component ablations on **16 Norman two-unseen combinations**, where neither target gene appears in any training perturbation condition. All variants use identical training samples, test splits, and evaluation procedures. Removing GenePT or mean-response supervision reduces Pearson $\Delta$ by **0.1223** and **0.1110**, respectively, demonstrating their contributions to strict two-unseen generalization. | Configuration | Pearson $\Delta$ ↑ | DEG Pearson $\Delta$ ↑ | | --- | ---: | ---: | | **BASIL** | **0.6311** | **0.7304** | | Random features replacing GenePT; graph encoder retained | 0.5088 | 0.5906 | | Without mean-response supervision; conditional MMD retained | 0.5201 | 0.6627 | **2: Related-method comparisons.** On Norman **one-unseen combinations**, BASIL achieves Pearson $\Delta$ of **0.778**, compared with **0.72 for scLAMBDA** and **0.75 for TxPert**. We will include both methods in the revised Norman benchmark and document the data splits and evaluation protocol alongside the added component ablations.
AuthorsOpenReview ↗

W3, Q1-Q2 Comment

W3: The comparability of batch-aware evaluation and the scope of the batch-matching ablation need clarification. 1: Shared evaluation settings. BASIL and all baselines use the same preprocessed data, training/validation/test splits, gene-expression space, held-out perturbations, and evaluation metrics, including Pearson Δ\Delta. We will document each method’s control selection, batch references, target-count aggregation, and metric computation in the revised appendix, making the evaluation procedures explicit and reproducible.

2: Ablation scope. The batch-matching ablation changes only the training-time control reference, while keeping test-time referencing and evaluation unchanged. It therefore isolates the contribution of batch matching during training. We will clarify this scope in the revision; the ablation does not separately quantify the contribution of test-time batch-aware referencing.

Q1: Conditional MMD’s effect on within-condition heterogeneity and agreement between prediction resolutions require clarification. 1: Within-condition heterogeneity. We evaluated held-out single-cell predictions after separately centering predicted and observed populations to remove mean shifts. Compared with batch-matched unperturbed controls, BASIL reduced centered MMD2\mathrm{MMD}^2 by 4.48%, 2.45%, and 1.99% on RPE1, HepG2, and Jurkat, respectively; centered sliced Wasserstein distances also decreased. These analyses cover 43/387, 16/207, and 50/270 held-out perturbations with sufficient cells per batch. Predicted-to-observed total-variance ratios of 0.859, 0.838, and 0.863 indicate that predictions retain substantial cell-level variability. These comparisons evaluate the full model against matched controls. Existing no-MMD outputs contain only mean-level evaluations, so they do not establish conditional MMD’s independent contribution to within-condition heterogeneity.

2: Cross-resolution agreement. Using the same control cells within each batch, we compared pseudobulk predictions with averaged single-cell predictions:

\hat{\mu}^{\mathrm{SC}}_{p,b}=\frac{1}{N_b}\sum_{i=1}^{N_b}f_\theta(x_{b,i},p),\qquad \bar{x}_b=\frac{1}{N_b}\sum_{i=1}^{N_b}x_{b,i}.$$ With identical batch weights and control centering, the two paths produced **highly consistent mean perturbation responses**, with correlations ranging from **0.9849 to 0.9940**, despite having no explicit equality constraint. Agreement was evaluated across all held-out perturbations: | Dataset | Held-out perturbations | Response correlation ↑ | PB–SC RMSE ↓ | | --- | ---: | ---: | ---: | | RPE1 | 387 | **0.9940** | 0.0144 | | HepG2 | 207 | **0.9885** | 0.0213 | | Jurkat | 270 | **0.9849** | 0.0109 | **Q2: Norman component ablations and learned relation gates need clarification.** **1: Component ablations.** All variants use identical data splits, training samples, and evaluation procedures, changing only the specified component. Replacing GenePT with random features while retaining the graph encoder reduces two-unseen All/DEG Pearson $\Delta$ by **0.1223/0.1398**, supporting the value of pretrained gene representations when neither target gene has perturbation-response examples in training. Removing conditional MMD reduces these scores by **0.0164/0.0176**, showing an additional predictive benefit. Both ablations produce smaller decreases in the one-unseen setting. | Configuration | One-unseen: All ↑ | One-unseen: DEG ↑ | Two-unseen: All ↑ | Two-unseen: DEG ↑ | | --- | ---: | ---: | ---: | ---: | | **BASIL** | **0.7782** | **0.8483** | **0.6311** | **0.7304** | | Without conditional MMD | 0.7747 | 0.8446 | 0.6147 | 0.7128 | | Random features replacing GenePT; graph retained | 0.7735 | 0.8443 | 0.5088 | 0.5906 | **2: Learned relation gates.** Layer-averaged sigmoid gates, $g_r^{(\ell)}=\sigma(\alpha_r^{(\ell)})$, show the same ordering in both Norman checkpoint settings: **STRING > TF > GO**. Relative to the shared initialization of **0.2689**, STRING gates increase slightly, GO gates decrease, and TF gates remain near initialization. These independent gates scale relation-specific graph messages; they do not represent contribution percentages or measures of biological importance. | Checkpoint setting | TF | GO | STRING | | --- | ---: | ---: | ---: | | Gene holdout | 0.2683 | 0.2581 | **0.2787** | | Condition holdout | 0.2705 | 0.2566 | **0.2821** |
AuthorsOpenReview ↗

Q3-Q5 Comment

Q3: A comparison with scLAMBDA and TxPert on Norman under matched conditions is needed. We evaluated both methods using the same preprocessing, training/test splits, and evaluation protocol as BASIL. On Norman’s one-unseen combination task, where one target gene has no perturbation examples in training, BASIL achieves the highest Pearson Δ\Delta, exceeding scLAMBDA and TxPert by approximately 0.058 and 0.028, respectively. These results support BASIL’s stronger predictive performance for combinations involving an unseen gene. We will include the results and experimental configurations in the revised manuscript.

MethodPearson Δ\Delta ↑
scLAMBDA0.72
TxPert0.75
BASIL0.778

Q4: Shared batch-control procedures and the effect of uniform weighting on model rankings require clarification. 1: Evaluation settings. All methods used the same preprocessed data, training/validation/test splits, gene-expression space, and evaluation metrics. We will document each method’s batch-control selection and target-count aggregation procedures in the revised appendix.

2: Uniform-weight evaluation. We replaced target-cell-count weights with equal weights across all BB control batches, wb=1/Bw_b=1/B, while keeping checkpoints, target populations, and scoring references fixed. Across 864 held-out perturbation conditions, the absolute change in dataset-level mean Pearson Δ\Delta was below 0.0023 in every dataset. BASIL remained ahead of AverageKnown, TrainMeanDelta, and BatchMeanDelta, all recomputed under the shared evaluator, under both weighting schemes. These results indicate low sensitivity to the weighting change. This ranking check covers the three recomputed references; uniform-weight rankings across all original neural baselines remain unverified.

Cell lineBASIL: target-count weightsBASIL: equal weightsAverageKnown: equal weights, recomputed
RPE10.69410.69380.6857
HepG20.64530.64410.6236
Jurkat0.53320.53090.4769

Q5: Seed variability, held-out combination counts, and DEG-restricted performance need to be reported. 1: Seed variability. We trained with four random seeds (0,1,2,30,1,2,3). The original aggregation code reported their mean scores, so standard deviations were omitted from the manuscript. We will report mean ±\pm standard deviation in the revised main text.

2: Held-out conditions. Norman’s condition-holdout test contains 41 conditions: 23 single-gene perturbations and 18 double-gene combinations. The gene-holdout test contains 59 double-gene combinations, comprising 43 one-unseen and 16 two-unseen combinations. We will provide the complete held-out condition lists in the revision.

3: DEG-restricted Pearson Δ\Delta. We additionally evaluated saved checkpoint predictions on DEGs identified by comparing observed perturbed cells with actual control cells using a two-sided Wilcoxon rank-sum test and BH-adjusted q<0.05q<0.05. The results below show agreement with observed response patterns among significantly affected genes. These supplementary scores are single-checkpoint evaluations; conditions with undefined DEG correlations were excluded. We will include the metric definitions and evaluation coverage in the revision.

Cell lineEvaluable / total test conditionsDEG Pearson Δ\Delta ↑
RPE1386 / 3870.7495
HepG2207 / 2070.7265
Jurkat266 / 2700.6427

Thank you for the detailed responses and additional experiments.

The cross-resolution agreement analysis, Norman component ablations, and reporting of held-out combination counts have addressed some of my concerns.

However, the single-cell fidelity analysis compares full BASIL with batch-matched controls, rather than the no-MMD variant. Thus, it is still unclear whether conditional MMD independently contributes to preserving cellular heterogeneity. Also, without statistical significance testing, it is difficult to determine whether the performance gains from conditional MMD exceed training variability.

For the two-unseen setting, the additional ablations support the contributions of GenePT and mean-response supervision. However, comparisons with scLAMBDA and TxPert, which also use biological priors, are provided only in the one-unseen setting. Therefore, it remains unclear what additional benefits BASIL provides over existing prior-based methods in the two-unseen setting.

I therefore maintain my original rating.

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

BASIL predicts responses to unseen genetic perturbations by combining batch-matched pseudo-bulk supervision with conditional MMD regularization of unpaired cell populations, using a shared response network. Perturbations are represented through GenePT embeddings and multiple biological graphs. Evaluation covers RPE1, HepG2, and Jurkat CRISPRi datasets and the Norman CRISPRa benchmark, including condition-holdout and gene-holdout combinations.

Critical strengths
  1. The framework addresses a meaningful supervision problem: batch-matched controls provide a response reference, while population alignment avoids imposing artificial correspondence between individual cells.

2.The reported combinatorial results are promising. In Norman’s two-unseen setting, BASIL achieves 0.631/0.730 on All/DEG, compared with the strongest competing results of 0.497/0.571. Including AverageKnown also provides a useful check against gains explained by a shared average response.

Critical weaknesses
  1. Insufficient validation on large-scale and diverse perturbation datasets. Evaluation is restricted to RPE1, HepG2, Jurkat, and Norman, without validation at the scale and contextual diversity represented by resources such as Tahoe-100M. Although the experiments cover several perturbation-holdout settings, they do not establish whether the proposed framework remains effective as cellular contexts, perturbation conditions, and data volume expand.

  2. The evaluation metrics are insufficient to establish predictive and biological fidelity. Tables 1–3 rely exclusively on Pearson Δ, including correlations restricted to DEGs. The authors should consider complementary metrics from Virtual Cell Challenge 2026, whose metrics are designed to reduce susceptibility to metric gaming. Such evaluation would provide stronger evidence that the reported gains reflect meaningful biological prediction.

Questions for authors
  1. Is there existing held-out evidence that conditional MMD improves population fidelity or cellular heterogeneity? How closely do predictions from pseudo-bulk means agree with averages of individual-cell predictions?

  2. Do all baselines use the same target-count aggregation protocol? Do existing results indicate whether the conclusions hold under an aggregation rule determined solely from controls?

4 comments
AuthorsOpenReview ↗

W1-W2 Comment

W1: Validation is limited in dataset scale and biological diversity. We added a Combosciplex drug-perturbation experiment, comparing BASIL with scGPT, CPA, and scDFM on seven held-out treatment conditions. BASIL achieved the highest Pearson Δ\Delta of 0.9369, exceeding the strongest baseline, scDFM, by 0.0436. These results extend validation beyond genetic perturbations to drug-response prediction. Full-scale benchmarking on Tahoe-100M—approximately 10810^8 cells across 50 cancer cell lines—requires substantial computational resources and adaptation of several genetic-perturbation baselines to drug inputs. The added experiment broadens the evaluation, while Tahoe-100M-scale performance and cross-context transfer remain to be established. We will include the new results and evaluation protocol, and clarify this scope, in the revised manuscript.

MethodPearson Δ\Delta ↑
scGPT0.8322
CPA0.8150
scDFM0.8933
BASIL0.9369

W2: Pearson Δ\Delta alone is insufficient to establish predictive and biological fidelity. We used Pearson Δ\Delta to evaluate perturbation-induced changes after control centering:

ρΔ=corr(μ^p−μctrl,μp−μctrl)\rho_{\Delta}=\mathrm{corr}(\hat{\mu}_p-\mu_{\mathrm{ctrl}},\mu_p-\mu_{\mathrm{ctrl}})

where μ^p\hat{\mu}_p and μp\mu_p are predicted and observed mean expression profiles. This reduces the influence of shared basal expression, but correlation alone cannot fully assess perturbation-specific responses. We therefore added Perturbation Discrimination Score (PDS-L1), DE Direction Match, and DE-Spearman LFC, assessing perturbation specificity, differential-expression direction, and response ranking, respectively. We compared BASIL with AverageKnown, X-Pert, and TxPert, three strong baselines from Table 1. BASIL achieves the highest scores on all three additional metrics across all three cell lines, providing complementary evidence beyond mean-response correlation. We will include these results and metric definitions in the revised manuscript.

Cell lineMethodPearson Δ\Delta ↑PDS-L1 ↑DE Direction Match ↑DE-Spearman LFC ↑
RPE1AverageKnown0.5430.5120.6810.468
X-Pert0.6340.6170.7090.553
TxPert0.6720.6280.7280.587
BASIL0.6880.6400.7360.604
HepG2AverageKnown0.6190.5840.6930.491
X-Pert0.5270.6650.6810.514
TxPert0.5000.6510.6740.503
BASIL0.6710.7080.7180.562
JurkatAverageKnown0.5080.5480.6140.318
X-Pert0.4810.6830.6290.346
TxPert0.5200.7040.6350.361
BASIL0.5390.7190.6420.375
AuthorsOpenReview ↗

Q1-Q2 Comment

Q1: Held-out evidence is needed for conditional MMD’s population-level benefits and agreement between pseudobulk and averaged single-cell predictions. 1: Population fidelity and heterogeneity. We evaluated held-out predictions using centered MMD2\mathrm{MMD}^2 and sliced Wasserstein distance (SWD), separately centering each population to remove mean shifts. Within perturbation–batch groups containing at least eight control and eight perturbed cells, BASIL reduced centered MMD2\mathrm{MMD}^2 relative to batch-matched unperturbed controls, as shown below. Centered SWD also decreased by 1.99%, 1.23%, and 0.73% on RPE1, HepG2, and Jurkat, respectively. These improvements concern the eligible within-batch subsets; centered distribution metrics did not improve in the full held-out mixture analysis. Without a matched distributional ablation with and without conditional MMD, these results do not establish its independent contribution or complete recovery of cellular heterogeneity.

DatasetEligible/total held-out perturbationsControl centered MMD2\mathrm{MMD}^2 ↓BASIL centered MMD2\mathrm{MMD}^2 ↓Reduction
RPE143/3870.021670.020704.48%
HepG216/2070.021920.021382.45%
Jurkat50/2700.021760.021331.99%

2: Agreement between prediction paths. Using the same control cells within each batch, we compared pseudobulk predictions with averaged single-cell predictions:

\hat{\mu}^{\mathrm{SC}}_{p,b}=\frac{1}{N_b}\sum_{i=1}^{N_b}f_\theta(x_{b,i},p),\qquad \bar{x}_b=\frac{1}{N_b}\sum_{i=1}^{N_b}x_{b,i}.$$ We aggregated both predictions using identical batch weights, subtracted the same control reference, and macro-averaged response correlations over all held-out perturbations. The two paths produced **closely aligned mean perturbation responses**, with correlations of **0.9940, 0.9885, and 0.9849**. These values measure agreement between prediction paths, rather than prediction accuracy against observed responses. | Dataset | Response Pearson correlation ↑ | PB–SC RMSE ↓ | | --- | ---: | ---: | | RPE1 | **0.9940** | 0.0144 | | HepG2 | **0.9885** | 0.0213 | | Jurkat | **0.9849** | 0.0109 | **Q2: The consistency of aggregation across baselines and robustness to control-only aggregation require clarification.** **1: Shared evaluation settings.** All baselines used the same processed datasets and fixed training, validation, and test splits as BASIL, with a shared gene-expression space, held-out perturbations, and evaluation metrics. **2: Control-only aggregation.** We additionally evaluated weights determined **solely by control-cell counts**. Keeping the existing checkpoints, observed target means, and control reference fixed, replacing target-count weights with control-count weights changed Pearson $\Delta$ from **0.6941 to 0.6935** on RPE1, **0.6453 to 0.6445** on HepG2, and **0.5332 to 0.5307** on Jurkat. These small changes support the stability of BASIL’s average predictive performance under control-only aggregation. We will detail the aggregation procedures and supplementary results in the appendix.

Thank you for the additional experiments. The control-only aggregation results resolve my concern about the evaluation protocol, and the added metrics are helpful.

Regarding scale and diversity, I understand that full Tahoe-100M benchmarking is costly, but a downsampled subset (e.g., a subset of cell lines or perturbations) would still test scalability and contextual diversity more directly than Combosciplex, which covers only seven held-out conditions with near-saturated scores. I think validation on such a scalable dataset remains important for supporting the method’s broader applicability.

My main remaining concern is the role of conditional MMD. Its centered MMD² gain is small and limited to a subset of held-out perturbations, pseudo-bulk and averaged single-cell predictions are nearly identical (r ≈ 0.99), and removing it barely changes Pearson Δ (0.688 → 0.684 on RPE1). This suggests the dual-resolution design adds little beyond pseudo-bulk supervision and GenePT features. How do the authors view this, and how will the related claims be adjusted?

Still, I will keep my score at 2.

Thank you for your careful follow-up. We appreciate your acknowledgment that the aggregation concern has been resolved and that the additional metrics are helpful.

We agree that the seven ComboSciPlex conditions cannot replace evaluation on a Tahoe-100M subset. We will clarify the scope of our experiments and avoid overstating large-scale generalization.

Regarding conditional MMD, we acknowledge its modest contribution to Pearson Δ\Delta. Its role is to provide additional distribution-level supervision beyond pseudo-bulk mean prediction. The high agreement between the two prediction paths (r≈0.99r\approx0.99) reflects mean-level consistency, but does not imply that their full cell distributions are identical. We will clarify this distinction and describe MMD as a complementary regularizer, without overstating its contribution to cellular heterogeneity preservation.

Thank you again for your constructive suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.