acceptodds
Withdrawn from ICLR 2027

When Not to Correct: Evidence-Based Interaction Prediction for Combinatorial Perturbations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Predicting the transcriptomic effects of unseen combinatorial genetic perturbations from observed single-gene responses is a central challenge in functional genomics. A key difficulty is not whether genetic interactions exist, but whether interaction structure learned from observed pairs transfers reliably to new combinations. We introduce Hierarchical Null Regression (HNR), an evidence-based framework that adds interaction corrections only when supported by the training data. HNR separates residual interactions into a low-rank Core and full-rank Tail, with structure selected by permutation testing and predictions regularized through empirical Bayes estimation and uncertainty-aware shrinkage. When transferable structure is unsupported, HNR suppresses the interaction correction rather than forcing it. Across four Perturb-seq datasets, HNR consistently improves prediction of unseen combinations over strong linear, neural, graph-based, and pretrained baselines, while requiring no pretraining, external knowledge graph, validation-set hyperparameter tuning, or GPU computation. These results suggest that reliable combinatorial perturbation prediction depends not only on modeling interactions, but on determining when those interactions can be trusted. Code is provided as supplemental material for review and will be publicly released upon acceptance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
perturbation prediction, genetic interactions, empirical Bayes, compositional generalization, single-cell genomics

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

This paper explores the problem from the perspective that observing non-additive effects does not necessarily imply an ability to predict such effects in unseen combinations. It focuses on the OOD double-gene perturbation prediction setting: All single-gene perturbation responses are observed, together with a subset of double-gene perturbations, and the goal is to predict unmeasured double-gene combinations.

Critical strengths

1. The paper examines when modeling and correcting non-additive interactions improves single-cell perturbation prediction. This is a practically relevant question, as recent studies have found that complex neural networks and foundation models do not consistently outperform simple linear baselines, particularly for unseen or combinatorial perturbations.

2.The evaluation covers four double-gene perturbation datasets. These datasets span different perturbation technologies and experimental settings, providing a relatively broad assessment of robustness and generalization across combinatorial perturbation scenarios.

3. HNR is a lightweight, computationally efficient statistical method. It requires no pretrained models, external biological knowledge graphs, or GPU training, yet remains competitive with more complex neural and graph-based models across multiple datasets. Its simplicity is well suited to the paper’s focus on when additional modeling complexity improves prediction.

Critical weaknesses

1. Missing relevant references. The paper asks when non-additive corrections improve double-gene perturbation prediction, motivated by the observation that complex models do not necessarily outperform simple linear baselines in single-cell prediction. However, the related-work discussion focuses mainly on GEARS, CPA, and single-cell foundation models. It gives limited attention to recent approaches based on causal representation learning, soft interventions, and sparse mechanism shifts, including Discrepancy-VAE [1] , sVAE+ [2] , and PerturbedVAE [3], these three methods are all related to double gene perturbation prediction.

2. Poor metrics. Although the paper reports six metrics, most are variants of MAE computed over different gene subsets or using different weighting schemes. The evaluation therefore focuses on absolute prediction error, with limited assessment of whether the model recovers the structure of perturbation responses. Common complementary metrics, such as Δ\DeltaPearson and ΔR2\Delta R^2, are not reported.

[1] Zhang J, Greenewald K, Squires C, et al. Identifiability guarantees for causal disentanglement from soft interventions[J]. Advances in Neural Information Processing Systems, 2023, 36: 50254-50292.

[2] Lopez R, Tagasovska N, Ra S, et al. Learning causal representations of single cells via sparse mechanism shift modeling[C]//Conference on Causal Learning and Reasoning. PMLR, 2023: 662-691.

[3] Jiang W, Liu Y, Cai Y, et al. What Makes a Representation Good for Single-Cell Perturbation Prediction?[J]. arXiv preprint arXiv:2605.19343, 2026.

Questions for authors

Q1. The paper studies combinatorial perturbation prediction and is partly motivated by the observation that complex models do not necessarily outperform simple linear or additive baselines. However, the related work gives limited attention to causal representation learning approaches such as Discrepancy-VAE, sVAE+, and PerturbedVAE. In particular, PerturbedVAE systematically examines the competitiveness of simple linear models and evaluates OOD prediction of double-gene perturbations. Could the authors explain why these works are not discussed and clarify how HNR differs from them in problem setting and methodological objective?

Q2. The evaluation relies mainly on MAE-family metrics. Could the authors also report complementary metrics, such as ΔPearson\Delta\mathrm{Pearson} or ΔR2\Delta R^2, to assess whether the model recovers the structure of perturbation responses?

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

This paper addresses the problem that predicting the transcriptomic effects of unseen two-gene combinatorial genetic perturbations given observed single-perturbation responses and a training set of measured perturbation pairs. So unlike existing assumptions that non-additive genetic interactions learned from observed pairs can always be reliably transfered to unseen pairs, the authors propose Hierarchical Null Regression, which is an evidence-based framework that treats additivity/base effects as a null model and introduces interaction corrections only when supported by the training data.

From my understanding, the HNR includes three stages: 1. an adaptive base predictor. 2. decomposition of residuals after fitting the base. 3. Kernel regression and an empirical Bayesian shrinkage.

The frameworks are evaluated across 4 perturb-seq datasets with 6 metrics and the proposed HNR achieves the lowest error on most of the dataset×metric combinations while running in seconds on CPU without pretraining or GO graphs required.

Critical strengths

1. Strong motivation

I believe the motivation for “When Not to Correct” is sufficient since recent benchmarks have repeatedly shown that complex deep learning / foundation models often fail to beat simple linear or additive baselines on unseen perturbations.

The task setting is also legit scientifically, where the authors measure individual perturbations and a subset of pairs, then predict additional combinations. This setting has established existing works, including SALT/PEPER[1] and recent combinatorial prediction studies, e.g, CLIPR[2].

2. Reasonable inductive biases for small datasets

It is a reasonable design to separate a strong base predictor from a regularized residual model when the number of training pairs is small and the expression output is high dimensional.

The pair kernels are also mathematically sensible. Eq 6-7 are symmetric under swapping the two perturbations and correspond to valid positive-semidefinite kernels.

Also regarding the performance, HNR achieves SOTA results across most evaluations settings (Table 1), and more importantly, since it relies on closed-form linear algebra, eigendecomposition, and low-dimensional evidence maximization, it runs in 3.8 secs on a single CPU on Norman (compared to 14 mins on GPU for GEARS).

Critical weaknesses

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

1. Significant result inconsistency

For the authors, please carefully check the following inconsistency in the main text tables: for DirectCapture, 0.40 in Table 4 differs by two orders of magnitude from 42.89 in Table 1, and even multiplying 0.40 by 100 gives 40, not 42.89.

Moreover, I found the text claims that "downstream MAE remains unchanged at the two extremes for all four datasets", but for Wessels2023, DirectCapture, and GSE, the unchanged values at the 90th and 99th percentiles (1.58, 0.40, 1.19) do not match the 95th-percentile values in Table 1 (1.60, 42.89, 1.14).

2. The strongest novelty claim is not sufficiently distinguished from existing work

The authors omitted a highly related paper SALT & PEPER [1] published on ICLR 2024 workshop. It already studies the same observed-single/unseen-combination setting and adds a learned nonlinear correction to additive single-effect predictions. It also examines failures under distribution shifts.

Therefore, HNR’s unique contribution should be reframed around statistical residual decomposition, evidence-based parameter estimation, or additional shrinkage. The “additive-reference + correction” strategy is already well established.

A direct PEPER comparison is more relevant than a generic MLP comparison.


References:

  1. Season combinatorial intervention predictions with Salt & Peper (ICLR 2024 Workshop)
  2. Latent Causal Diffusions for Single-Cell Perturbation Modeling (Arxiv)
Questions for authors

1. What the permutation test truly detects?

My understanding is the permutation test detects residual structure, not whether that structure is transferable. Throughout the paper, the authors claim that selecting K>0K > 0 vs K=0K = 0 via permutation testing (Eq 5) determines whether "transferable interaction structure" exists in the dataset. But a statistical distinction here is that Eq. 5 only tests whether the training residuals RcR_c are correlated across genes (low-rank co-expression structure), without ever looking at the input pair features EE.

A simple counterexample is $ R=z u^\top $ where uu is a shared expression program and the pair scores ziz_i are independent of all single-effect features. This matrix can have a very strong rank-one component, while $ \mathbb E[z_*\mid E]=0 $ for an unseen pair.

2. Double shrinkage is not guaranteed to be a Bayesian posterior estimator

The kernel stage is described through Gaussian marginal likelihood and predictive means and variances. If those predictions are standard GP posterior outputs, the posterior mean has already incorporated Bayesian shrinkage. I suspect the method is undergoing a double shrinkage.

For instance, let the posterior have mean mm and covariance Σ\Sigma. Under squared loss,

E[r−wm2∣D]=tr⁡(Σ)+(1−w)2∥m∥2\mathbb{E}[r-wm^2 \mid D]=\operatorname{tr}(\Sigma)+(1-w)^2\|m\|^2

So under that same posterior model, an additional factor 0<w<10<w<1 does not improve the Bayes estimator, instead, it is extra regularization.

3. The authors should revise the claims in Sections 1, 3, and 5 to clarify that Eq. 5 only identifies low-rank residual gene programs within the training set, whereas actual transferability to unseen pairs is determined only at the kernel regression stage.

4. The authors should either provide a rigorous hierarchical derivation showing why the GP posterior mean is not already properly shrinked, or reframe Eq. 9 as an empirical post-hoc regularization heuristic.

Reviewer_b6o8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies prediction of unseen double genetic perturbations from observed single perturbations and a subset of observed doubles.

It proposes Hierarchical Null Regression (HNR). HNR first builds a base prediction by blending a calibrated additive model with a one-hot Ridge model. It then adds an interaction correction: low-rank residual programs are selected by a permutation test, predicted for new pairs by GP regression on single-effect kernels, and shrunk toward zero when predictive uncertainty is high.

The main claim is that interaction corrections should be added only when the training data support them. On four Perturb-seq datasets, HNR is reported best in 22 of 24 dataset-metric pairs against eight baselines, with low compute cost.

Critical strengths

The paper takes the strong linear baseline results in this field seriously and builds a simple, closed-form method on top of them. The code is clean, runs on CPU in seconds, and I found no leakage between test pairs and model fitting. This makes the results easy to reproduce and gives a useful reference point for future work. The question of whether interaction structure transfers to new pairs, not only whether it exists, is also relevant for designing combinatorial screens.

Critical weaknesses
  1. It is unclear what the paper finds beyond the additive model. Each part of HNR is a standard statistical tool, so the contribution depends on showing that the framing works. Most of the gain comes from Ridge, which is already a baseline, and HNR uses Ridge alone as its base on three of four datasets. Over Ridge, much of the remaining gain seems to come from adding the average training residual, which is the same shift for every pair. HNR is not better than Ridge in MAE on Wessels2023 and is slightly worse on GSE164996 (Table 6). The programs in Table 7 mostly match cell-type axes already reported by Norman et al. (2019).

  2. The main claim, “when not to correct”, is not tested directly. The paper says it separates whether interactions exist from whether they can be predicted, but does not show this in the data. The scaling weight only depends on how similar a test pair is to training pairs. No analysis checks whether pairs with low weights are really pairs where the correction would hurt. Also, all metrics are MAE-based, so they cannot tell a global shift from pair-specific interactions.

Questions for authors
  1. Could you add “Ridge + average training residual” to Table 2? Could you also report IntR2 and IntSpearman (already in your code) after removing this average? This would show how much of the gain is pair-specific. Also, some bold marks in Table 2 are on the full model when an ablation is better.

  2. Could you group test pairs by their scaling weight (high or low) and by their true interaction strength, and compare the error with Ridge in each group? Please also show how the weights are distributed.

  3. Could you report DEG overlap between predicted and true top genes? It would help to split DEGs into those already seen in a single perturbation and those that appear only in the pair. The second group shows whether HNR learns anything beyond additivity.

  4. Could you add a no-change baseline, a tuned Ridge, and a neural model that takes the two single effects as input and predicts only the residual? On DirectCapture, all neural baselines give almost the same numbers. Did they collapse to zero? Reporting Norman results with the metrics of Ahlmann-Eltze et al. (2025) would also allow comparison with published numbers.

  5. The paper has no limitations section. One limitation is small training sets. On GSE164996, with only 32 training pairs, the model selects ten programs, but adding them hurts test performance. A learning curve with fewer training pairs would show whether HNR safely falls back to the base model.

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

Related papers

Discussion (0)

Sign in to comment.