acceptodds
Under review as a conference paper at ICLR 2027

SENTRY-BD: Bilevel Measurement Design for Cross-Context Perturbation Recovery

Abstract

We introduce Sentry-BD, a framework that learns which genes to measure by optimizing cross-context perturbation recovery after refitting the predictor. Compact panels could extend perturbation screens across cell states and time points, but useful measurements must predict the endpoint and remain informative alongside the other selected genes. We formulate panel learning as a budget-constrained bilevel problem: a shared gene ordering determines context-grouped measurement weights, an analytic inner model fits the response map, and disjoint query perturbations train the design. An exact gene-utility derivative accounts for decoder refitting; a signed response loss complements magnitude reconstruction. One training run produces nested panels at three budgets. On an independent confirmation cohort of 1,699 primary CD4 perturbations, a 128-gene panel recovers a 10,282-gene endpoint with 1.711% lower MSE than a representation-matched PERSIST adaptation (95% CI: 1.492–1.945%). The reduction is 3.498% over 1,676 K562-to-RPE1 perturbations. Five training seeds and component controls show that learning through query prediction provides the main gain. The framework connects measurement allocation to the response prediction task through a differentiable, computationally compact design objective.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
perturb-seq, gene-panel, feature-selection, transcriptomics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_EUdDOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SENTRY-BD, a bilevel optimization framework for selecting compact gene panels for cross-context perturbation response recovery. The method combines differentiable gene selection, an analytically refitted ridge predictor, and a signed-response loss. Experiments on CD4 T-cell and K562-to-RPE1 datasets report improvements over existing gene-selection methods, including PERSIST and residual selection. Additional theoretical derivations and ablation studies are provided to justify the proposed framework.

Critical strengths

The paper addresses a relevant problem in efficient perturbation measurement. Optimizing gene selection according to downstream prediction performance is a reasonable direction. The use of a common decoder, multiple baselines, and component ablations provides a useful basis for evaluating the proposed method.

Critical weaknesses

Weakness 1: The central methodological contributions are not supported by the experimental evidence, raising serious concerns about novelty and necessity.
a. The proposed bilevel differentiation provides negligible benefits. Table 2 shows that stopping gradients through the inner-model refitting achieves an MSE of 0.07824, nearly identical to the full method's 0.07821. Removing the signed-response loss similarly yields 0.07823. These results directly question whether the paper's emphasized technical components contribute meaningfully to performance.
b. Limited theoretical novelty. The proposed exact gradient and finite-change identities primarily follow standard differentiation and algebraic properties of ridge regression. The paper neither establishes meaningful theoretical guarantees for the resulting discrete gene panels nor demonstrates an empirical advantage attributable to differentiating through the inner optimization. The primary methodological contribution appears insufficiently novel and empirically unjustified, substantially weakening the paper's significance.
Weakness 2: The empirical improvements are marginal, inconsistent across evaluation metrics, and insufficient to establish superiority over strong baselines.
a. Minimal improvement over competitive methods. On CD4, SENTRY-BD reduces MSE by only 0.66% relative to the stronger 64D residual-selection baseline. On K562-to-RPE1, the improvement is approximately 0.16% and is not statistically significant. Moreover, the proposed method performs worse than residual selection on standardized MSE in both tasks, while simple variance selection achieves better signed top-20 overlap.
b. Insufficiently controlled comparisons. The headline comparison against PERSIST is potentially misleading because PERSIST uses a restricted 1,000-gene candidate pool, whereas SENTRY-BD searches the full gene space. The paper's own ablation shows that restricting SENTRY-BD to the same candidate pool noticeably reduces its advantage. Therefore, the reported improvements cannot be attributed solely to the proposed optimization strategy.
The empirical evidence fails to demonstrate a substantial or consistent advantage over simpler alternatives, making the additional methodological complexity difficult to justify.

Questions for authors

Q1. Given that stopping gradients through the inner optimization yields nearly identical performance, what concrete evidence supports the necessity of the proposed bilevel differentiation rather than simpler supervised gene-selection approaches?
Q2. How can the authors establish that the reported improvements originate from the proposed optimization method rather than differences in candidate-pool size or baseline configurations? Can all methods be compared under strictly matched candidate spaces and model settings?
Q3. How do the authors justify the significance of the marginal improvements over strong residual-selection baselines, particularly when the gains are statistically insignificant on one task and the method underperforms on several alternative metrics? Do these conclusions remain consistent across different decoder configurations?

3 comments
AuthorsOpenReview ↗

Response to Reviewer EUdD

Thank you for the detailed review. We begin by explaining what two of the controls measure, since this affects several of the points raised. We also report a test on 1,606 CD4 perturbations that had not been used before (general response, Appendix F).

Q1 (Weakness 1a). Evidence for the bilevel formulation. The stop-gradient variant in Table 2 is itself a bilevel method. At every step it solves the inner ridge problem and evaluates the refitted predictor on separate query perturbations; it removes only the second-order term of the derivative (Appendix A.2). Table 2 therefore compares two ways of differentiating the same bilevel objective.

The comparison with simpler supervised selection is the fixed-initialization control, which starts from a supervised marginal ranking:

  • Bilevel training improves on this starting point by 2.6% on CD4 confirmation (0.08028 → 0.07821) and by 2.56% [2.33, 2.79] on the new test set.
  • Separate query sets add a further +0.37% [0.24, 0.49] on the new test set.
  • The exact second-order term gives a further gain when training data are limited. With the exact leave-one-out solver, it improves on the stop-gradient variant on K562→RPE1 by +0.70% [0.32, 1.09] at B = 32 and +0.55% [0.12, 0.96] at B = 64 (Table 3A).

Weakness 1b. Role of the theory. Proposition 1 gives the effect of refitting in closed form: each gene's value is its query innovation times its endpoint coupling. This form makes exact leave-one-out training over all perturbations possible in seconds with G ≈ 10^4 genes (§4.2), and this is what allows the method to search the full gene set.

Q2 (Weakness 2b). Same candidate pool. Table 3B runs this comparison directly. Inside PERSIST's own 1,000-gene pool, with the same decoder and budget, SENTRY-BD outperforms PERSIST-64D:

B = 128, raw MSEPERSIST-64DSENTRY-BD
CD4, 1,000-gene pool0.079570.07874
K562→RPE1, 1,000-gene pool0.039250.03744
New test set, 1,000-gene pool—+0.92% [0.79, 1.06]

The gain over PERSIST therefore holds with matched candidates. Given all genes, PERSIST becomes slightly worse (0.07981) while SENTRY-BD improves. All other baselines already use the full gene set, the same decoder and the same budget.

Q3 (Weakness 2a). Significance and consistency.

  • New data. On the 1,606 new perturbations, all five main comparisons are positive with 95% intervals above zero. The submitted method gains +0.49% [0.34, 0.63] over Residual-64D and +1.80% [1.55, 2.07] over PERSIST-64D, and the coupled version gains +2.77% [2.57, 2.96].
  • K562→RPE1. With the exact leave-one-out solver, the gain over Residual-64D is +0.90% [0.16, 1.88].
  • Decoder settings. SENTRY-BD has the lowest raw MSE at linear decoder ranks 32, 64 and 128 (Table 4). With the shared three-stream predictor, its panel beats the Residual-64D panel by +1.25% [1.06, 1.43] on the new test set.
  • Other metrics. With the shared three-stream predictor, SENTRY-BD also leads on ρΔ (0.2950 vs 0.2368) and centroid accuracy (67.55% vs 62.43%).
  • Practical size. 96 genes reach the error of the 128-gene Residual-64D panel, a 25% saving in measurements, and the gain holds under gene-family resampling.

Thank you for the additional experiments. The matched candidate-pool comparisons and the fresh test set address some of my concerns, but the central methodological contribution remains weakly supported. Appendix Table 9 shows only a 0.09% improvement over the stop-gradient control, with a confidence interval that includes zero, while removing the signed-response loss has essentially no effect. The exact gradient helps only in some Replogle settings and performs worse at two smaller budgets on the fresh CD4 test. I therefore remain unconvinced that the paper establishes the necessity, consistency, or practical value of its main technical components.

Beyond the technical issues, I have serious concerns about the authors’ research attitude and the overall quality of the submission. The original manuscript strongly appeared to be AI-generated throughout: its organization was highly careless, its presentation was awkward and difficult to read, and it lacked the level of human verification and academic rigor expected of an ICLR submission. After revision, the methods, experiments, and narrative changed substantially, yet these changes were not clearly marked. In my view, this indicates that the original submission was not a mature piece of work and was instead extensively reconstructed during the review process. I consider this highly unprofessional.

More seriously, even the abstract spacing in the revised manuscript does not follow the ICLR template. This is not merely a formatting detail; it further reflects a lack of care toward the submission requirements and basic manuscript quality control. I do not believe a submission in this state should proceed through the normal review process.

I also do not understand the basis for the scores of 3 given by other reviewers. I strongly encourage them not to rely on AI-generated summaries or judgments, but to personally read and compare the original and revised manuscripts. The maturity, writing quality, and revision practices demonstrated here do not meet the standard I expect from an ICLR paper.

Considering the paper, the additional evidence, and the concerns above, I retain my original score of 1 as my final rating. I strongly recommend that the AC personally examine every version of the manuscript and its revision history before making a decision.

AuthorsOpenReview ↗

Response to Reviewer EUdD

Thank you for reading each version closely and for the specific points. Our responses follow.

Technical contribution. The paper’s main claim is that a panel should be chosen by the held-out error of the predictor refitted on it. The full and stop-gradient updates optimize the same objective and differ by one second-order term. The original submission already reported that this term has a small effect (§5.4). The evidence for the formulation is the comparison with the supervised starting ranking. On the 1,606 unused perturbations, learning the panel through the refitted predictor reduced error by 2.56% [2.33, 2.79], and scoring it on separate query perturbations gave a further 0.37% [0.24, 0.49]. The sparse selector added in the revision does not use the signed-response term.

Checks after the revision. After posting the revision, we checked the gradient code against explicit ridge refits and recomputed the reported comparisons from the stored per-perturbation results. The code for the sparse stop-gradient control left out the contribution of the held-out row. We fixed this. Table 3A now reports the corrected comparison: at 128 genes, the full and stopped updates are close on both tasks. The three affected rows of Table 9 keep their recorded values and are marked in Appendix F. None of the five primary comparisons uses this control, so they are unchanged.

We also found that the previous revised PDF had been compiled with a modified style file, which changed the spacing of the abstract, captions and lists. The PDF now posted uses the unmodified ICLR 2027 style file and includes both corrections.

Changes in the revision. The submitted method is unchanged and appears as “SENTRY-BD (submitted)” in every table, including the new test on 1,606 perturbations. To address the questions about decoder dependence and panel stability, the revision adds a leave-one-out selector and a three-stream predictor, reported in separate rows. We have sent the Area Chair a section-by-section list of the changes.

Verification. The AI-use statement describes how AI tools were used and what the authors checked. The supplement contains the locked protocol, the per-perturbation results, and verify_results.py, which recomputes the primary and secondary comparisons on the new cohort as well as the shared-pool comparisons.

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

SENTRY-BD learns a gene panel for recovering perturbation responses across contexts. It selects genes under a measurement budget and fits an analytic ridge-regression predictor. Disjoint query perturbations guide panel learning, and one ordering produces nested panels at several budgets. The paper derives a gene-utility gradient that accounts for predictor refitting and adds a signed-response objective. Evaluation uses held-out CD4 perturbations and K562-to-RPE1 recovery, with a common final decoder, stronger selector baselines, multiple seeds, and component controls. The gains are modest and demonstrated retrospectively on aggregate responses, rather than through prospective targeted measurements.

Critical strengths
  1. The panel-learning objective fits the problem well. Genes are selected for how well they support response recovery after predictor fitting, rather than for variability or marginal association alone. The analytic inner model and nested panels make this a compact approach to a useful measurement-design problem.

  2. The comparisons are unusually careful about what causes the improvement. A common final decoder and representation-matched selectors help separate panel quality from representation and decoder effects. The authors also credit query-driven learning as the main source of improvement, which is consistent with their ablations.

Critical weaknesses
  1. The gain over the strongest simple selector is small. On CD4, MSE improves by about 0.657% over the stronger residual selector (Table 1). The advantage is less clear in the cross-context task, and other endpoints do not uniformly favor SENTRY-BD. It remains unclear whether this translates into a useful reduction in measurement budget or a more effective biological panel.

  2. The exact gradient and signed loss add little to the reported prediction performance. CD4 MSE changes from 0.07821 to 0.07824 with a stopped inner gradient, and to 0.07823 without the signed loss (Table 2). The authors acknowledge that query-driven learning provides the main gain and select the signed term using construction-only validation. The derivative is still a useful analytic result, but the practical contribution of these additional components appears limited.

  3. The evaluation supports conclusions within two aggregate-response tasks. Confirmation perturbations come from the same underlying screen, and Appendix D limits its intervals to fixed-cohort variation and panel-learning stochasticity. The identity-level bootstrap does not assess dependence among related perturbations or shared controls. Assay noise and donor variation are also left to future work, as the discussion acknowledges.

Questions for authors
  1. Using the existing per-perturbation results, what fraction of CD4 perturbations improve over the 64D residual selector? Are gains spread across response magnitudes or concentrated in a few large responses? Figure 3 already provides budget curves; the remaining question is how broadly the small mean gain is shared.

  2. Across budgets and tasks, does the exact inner gradient improve convergence, runtime, or panel quality relative to the stop-gradient variant? This would clarify its practical value despite the nearly identical headline errors.

  3. Where suitable groups exist, how sensitive is the positive CD4 comparison to resampling related perturbations together? This check concerns dependence within the current screen and could use the same fixed-atlas setting described in Appendix D.

2 comments
AuthorsOpenReview ↗

Response to Reviewer oMdA

Thank you for the constructive review and the precise questions. We also added a test on 1,606 CD4 perturbations that had not been used before (general response, Appendix F). Below we answer each question on both the confirmation cohort and this new set.

Q1. How broadly the gain is shared. At B = 128, SENTRY-BD has lower error than Residual-64D for 61.9% of CD4 confirmation perturbations (1,051/1,699). By quartile of response size:

QuartileResidual-64DSENTRY-BDGainWin rateNew test set: gain (win rate)
Q1 (smallest)0.048280.04795+0.69%62.8%+0.85% (67.7%)
Q20.058250.05799+0.45%62.1%+0.44% (61.8%)
Q30.072920.07240+0.71%60.6%+0.57% (62.6%)
Q4 (largest)0.135440.13449+0.71%61.9%+0.35% (60.2%)

Every quartile improves in both cohorts, so the mean gain is not driven by a few large responses. On the new test set, 63.1% of perturbations improve overall. In budget terms, 96 genes from the learned ordering reach the error of the 128-gene Residual-64D panel, which saves 25% of the measurements.

Q2. Value of the exact inner gradient. Both the full and stop-gradient updates refit the predictor at every step, and this shared refitting provides the gain. MSE falls from 0.08028 with a fixed initialization to 0.07821 (Table 5), and on the new test set the learned panel improves on the fixed ranking by 2.56% [2.33, 2.79]. Convergence and runtime are similar for the two updates: the logged training losses differ by at most 0.002, and a run takes 8.2 s vs 7.7 s on one A100.

The revision replaces the 256-perturbation episodes with an exact leave-one-out fit over all perturbations (§4.2), which improves both panel quality and stability:

  • On the new test set, the sparse selector improves on the submitted method at every budget: +0.31% [0.16, 0.46], +0.15% [0.03, 0.26] and +0.17% [0.08, 0.26] at B = 32, 64 and 128.
  • Panel overlap across five seeds rises from a Jaccard index of 0.364 to 0.570, and the selector takes 2.9 s per fold on Replogle and 7.8 s on CD4.

Table 3A compares the full and stopped versions of this selector at three budgets.

Q3. Resampling related perturbations together. We grouped targets into HGNC gene families (940 clusters on CD4 confirmation) and resampled whole clusters 5,000 times. The gain over Residual-64D stays clearly positive:

Confirmation (1,699)New test set (1,606)
SENTRY-BD (submitted)+0.66% [0.50, 0.82]+0.49% [0.28, 0.66]
SENTRY-BD (sparse LOO)+0.73% [0.58, 0.87]+0.67% [0.49, 0.81]
SENTRY-BD (coupled, three-stream predictor)+3.23% [2.98, 3.46]+2.77% [2.54, 2.98]

These intervals are close to the per-perturbation ones, so dependence within gene families does not drive the result.

Thanks for the thorough follow-up. The per-perturbation and quartile results, especially their replication on the new CD4 test set, show that the gain is small but broadly shared, and the 96-versus-128-gene comparison makes the practical benefit easier to interpret. The grouped resampling also addresses my dependence concern; I’ll keep my weak-accept rating, while expecting the revision to describe the exact gradient’s practical contribution modestly and include these new analyses.

Reviewer_p96QOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SENTRY-BD uses bilevel optimization to select compact gene panels for recovering perturbation responses in another context. It is evaluated on an independent CD4 confirmation cohort and on K562-to-RPE1 transfer, with all selectors compared through the same decoder.

Critical strengths

The experimental protocol is careful. It separates construction and confirmation identities, records the recipe before confirmation, and uses the same final decoder for all selectors. The controls are also informative. They show both where SENTRY-BD helps and where stronger residual-selection controls close the gap.

Critical weaknesses

The practical effect is modest and varies by task. The CD4 result is positive, while the K562-to-RPE1 comparison is less conclusive.

The evaluation is retrospective and does not test a physically measured targeted panel. This mainly limits the deployment claim rather than the current methodological result.

Questions for authors
  1. Does the MSE reduction change any downstream scientific conclusions?

  2. The stop-gradient control is close to the full method. Could the authors clarify when exact differentiation through refitting is expected to matter?

  3. How sensitive might the learned ordering be to measurement noise, missing genes, or gene-specific costs?

  4. A random-panel control or a matched learned-selector baseline would make the comparison more complete. Are such results available?

1 comment
AuthorsOpenReview ↗

Response to Reviewer p96Q

Thank you for the positive review and the helpful questions. We also added a test on 1,606 CD4 perturbations that had not been used before (general response, Appendix F). On this new set, the submitted SENTRY-BD outperforms Residual-64D by +0.49% [0.34, 0.63] and PERSIST-64D by +1.80% [1.55, 2.07], so the CD4 result holds on new data.

Q1. Downstream conclusions. The reduction has three practical consequences:

  • 96 genes from the learned ordering reach the error of the 128-gene Residual-64D panel (0.07857 vs 0.07873), so the same accuracy needs 25% fewer genes.
  • Prediction for individual perturbations improves. 61.9% of perturbations have lower error, ρΔ rises from 0.231 to 0.243, and centroid accuracy rises from 0.594 to 0.600. The new test set shows the same pattern (ρΔ 0.247 vs 0.235).
  • On the published GWCD4i aging signature, the coupled panel recovers the measured ranking better than Residual-64D with the same predictor (Spearman 0.5642 vs 0.5377; difference 0.0265 [0.0004, 0.0532]; Table 2B).

Testing a physically measured panel is the natural next step.

Q2. When exact differentiation matters. The stop-gradient variant still refits the ridge predictor at every step and drops only the derivative through the refit (Appendix A.2). Our results place the gain in learning through the refitted predictor. MSE falls from 0.08028 with a fixed initialization, to 0.07850 when the support and query sets overlap, and to 0.07821 with separate query sets (Table 5). On the new test set, the learned panel improves on the fixed ranking by 2.56% [2.33, 2.79]. The exact second-order term refines this procedure at little extra cost (8.2 s vs 7.7 s per run), and at 128 genes the full and stopped updates are close on both tasks (Table 3A).

Q3. Noise, missing genes and costs. The learned ordering is robust to measurement problems (Table 7). Query noise up to σ = 0.20 changes MSE by less than 0.05%. With 5%, 10% or 20% of panel genes missing, MSE rises gradually (0.07830, 0.07848, 0.07888) and stays below Residual-64D at every level. Gene-specific costs enter through the budget constraint Σ c_g w_g = C, with the same closed-form gate (§4.1, Appendix A.1).

Q4. Random and matched baselines. Both are now included:

  • SENTRY-BD outperforms 20 random panels, refitted with the same decoder, by 2.5% on CD4 and 11.2% on K562→RPE1 (Table 1). On the new test set the margin is 2.55%.
  • Inside PERSIST's own 1,000-gene pool, SENTRY-BD outperforms PERSIST-64D on both tasks: 0.07874 vs 0.07957 on CD4, and 0.03744 vs 0.03925 on K562→RPE1 (Table 3B). On the new test set the margin is +0.92% [0.79, 1.06]. Giving PERSIST all genes does not help it (0.07981).

For K562→RPE1, the sparse leave-one-out selector gives +0.90% [0.16, 1.88] over Residual-64D (§5.2).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank all reviewers for their careful and constructive reviews. The revised PDF adds the analyses they requested (§5.4, Tables 3–9, Appendices E–F). The main points are below.

1. A new test on 1,606 unused perturbations (§5.4, Appendix F). The original submission set aside 1,606 eligible CD4 perturbations and did not use them (original Appendix C). We used them as a new test set. First, we fixed every panel and model using the construction data only. Next, we recorded the full list of comparisons, along with hashes of the code and of the fixed models. Finally, we evaluated once.

All five main comparisons are positive, and their 95% intervals stay above zero, including when related genes are resampled together:

Comparison (B = 128, raw MSE)Reduction95% CIWin rate
SENTRY-BD (submitted) vs Residual-64D+0.49%[0.34, 0.63]63.1%
SENTRY-BD (submitted) vs PERSIST-64D+1.80%[1.55, 2.07]74.3%
SENTRY-BD (LOO solver) vs Residual-64D+0.67%[0.53, 0.79]65.5%
SENTRY-BD panel vs Residual-64D panel, same three-stream predictor+1.25%[1.06, 1.43]70.4%
SENTRY-BD (coupled) vs Residual-64D, same predictor+2.77%[2.57, 2.96]83.6%

For each method, all five training seeds give a positive result (Table 9).

2. Breadth and robustness.

  • On the CD4 confirmation cohort, 61.9% of perturbations improve, and so does every response-magnitude quartile.
  • The gain holds under gene-family resampling, at linear decoder ranks 32, 64 and 128, and with noisy or missing measurements (Tables 4 and 7, Appendix E).
  • SENTRY-BD outperforms 20 random panels by 2.5% on CD4 and by 11.2% on K562→RPE1.
  • 96 genes from the learned ordering reach the error of the 128-gene Residual-64D panel.

3. Same candidate pool (Table 3B). Inside PERSIST's own 1,000-gene pool, SENTRY-BD still outperforms PERSIST-64D on both tasks. On the new test set the margin is +0.92% [0.79, 1.06].

4. Role of bilevel learning (Tables 3A and 5). The stop-gradient variant is also bilevel: it refits the predictor at every step and only drops the second-order term. Learning through the refitted predictor improves on the supervised starting point by 2.6% on CD4 and by 2.56% [2.33, 2.79] on the new test set. With the exact leave-one-out solver, the full gradient also beats the stop-gradient variant on K562→RPE1 (+0.70% at B = 32 and +0.55% at B = 64), and panel overlap across seeds rises from 0.364 to 0.570.

5. K562→RPE1. With the exact leave-one-out solver, SENTRY-BD improves on Residual-64D by +0.90% [0.16, 1.88].

In §5.1, the two earlier CD4 cohorts are now described as held-out cohorts. The protocol, the fixed panels and the per-perturbation results are in the supplementary material.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.