acceptodds
Withdrawn from ICLR 2027

DACL: Predicting Responses to Pairs of Unseen Genes by Reusing Single-Perturbation Supervision

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Combinatorial CRISPR screens measure a small fraction of possible gene pairs, so models must predict responses to gene pairs never perturbed in training. In this strict both-unseen setting, the only information about the test genes is a response-free descriptor such as Gene Ontology annotations. Learned structure helps only if these descriptors can retrieve it. We introduce Deployment-Addressable Constituent Learning (DACL), which learns one map from a gene's descriptor to its perturbation effect. Single perturbations supervise the map directly, while pairs supervise the sum of two predicted effects. An unseen pair is predicted by that sum. On the three official Norman Perturb-seq splits, DACL reaches a Pearson- of 0.401 on pairs of unseen genes, against 0.337 for biolord and 0.191 for GEARS. Over 70 further splits it outperforms biolord and direct pair models given the same Gene Ontology descriptors, and the advantage stays positive, though unresolved at the gene level, with language-model gene embeddings. On a genetic-interaction screen, matched controls show that single perturbations help through the additive path, not a larger function class. Direct pair models remain preferable on some screens with strong pair-specific structure. In two prospective evaluations, training data alone always selected the direct model, which had lower normalized pair error in all 45 panels.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
perturbation response prediction, combinatorial CRISPR screens, Perturb-seq, single-cell transcriptomics, compositional generalization, Gene Ontology, genetic interactions, kernel ridge regression

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

The paper presents DACL, a kernel ridge-regression approach to predict responses to pairs of unseen perturbations, using known gene descriptors. The paper compares performance to naive linear baselines as well as previously published approaches that support the task. The Norman dataset is reported as the primary benchmark evaluation setting. Six additional datasets are used to study different performance aspects. The authors conduct extensive evaluations demonstrating comparative or slight gain in most cases.

Critical strengths

The paper targets a less studied case, double-unseen perturbation. It backs the modeling approach with mathematical considerations/explanations, and reports extensive evaluations and ablations over different datasets.

Critical weaknesses

Overall performance and Contribution.

(performance) Although the authors present extensive evaluations, DACL performance is often marginally better than the direct pair-response baselines or compared baselines.

(contribution) Therefore, considering that the success of linear models in perturbation modeling, was already reported, and acknowledged by the authors (citing Ahlmann-Eltze et al., 2025; which in general represents a line of works that reported this), and that the DACL approach is only marginally better than the simple controls with same information, it is unclear what is the contribution of this work. This is well captured by the authors in the four line conclusion paragraph of the paper.

Questions for authors

Can the authors clarify what they see as the contribution of this work ? it will be valuable to address both:
a. the importance of the work to the single-cell perturbation modeling community? b. the relevance for the broader ML community and fit to ICLR.

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

The paper studies strict both-unseen prediction: predicting the response to a pair of genes when neither gene has appeared as an intervention target in training. So the only test-time information is a response-free descriptor (mainly a GO membership vector). The proposed method DACL is kernel ridge regression from the descriptor to gene’s perturbation effect prediction.

Two DACL variants are proposed that differs in the supervision. The DACL-S fits the mapping using single perturbations, while DACL-C fits pair responses through the sum of the two constituent predictions, with a fixed pair-block weight λ=1\lambda=1. And the method predicts condition-mean responses, rather than individual-cell response distributions.

On the three official Norman dataset splits, DACL-C reaches delta-pearson ~0.4 on SEEN-0 and on 70 further splits it reaches 0.294, outperforming the DirectJoint Ridge / MLP with the same descriptor. It also includes extensive control experiments.

Overall, I think this paper is a useful, carefully controlled study of a simple additive predictor, but the case for a substantial methodological advance is less convincing. My main concern is the limited/incremental novelty, the contribution of pair supervision on the main task, the interpretation of the experiment and the logic, and readability of the appendix.

Critical strengths

1. The prediction problem is motivated and distinct from much easier combination prediction settings

Predicting a pair whose two constituent perturbations are unseen is a real and useful setting. The paper carefully distinguishes it from predicting a new combination of genes whose individual effects have already been observed, and to my knowledge, most of the related works focus on the latter, easier combination settings.

I appreciate the careful removal of training perturbations involving held-out genes, the use of target-disjoint validation, and the reporting of SEEN-0 and SEEN-2. This distinction is important because success on combinations of observed singles does not necessarily indicate the success on pairs of entirely unseen intervention targets.

2. Simplicity and efficiency

Closed-form ridge can fit in 2 to 4 sec on CPU that beats GEARS, scGPT, Scouter and biolord on SEEN-0 Norman, which is a practical, cost-efficient contribution to this area, and it is also consistent with the recent benchmarks showing linear baselines are hard to beat.

Moreover, the paper also provides several control experiments that can rule out the potential alternative explanations, for instance, gene-to-descriptor assignments are permuted, neural head share the same encoder and training protocol, pair targets are compared with shuffled or synthetic additive targets.

Critical weaknesses

Due to the 1-2 critical weakness limit of this block, I moved few weaknesses to the Question section.

1. Originality and significance

This is my biggest concern. The paper identifies its addition over existing constituent approaches as pair-sum supervision of the same descriptor-to-effect map. This is a narrow modification, Eq. 4 is a weighted ridge regression with additional rows whose features are sums of constituent features.

The paper also states that DACL-S “combines two baselines of Ahlmann-Eltze” and that constituent prediction plus additive composition already exists in biolord and CPA. PerturbNet [1] paper also implements a learned linear single-perturbation predictor and predicts doubles by adding its two predicted single responses and subtracting the control.

GenePert [2] uses simple l2-regularized linear regression on pretrained GenePT embeddings that is designed to be permutation-invariant over the two genes of a double perturbation and to reuse single-perturbation information, so this is essentially DACL-S with a GenePT descriptor, but it is not cited or compared. scLAMBDA [3] also uses GenePT embeddings by default and targets multi-gene perturbation prediction

The claimed addition is pair-sum supervision of the shared map (DACL-C). But across screens the pair term is not decisive, for instance, DACL-C vs. DACL-S is +0.008 on Norman SEEN-0 and not significant (+0.006 [−0.011, 0.024]) at the gene level. It also lowers Pearson on GSE116198, GSE284018 Gamma (0.376 to 0.357) and Tau (0.314 to0.272), it is ~0 on MCF10A and GSE90063. The paper’s main methodological contribution therefore does not have consistent support.

2. Confounded comparison with biolord

Tab. G.1 mentions that biolord is trained on single perturbations only, whereas DACL-C uses both single and pair responses. So, the DACL-C versus biolord comparison differs in both model design and training supervision, making it difficult to attribute the performance gain to the proposed method itself.

Moreover, using pair observations to learn compositional perturbation effects has precedent in CPA, although its architecture is different from DACL. I suggest a more informative comparison would include a biolord variant trained with both singles and pairs under a comparable additive objective.

References:

[1] Yu, Hengshi, et al. "Perturbnet predicts single-cell responses to unseen chemical and genetic perturbations." Molecular systems biology 21.8 (2025): 960-982.

[2] Chen, Yiqun, and James Zou. "GenePert: Leveraging GenePT embeddings for gene perturbation prediction." bioRxiv (2024): 2024-10.

[3] Wang, Gefei, et al. "Modeling and predicting single-cell multi-gene perturbation responses with sclambda." bioRxiv (2024).

Questions for authors
  1. Evidence gap why pair-specific capacity helps

    TheTab. F.1 results show a large disadvantage for DACL-C under PRJNA691742 and Shen2017. The issue is the explanation of this gap.

    The main DACL formulation fixes the pair offset using twice the mean single effect. It has no independently fitted pair intercept. By comparison, the richer direct head described in Table A.1 has a count feature and a free bias, allowing more flexibility in the relationship between single and pair offsets.

    A direct model can consequently improve absolute error by correcting a systematic offset or amplitude mismatch, even without predicting useful variation specific to individual pairs.

  2. The appendix is piecemeal and hard to read or navigate, it includes excessive repeated information and very defensive scope explanations which is not necessary. For instance, the tables appear out of section order, Appendix C.1 is only a single sentence, Table A.8 names no metric and uses a different sign convention per row. Figure 1 tries to carry the setting, the theory and the model at once and is difficult to read.

  3. Also, I wonder what the intuition of introducing Eq. H.1?

Reviewer_Zzk8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces DACL, which is specialized in predicting responses when two genes have been perturbed that haven't been seen during training. In order to do so, they use kernel regression to map Gene Oncology (GO) descriptors to single-gene perturbation responses. These are then summed up to provide the predicted effect of perturbing two genes. The method is jointly trained on both single and double perturbations. DACL outperforms competing methods in the both-unseen generalization regime in the Norman Perturb-seq dataset. Furthermore, the authors investigate the role of gene descriptors and gene interactions across datasets using controlled experiments.

Critical strengths

Careful experimental analysis

The authors provide an extensive empirical analysis with descriptor permutation controls, gene level uncertainty estimates, matched baselines, and evaluations across multiple datasets and screens. I find the investigation of when additive models outperform direct pair predictors and whent they fail valuable. It provides a useful insights into the inductive biases governing generalization to the unseen-gene-pair regime in the Norman dataset.

Critical weaknesses

1. Insufficient evidence and limited methodological novelty

The main methodological contribution of the paper is additive supervision of a descriptor-to-effect map, yet, the single-only method, which relies on established ideas (single-perturbation prediction using linear prediction from GO descriptors followed by summation), achieves nearly identical performance. The oracle analysis also shows that prediction errors from constituents dominate interaction errors. In conclusion, the proposed main method seems to provide negligible demonstrated benefits.

2. Limited evidence of generalization beyond Norman and important methods missing

Norman is a particular dataset which is known to exhibit predominantly additive affects, which favors DACL's additive setup. On other datasets, such as Wessels, DACL achieves negative Pearson, while direct models outperform it on additional screens. Furthermore, PerturbNet, which explicitly targets the both-unknown perturbation pairs, is not evaluated despite being acknowledged.

Questions for authors

1. Clarification: Contribution of pair supervision.

DACL-C seems to only improve marginally compared to DACL-S, with a gene-level confidence interval including zero. What evidence supports the claim that the proposed pair-sum supervision provides a meaningful advantage for the both-unseen regime?

2. Clarification and additional evicence: Missing methods.

Why is PerturBnet, which explicitly addresses unseen-gene-pair prediction using GO representations, omitted from the evaluation? Could the authors kindly provide such a comparison or explain what prevents them from doing so?

3. Additional evidence: Generalization beyond additive effects

Given that Norman is known to contain predominantly additive effects, and that DACL performs poorly on the Wessels dataset, can the authors provide evidence that the proposed method is useful for predicting non-additive effects, which may be rare in the specific Norman dataset, but much more prominent in others? How does it compare to direct pair models as interaction strength increases?

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

Related papers

Discussion (0)

Sign in to comment.