acceptodds
Under review as a conference paper at ICLR 2027

Source Information for Few-Shot Prediction of Donor-Dependent Regulatory Responses

Abstract

Perturbation effects on a transcriptional program vary across donors. Can a few measured responses predict differences between other regulators' effects, and what source data enable this? Such predictions could extend regulatory comparisons within each donor. We model reusable variation through a background-specific transformation of shared perturbation representations. Our response model LHB learns these representations and a nonlinear decoder from source donors, then fits the transformation using target RNA responses. We show that identifiable perturbation relationships can leave nonlinear query responses undetermined, so repeating observed conditions cannot resolve this ambiguity. Across ten induced pluripotent stem cell donors, LHB with ten measured perturbations reduced mean squared error in differences between program effects by 42.0 percent relative to a baseline without target perturbation responses. Cross-line transfer supported partial reuse. Numerical experiments linked source additions to improved prediction after refitting.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Few-shot Learning, Representation Learning, Perturbation Response

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_VyXBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates using limited perturbation measurements from a new donor to predict differential transcriptional responses across unseen perturbations. The authors propose the LHB response model alongside a staged learning framework. The architecture integrates shared perturbation representations, context-specific transformations, and a nonlinear decoder whose input coordinates are recoverable via linear projection. The theoretical analysis characterizes the specific information required to recover perturbation relationships, identify nonlinear query responses, and adapt to target donors. Across diverse experimental settings, evaluations on few-shot prediction and cross-context transfer demonstrate lower prediction errors compared to baselines that omit target perturbation responses. Controlled experiments further analyze how informational coverage across source datasets and target supports influences predictive performance.

Critical strengths

Strengths

  1. The paper clarifies what data to acquire when predictions fail. The theoretical analysis distinguishes the information needed for perturbation relationships, nonlinear responses, and target donor adaptation. Controlled experiments vary coverage of these information types to track changes in prediction error. These results separate measurement noise from informational deficiency, explaining why certain errors cannot be eliminated by simply adding replicate measurements.
  2. Experiments systematically examine the sources of empirical gains. Cell lines from the same donor are held out entirely to prevent donor overlap from inflating generalization estimates. The paired MSE difference against a compressed nonlinear predictor is −0.00197, with a 95% donor bootstrap interval of [−0.00319, −0.00080]. Comparisons across context transformations, decoders, and projection constraints validate key designs, while structural mismatch experiments identify conditions where the method fails.
Critical weaknesses

Weaknesses

  1. Diminishing gains when acquisition metrics are estimated independently. When selection metrics are independently estimated, the average improvement of S7 over S5-cal drops sharply from 0.000300 to 0.000030, with lower coverage of prediction intervals. Comparison intervals for reordering measurements contain zero under a fixed stopping rule. The acquisition tests also rely on a synthetic eight-dimensional benchmark. These results offer weak evidence for improved experimental design and leave a gap for real biological tasks.
  2. Unclear predictive gains over few-shot baselines and narrow evaluation scope. The 42% improvement comes from comparisons against baselines omitting target responses, while paired comparison intervals against HyperMap and PerturbMap contain zero. Experiments are limited to fixed pluripotency programs. Both the linear encoder and orthogonal complement decoder follow existing designs. These findings leave the incremental predictive value and broader generalizability of the model unconfirmed.
Questions for authors

Questions

  1. Confidence intervals and matched comparisons. Do the improvements of S7 over S5-cal and S6-cal hold when computing confidence intervals clustered by donor? How do these comparisons change under identical measurement budgets and under identical stopping rules?
  2. Empirical estimation and gain attribution. How much of the performance gain remains when relying solely on available measurements to estimate coordinates, noise, and expected candidate gains? Please disentangle how much improvement comes from selecting which targets to measure versus deciding when to stop measuring.
Reviewer_SeB2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies prediction of perturbation-response contrasts in a held-out donor after observing a small target support set; the query perturbations have already been observed in source donors. LHB combines shared perturbation embeddings, a target-specific matrix and a nonlinear decoder with a linear inverse projection. Its conditional analysis separates information needed to recover perturbation relationships, determine query-relevant decoder responses and constrain target adaptation. Validation uses a fixed three-gene iPSC score across ten donors, additional biological evaluations, controlled information and misspecification experiments, and numerical acquisition policies. Target support improves prediction at moderate support counts, while the incremental advantage over the strongest adapted comparators remains uncertain.

Critical strengths

S1. The analysis distinguishes linear perturbation coverage from nonlinear decoder information and target-support coverage. The appendix constructions make clear why one kind of information need not substitute for another. The results appear valid under their stated assumptions and give useful guidance about what additional measurements can resolve; they are conditional identification results rather than an end-to-end guarantee for the learned model.

S2. The donor-held-out evaluation keeps each donor’s lines together and includes simple baselines, compressed predictors and useful negative controls. The reported benefit over an independently learned compressed nonlinear predictor is supported by a paired donor interval. The appendix also exposes structural-mismatch failures and cases where support optimization worsens query error, helping define the method’s limits.

Critical weaknesses

W1. The incremental empirical benefit over the strongest adapted comparators is modest and uncertain. The headline 42.0% contrast-MSE reduction uses a control-only reference; adapted HyperMap already obtains 40.0%. LHB’s MSE is 0.03557 versus 0.03678 for adapted HyperMap, with a paired LHB-minus-HyperMap difference of −0.00121 and 95% donor-bootstrap interval [−0.00308, 0.00099] (Table 1; Appendix H). The PerturbMap comparison is also unresolved. These results support the value of target measurements more clearly than a decisive advantage of LHB over these alternatives. The favorable compressed nonlinear comparison is useful, but does not remove this uncertainty. The contribution should be framed around the support-matched comparisons and their uncertainty.

W2. The acquisition evidence is difficult to reconstruct and does not yet establish the usefulness of the described fitted-model pipeline. Appendix K distinguishes a numerical variance-proxy evaluation from a separate implementation computing diagnostics from fitted representations. The supplied library lacks the experiment driver, generators, configurations and action records needed to connect them. Moreover, Table S6 shows a budget-dependent cost–error tradeoff, and independent diagnostics shrink the gain and reduce selected-candidate coverage to about 89–91%. The evidence therefore supports a conditional numerical illustration; stronger practical conclusions need the exact evaluated pipeline and a clearer account of calibration transfer. This is a material gap because choosing useful source measurements is a central contribution.

W3. The paper should introduce the source and target data, the support/query split, and the sizes and learning roles of the model components together. Currently, understanding how these fit together requires repeatedly consulting different sections throughout the paper and the appendix. Figures 4–5 use terms and labels that are not clearly explained in the main text or captions. Readers must consult several appendix sections to understand the experimental settings, acquisition policies, budget units and which model components are updated. Furthermore, the legends as currently presented are unclear. A worked example, notation table and panel-local definitions would substantially improve assessment.

Questions for authors

Q1 — Clarification of the existing comparison. Can the the Table 1 configurations, adapted comparator implementations and tuning budgets, donor/line/guide/cell manifests, final post-filtering support/query masks and per-donor losses be provided? Appendix G shows pre-filtering coverage. Please identify the complete versus masked estimator in each branch and confirm that comparisons use the same scored observations. These details could change the interpretation of the small differences in W1.

Q2 — Clarification of the acquisition evidence. Which code path and data generator produced Figure 5 and Table S6, and which diagnostics came from fitted observations, paid pilots or numerical proxies? Please define the 128 evaluation units and their donor/dependency structure, expose the seven nonzero complete-policy differences (121 ties), and distinguish ranking from stopping effects. Please also report selected-candidate and executed-action coverage together. This would determine how much weight to place on the reported gains and calibration.

Q3 — Clarification of biological validation. What are the 48 additional prediction units and the 48 RNA–OCT4 pairing units, where do the measurements come from, and how are they separated from development? Please supply the unit hierarchy, paired measured records, actual event counts, gates and exclusions. The 175/192 agreement concerns measured RNA and protein, rather than model-to-protein prediction; traceable provenance would determine the weight of this supporting evidence.

These questions request clarification and artifacts for existing results, rather than proposing new biological experiments.

Reviewer_3TNbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LHB for predicting held-out perturbation-induced program effects in a new donor or cell line from a small support panel of measured perturbation responses. LHB combines shared perturbation embeddings, a context-specific transformation, and a nonlinear decoder whose latent coordinates can be recovered by linear projection.

Critical strengths

The paper distinguishes three requirements for cross-context transfer: recovering shared perturbation relationships, covering the query-relevant nonlinear decoder features, and spanning the target support coordinates. The decoder permits exact recovery of its input coordinates by linear projection of the output, allowing full response vectors, rather than only the scalar program score, to constrain the target context transformation.

Critical weaknesses

1.The methodological contribution is largely an integration of existing components. LHB combines a prior coordinate-recoverable decoder with shared perturbation embeddings and a context-specific target adaptation. The paper should clarify what is methodologically new beyond this integration and the accompanying identifiability analysis.

2.The empirical evidence is limited. The main evaluation is restricted to a predefined pluripotency program in one iPSC atlas, and LHB’s paired intervals against HyperMap and PerturbMap include zero at the main support budget. These results do not yet establish a consistent advantage across contexts or readouts.

Questions for authors

1.Could the authors report donor-wise paired errors and support-panel stability to clarify the advantage over HyperMap and PerturbMap?

2.Does LHB retain its advantage after excluding the POU5F1 query, which directly targets one of the genes defining the primary readout?

3.Can the held-out predictions be validated against known regulator roles or the independent OCT4 assay? The current RNA-OCT4 comparison concerns agreement between measured RNA and protein contrasts, rather than prediction accuracy.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.