acceptodds
Under review as a conference paper at ICLR 2027

PAD: JOINT INTERPRETATION OF PERTURBATION EVIDENCE

Abstract

Inferring causal direction from single- and double-perturbation assays requires understanding how a response depends on its background. The question is how much predictive value independent scoring must lose, and which measured sources change a decision. The central insight is that exchange-invariant evidence can carry no directional sign on its own yet determine how another channel should be read. PaD implements this conditional interpretation with channel-specific and joint kernel terms. Under balanced exchange symmetry with invariant context, arbitrary nonlinear additive scoring recovers at most half of the classification gain available through joint access; exchange-consistent additive decisions recover none. A common-threshold condition makes the half-gain bound exact. Observation-level bounds and independent-test certification connect this limit to fitted learners: nine simulated classifiers attain risk upper bounds at most 9.534%, below the sharp additive limit of 16.361%. Measured yeast comparisons attribute the empirical benefit to joint access rather than to a specific kernel mixture. On the Jonikas development panel, exchange-invariant measurement metadata can change directional decisions, with contributions that depend on the other retained inputs. Source removal, refitted block-mismatch controls and resampling of the evaluation records assess these contributions and their sensitivity to record composition. Together, the results connect a sharp limit on independent scoring to the practical evaluation of which evidence changes which decision, and under which retained context.

Primary area:
causal reasoning
Keywords:
causal inference, perturbation learning, genetic interactions, nonlinear data fusion, kernel methods, statistical learning theory

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_mgTrOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

this work studies how to combine two types of perturbation experiments to tell which gene regulates which ( does gene A control B, or B control A?). The motivation is that exchange-invariant evidence has no direction on its own, and can determine or flip how another directional signal is interpreted; so scoring the two data sources separately and adding them up fails, thus we must model them jointly. The paper prove a theoretical bound for this limitation, propose a simple product-kernel method to combine the two sources, and test it on simulations and two small yeast datasets.

Critical strengths

the authors start from exchange-invariant evidence has no directional sign on its own yet can change how another signal is read, and combining channel-specific and joint product kernels, they provide sufficient theoretical analysis (honestly, hard for me to validate its correctness) and a sharp bound (theorem 1)

Critical weaknesses
  • my main concern is the poor readbility, really hard to understand; such as in the title / intro, the author does not introduce background about bioinformatics. only after picking up few words in related works or experiment so that I figure out what domain and context this paper is actually applied to.

  • entire intro section without any reference.

  • lots of non-standard, some details don't make sense such as the abstract reports hyper-precise numbers down to three decimal places (e.g., 9.534% and 16.361%);

  • weird statement such as line 451 “V's observed F increment”.

  • the paper dumps dense, unmotivated equations (eg line 347), also lots of notion do not have explanations. such as Eq.1 M_A.

  • figure 1 might be ai-generated with crop-outs and awkward spacing

  • incremental method: PaD sound like a standard kernel ridge regression with a mixture of linear and ref product kernels.

  • the evaluation jonika dataset has only 21 positive relations; which is no convincing and hard to prove the generalizes of method.

Questions for authors

please refer to the weakness.

also the main body constantly points to supp sections and tables (eg S1–S13, Table S26); pls concretely indicate appendix when referring.

Reviewer_pC2wOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes a kernel based method for inferring causal direction for pertubation assays under certain symmetry conditions. They show some theoretical results and evaluate their method on simulations and published yeast data.

Critical strengths

Inferring causal direction for pertubation assays and in general is an interesting and relevant task.

The paper provides evaluation on simulations and published yeast data, and the authors acknowledge limitations of the biological evaluation.

Critical weaknesses

The paper is hard to understand in general (maybe partly because it is a bit out of my expertice). The paper seems unfinished. Figure 1 looks unfinished. There are a lot of significant digits for each number.

The proposed method is not justified by the empirical results. And the benefit of the method remains unclear.

Questions for authors

What is the advantage of the method when simpler baselines beat it on 2/3 metrics?

Why do you think the baselines beat PaD on 2/3 metrics in table 1?

Why are there so many significant digits?

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

This paper investigates how to integrate gene expression response fingerprints-obtained following gene knockouts--with phenotypes resulting from single- and double-gene perturbations to infer regulatory, upstream-downstream relationships between genes and to rank these candidate relationships. The core challenge addressed is that the directional implication of one type of evidence may depend on the context revealed by another; consequently, simply calculating scores for each data channel independently and summing them may fail to fully exploit the combined information.

The paper demonstrates that additive scoring can recover, at most, half of the classification gain achievable through joint access. If additive decision-making is further required to satisfy exchange consistency, this additional benefit cannot be attained. The authors also specify the conditions under which the lower bound is exactly reached, and analyze the impact of a prescribed form of observation noise on these conclusions.

In terms of methodology, PaD combines channel-specific linear kernels with Gaussian product kernels, allowing background information to modulate directional evidence, and derives directional predictions through antisymmetrization via swapping.

The experiments utilized two yeast datasets: Costanzo and Jonikas. On the Costanzo dataset, PaD correctly predicted the direction for 611 out of 657 cases, compared to 613 out of 657 for the matched additive control. On the Jonikas dataset, PaD achieved an F-AUROC of 0.8811 and a d-AUROC of 0.8448, both of which outperform the matched additive control (0.8277 and 0.7143, respectively).

Critical strengths

1.This paper not only demonstrates the poor performance of additive baselines but also characterizes the optimal risk bounds for additive scoring models that allow for arbitrary intra-channel non-linearities under explicit assumptions. Risk certification on an independent test set further reveals that the jointly-fitted models can surpass these bounds. Consequently, the performance gap cannot be attributed simply to the limited expressive capacity or inadequate optimization of a given additive baseline. This connection-linking lower bounds derived for a specific function class to the actual performance of a model-offers a more compelling argument than standard empirical performance comparisons.

2.PaD employs a concise kernel-based approach--aligned with theoretical motivations-to represent cross-channel interactions. Its parity-based decomposition clearly demonstrates that permutation-invariant background features can modulate permutation-antisymmetric directional signals via product terms, rather than being simply discarded. While product kernels themselves are not novel mathematical tools, their integration with the permutation constraints inherent in directional tasks provides a clear and principled framework for leveraging background evidence.

Critical weaknesses

1.While the theoretical limitations of independent scoring are significant, the paper does not sufficiently demonstrate whether this paradigm constitutes the primary bottleneck in practical perturbation-based inference. Existing literature already covers joint factor modeling and the use of complementary phenotypes for pathway inference; therefore, it is inaccurate to categorize all prior approaches under the restrictive a(P)+b(D) framework. The paper needs to more explicitly identify which specific practical methods or application constraints fall into this category, and explain how this theoretical analysis refines our understanding of these methods--going beyond merely pointing out the existence of constraints imposed by additive interactions.

2.The experiments involved only two yeast datasets sharing the same expression data source; this is insufficient to determine whether the conclusions can be generalized to different biological systems or measurement conditions. In the Jonikas dataset, there are only 21 positive examples supported by annotations, while unannotated records--used here as reference negative--do not necessarily represent confirmed inverse relationships. As for the Costanzo dataset, the labels are derived from the direction of expression response rather than from independently verified, direct causal edges. Consequently, while the current experiments provide strong support for predictive capability within specific datasets, the evidence for the ability to recover causal ordering in a general sense remains limited.

Questions for authors

1.Which existing perturbation-based inference methods or practical operational workflows--that utilize input interfaces comparable to those in this paper actually fall into the category of independent additive scoring models, defined as a(P)+b(D)? For approaches already employing joint modeling (as discussed in the "Related Work" section), what additional insights or practical guidance can the theoretical analysis presented in this paper provide?

2.Given that the reference negative examples for Jonikas are not confirmed inverse relationships, which results can serve as evidence for the recovery of directional orientation (upstream vs. downstream), rather than merely as evidence for the ranking of annotated supportive relationships? Have the directional sign accuracy results for PaD, joint kernels, and other baselines already been compared specifically on relationships that possess independent evidence for their directionality?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.