acceptodds
Under review as a conference paper at ICLR 2027

HICON: Hierarchical Technical Conditioning for Contrastive Cell Painting Batch Correction

Abstract

Batch effects remain a major challenge in analyzing Cell Painting data. They introduce hierarchical and spatially structured technical variations that are entangled with perturbation-induced biological variations. Existing methods rely on global feature alignment or normalization over predefined batch groups. However, these approaches may fail to capture the heterogeneous technical variations, which can disrupt the biological signals. In this paper, we formulate Cell Painting batch correction as a conditional representation learning problem, aiming to maximize biological information given latent technical states. We propose HICON, a hierarchy-aware technical conditioning framework for contrastive biological representation learning. HICON uses a two-branch architecture to disentangle biological and technical representations. The technical branch learns a hierarchical technical embedding with a confidence-aware mechanism to capture sample-specific technical variation beyond metadata labels. The biological branch leverages perturbation replicates to define positive and negative samples and adaptively reweights the contrastive differentiation between the sampled pairs according to their inferred technical similarity, thereby emphasizing perturbation variation under comparable technical conditions. Extensive experiments across JUMP-CP, BBBC, CPG0036, and RxRx1 show that HICON achieves strong batch correction, preserves biological structure, and improves downstream biological utility.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cell Painting, Batch Effect Correction, Representation Learning, Contrastive Learning, Technical Variation, High-Content Screening, Morphological Profiling

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_2JrMOpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes HICON to address hierarchical technical variations across batches, plates, and spatial positions in Cell Painting data. The framework couples a technical branch with a biological branch: the former captures technical variations through hierarchical prototype assignments and confidence weighting, while the latter contrasts perturbation replicates weighted by technical similarity. A CLUB-style regularizer further decouples the two representations, effectively casting the objective as preserving biological signals conditioned on latent technical states. Extensive evaluations across four major benchmarks show improvements in batch correction, biological fidelity, and downstream retrieval across diverse settings.

Critical strengths

Strengths

  1. Practically grounded problem formulation with well-aligned hierarchical modeling. HICON jointly addresses batch, plate, and spatial effects, reducing reliance on unreliable metadata through soft assignments and confidence weighting. Hierarchical technical prototypes, biological supervision based on replicate perturbations, and technical similarity weighting provide a clear modeling framework for simultaneously correcting technical differences and preserving biological signals. Component ablations also support the contribution of several designs to performance.

  2. Extensive evaluation demonstrating robustness and generalization. The experiments cover multiple datasets and feature sources, examining scenarios such as unseen perturbations, unseen imaging sources, reduced replicate counts, and metadata corruption. Combined with synthetic experiments with known latent structures, these results provide valuable evidence for assessing the method's effectiveness and scope of applicability. Improvements across multiple individual biological metrics also support its ability to retain biological signals.

Critical weaknesses

Weaknesses

  1. Discrepancy between overall score definition and reported values. The overall score is defined in the text as an unweighted average of three technical and four biological metrics. However, calculating the arithmetic mean from the seven printed values for HICON in Table 32 and Table 34 yields 0.57143 and 0.61714, whereas the reported numbers are 0.56 and 0.63. The reported figures instead match an equal weighting between the technical and biological group means, which alters the implicit weighting across individual metrics. This inconsistency undermines the auditability of the composite metric and affects comparative claims. The authors should standardize the aggregation protocol, clarify the handling of missing kBET values, and recompute the affected results.
  2. Theoretical contradiction in the mutual information lower bound and incomplete CLUB training details. While Eq. (28) sums over multiple positive samples, Eq. (30) continues to use a constant log⁡K\log K. When technical conditions are identical, biological representations are constant, and the positive sample count m>1m > 1, the loss evaluates to log⁡(K/m)\log(K/m). The stated bound then yields log⁡m>0\log m > 0 while the true mutual information is zero, directly contradicting Eq. (30).
Questions for authors

Questions

  1. Clarification on aggregation protocol and necessary corrections: What exact formula was used to compute the Overall scores in Tables 32 and 34? Please recompute the overall results in both the main text and appendix tables in accordance with the protocol stated in Appendix D. In addition, please clarify how datasets lacking kBET were normalized, and whether method rankings and comparative conclusions remain unchanged. Providing per-run individual metrics alongside aggregated values would greatly facilitate auditability and verification.

  2. Theoretical derivation clarifications: How is KK in Eq. (30) formally defined in the presence of multiple positive samples? How does the framework reconcile the constant-representation counterexample, as well as the discrepancy between selecting negative samples based on perturbation labels versus sampling from the true conditional marginal distribution? Without formal theoretical guarantees, could this section be explicitly repositioned as a heuristic motivation instead of an unestablished lower bound?

Reviewer_Ms6hOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a new deep-learning based method for learning a representation of Cell Painting data. It consists of two branches: technical and biological, which aim to disentangle the technical from the biological differences in the cell.

Critical strengths

The paper introduces an interesting idea of learning representations of single-cell data by utilizing a two-branch approach and contrastive learning approach. They show good results in benchmarks.

Critical weaknesses

My main concerns are about a clean introduction of the problem formulation, and surprisingly good performance in the presence of no batch information labels. I would consider raising my score if the authors address these two main weaknesses:

  • The paper needs to explain more clearly what it is referring to as hierarchical technical effects. Fig 1. is missing a description of what it is trying to depict concretely, and, additioanlly, the hierarchical technical effects need a more formal definition.
  • Fig. 6. seems to indicate that HICON, without batch label information (noise fraction 1.0), outperforms any other method by a large margin (0.65 to 0.55). How can these performance gains be explained when utilizing random labels?

Minor comments:

  • The in-text citation formatting should use parentheses, e.g. lines 35-39.
  • G.1 COMPLETE TRAINING PROCEDURE in the appendix is an empty chapter.
Questions for authors

Main question:

  • See Weakness 2: How can these performance gains be explained?

Other questions:

  • Will the authors make the experiments / source code available?
  • HICON introduces many hyperparameters. How robust is it to changes in these, and how should practitioners set these hyperparameters?
  • How were the hyperparameters of the baselines tuned?
Reviewer_d4bUOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HICON, a hierarchical technical conditioning framework for Cell Painting batch correction. It learns separate biological and technical representations using confidence-aware hierarchical embeddings and technically weighted contrastive learning to reduce batch effects while preserving biological information.

Critical strengths

The paper is generally well-written, well-motivated, and addresses an important problem in Cell Painting analysis. The authors conduct extensive experiments across multiple datasets, including JUMP-CP, BBBC, CPG0036, and RxRx1, demonstrating strong performance in batch correction, biological preservation, and downstream biological tasks.

Critical weaknesses

Fairness of baseline comparisons and biological supervision. HICON uses perturbation labels for supervised contrastive learning, whereas many compared batch-correction methods do not receive equivalent biological supervision. Although Tables 2 and 10 include replicate-based contrastive ablations, these experiments are limited to JUMP-CP Scenario 4. More comprehensive comparisons under matched biological supervision across datasets would help establish the contribution of the proposed technical conditioning.

Insufficient validation of biological–technical disentanglement. A central assumption of HICON is that the learned technical representations capture technical variation independently of biological information. However, Appendix E evaluates this assumption using synthetic data with orthogonal biological and technical factors, which may not reflect their entanglement in real Cell Painting data. More direct evaluations of biological information leakage into technical embeddings, particularly when perturbations are correlated with experimental batches, would strengthen the claimed disentanglement.

Relation to prior work. Several components of HICON build upon established techniques, including hierarchical technical modeling, Gaussian mixture-based representation learning, and contrastive learning. In particular, cpDistiller [1] already addresses hierarchical batch and spatial effects using contrastive and domain-adversarial learning. Although HICON introduces confidence-aware technical proxies for pairwise contrastive weighting, the authors could better establish the advantages of this formulation over simpler metadata-based or existing contrastive conditioning strategies.

Reference

[1] Yan, Chengwei, et al. "Triple-effect correction for Cell Painting data with contrastive and domain-adversarial learning." Nature Com 2025

Questions for authors
  1. How do the authors verify that the technical embeddings do not retain perturbation-specific biological information?
  2. Why does the contrastive objective emphasize technically similar pairs rather than aligning replicates across different batches?
  3. How much of HICON's improvement comes from perturbation-label supervision versus the proposed technical conditioning?
Reviewer_eCMcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses batch correction of well-level Cell Painting profiles and proposes HICON, a two-branch model trained on pre-extracted features. A technical branch learns a hierarchical (batch, plate, row/column) technical embedding as a proxy for the unobserved technical state, using metadata as weak supervision with a confidence mechanism that down-weights wells whose features disagree with their metadata. A biological branch is supervised contrastive learning on perturbation identity, with pairwise terms reweighted by technical similarity from the proxy, and a CLUB term reduces the mutual information between the two branches. The paper frames this as maximizing perturbation information conditioned on the latent technical state. Evaluation follows the protocol of Arevalo et al. (2024) on five JUMP-CP scenarios, BBBC022/036, five CPG0036 subsets, and four RxRx1 cell lines, with perturbation-level train/test splits, against 13 baselines. HICON reports the best aggregate score in nearly every setting, gains on downstream retrieval (Fig. 5), robustness to batch-label corruption (Fig. 6), better row/column LISI within a batch (Fig. 7), and ablations on a 30% subset of JUMP-CP Scenario 4 (Tables 1, 2, 10, 11). A synthetic study (Appendix E) shows the learned proxy predicts ground-truth technical states better than corrupted metadata.

Critical strengths
  1. The problem framing is reasonable. Treating batch, plate, and row/column effects as a hierarchy, treating metadata as weak rather than exact supervision, and posing correction as conditional information maximization rather than invariance are sensible positions, and the within-batch triple-effect evaluation (Fig. 7) and the metadata-corruption protocol are useful additions.

  2. The evaluation is broad in coverage. The paper provides four datasets, three feature types, multiple settings, full per-metric tables, a rank-based aggregation check, a hyperparameter sweep with three seeds, etc.

Critical weaknesses
  1. The main comparison is partly confounded by supervision, and the paper's own ablation attributes most of the gain to it.

    • HICON's biological branch is weighted SupCon on perturbation labels. But it seems that none of the compared baselines take advantage of such supervision. The biological metrics (label LISI, Leiden ARI/NMI, replicate mAP, graph connectivity by biological label) directly reward an embedding trained to collapse replicates of the same perturbation, leading to potential hacking, and five of the seven metrics in the aggregated score are computed against perturbation labels, including graph connectivity, which sits in the batch-correction group. The disjoint-perturbation split does not remove this asymmetry, since a metric-learned embedding transfers to held-out perturbations from the same library. The large jumps that drive the headline (Table 13, mAP 0.95 vs. others at most 0.66) are a potential signature of a supervised advantage.
    • Table 2(A) and Table 10 also support this point. In the ablation setting, replicate-based SupCon alone moves Overall from 0.43 to 0.56 (30% data), 0.53 to 0.62 (50% data), 0.54 to 0.64 (75% data), while the full model adds 0.08, 0.05, and 0.04 on top of that (Table 10), and that increment is measured on one scenario, one feature type, with no error bars. The SupCon-only variant is absent from comparison with other baselines, so the reader cannot see how much of the reported advantage on the other 12 settings survives once supervision is provided.
    • More potential baselines are missing besides plain SupCon on perturbation labels. These include SupCon after a standard correction (Harmony or Sphering + Harmony, as recommended by Arevalo et al.), SupCon with batch-balanced sampling, and potentially scANVI with perturbation labels, which the related work cites but does not evaluate. Without these, the claim of strong batch correction is not separable from the fact that HICON is supervised.
  2. The formulation includes conditioning on T and independence from T at the same time, which may create competition. Additionally, the method contains soft proxies whose designated roles are neither guaranteed by construction nor evaluated.

    • The stated objective is max I(Z^{bio}; P | T) in Sec. 3.1, which by itself does not ask for Z^{bio} to be disentangled from T. Appendix I.6 motivates the CLUB term from a particular variational lower bound (Eq. 34) in which a -I(U; C) term appears, and reads it as a reason to minimize I(Z^{bio}; Z^{tech}). However, minimizing I(Z^{bio}; Z^{tech}) is actually an additional invariance prior, of the same kind as the global alignment approaches that Sec. 1 argues disrupt biological structure, applied to a learned proxy instead of predefined batch groups, as it pushes the biological and technical branches to be independent, conflicting with the intuition of conditioning on T. The paper needs to clarify which of the two it is proposing.
    • Following the intuition in the first point, the two mechanisms built on the proxy are mutually exclusive in the limit, at their respective optima. The pair weights beta_ik are justified as an approximation to sampling from the conditional p(v | c). If the CLUB term succeeds and Z^{bio} is (nearly) independent of Z^{tech}, then p(v | c) = p(v), and the conditional InfoNCE in Eq. 24 reduces to plain InfoNCE with weights beta_ik that are (nearly) independent of Z^{bio}. Conversely, if the reweighting matters, Z^{bio} still depends on Z^{tech} and CLUB has not done its job. Each component is useful only in the regime where the other is suboptimal. Appendix O concedes that a loose CLUB bound may leave residual technical information, which places the trained model somewhere in between, but the paper never diagnoses where. The core problem, in my view, is that the paper wants to disentangle technical and biological factors, yet still assumes, from experimental prior knowledge, that biological effects should be estimated under similar technical conditions, and it offers no insight into how to handle the tradeoff.
    • The confidence mechanism has no basis for its split. A one-dimensional two-component GMM on s_i in [0, 1] will partition the scores whether or not a genuine low-consistency mode exists, and the paper gives no evidence that s_i is bimodal on real data. The reference it follows (Huang et al. 2023) applies the same mixture to per-sample losses under the small-loss premise that noisy labels produce large losses, and s_i here is a monotone transform of the per-sample cross-entropy, so the same premise is being imported. With correct labels, a low s_i means only that Z^{tech} has not yet fit that well, and the mechanism reduces supervision on hard samples. The paper asserts that metadata is coarse and noisy (Sec. 3.3, Appendix H.2, Fig. 1) but gives no quantitative evidence of it on real data, such as a histogram of s_i.
  3. The ablations do not properly show component contribution in their desired scenario.

    • Every ablation row is run on JUMP-CP Scenario 4 with clean metadata, where confidence weighting and soft targets, which exist to handle noisy or incomplete metadata (Appendix M.4), have no label noise to act on. Removing confidence weighting nevertheless costs 0.09 Overall and 0.13 Batch (Table 2(B)), so the mechanism is possibly changing the supervision of correctly labeled wells, and the paper reports nothing about how many wells are down-weighted or what the down-weighting reflects.
    • The hyperparameters that govern the proxies' behavior over training are not specified and not tested. The lower bound eta_l(tau) on the confidence weight (Sec. 3.3, Eq. 15) is called training-dependent and never given. The momentum m_l(tau) is given only as start and end values per level (Appendix G) without the form of the increase. The switch to confidence weighting at 50% of training, the 30% warm-up before reweighting, the two temperatures, the augmentation noise and dropout, and the replicate-count reweighting are all stated once in Appendix G and absent from Table 11. Since the component-group correspondence is created during the warm-up and the confidence weighting only begins after it, these schedules are the parameters that shape the technical prototype training.
    • Fig. 6 and Appendix E do not test the proxy either. In Fig. 6, the biological loss itself uses only perturbation labels and the clipped pair weights, so corrupted batch labels reach the bio score only indirectly, through the technical branch they supervise and then through beta_ik and the CLUB term. The bio score is therefore largely insulated from corruption (HICON drops 0.6% at 50% corruption, the annotated point, and about 5% at 100%), whereas baselines align to whatever labels they are given and must fall. The Baseline row and SupCon without a technical branch are both absent. In Appendix E, the technical signals are constructed in subspaces orthogonal to the biological signal and added to it, so a metadata-free reference such as k-means or a GMM on the raw features is needed to tell whether the metadata supervision helps or hurts relative to the features alone. The paper reads 0.794 against the metadata's 0.599 as recovering structure beyond the labels, but the same table shows the proxy falling from 0.985 to 0.794 as labels are corrupted, which is the proxy following the labels. Which reading holds depends on what the features alone can give, and that reference (k-means or a GMM on the raw features) is not reported.
    • None of the main-text results carry error bars. Tables 1 and 2 report single numbers, and Figs. 3, 5, and 6 show single points, while Table 11 shows run-to-run standard deviations of 0.01-0.06 for the full model on the same ablation setting. The differences the ablations rest on (0.03 for CLUB in Table 2(C), 0.06-0.12 on the within-batch metrics in Table 1) are of that order, so none of them can be judged from what is reported.
  4. Several reported numbers point to potential evaluation errors that need clarification and discussion.

    • Table 16 (Scenario 4, CellProfiler) reports Leiden ARI of 0.92 for Scanorama, 0.89 for DESC, and 0.89 for cpDistiller alongside Leiden NMI of 0.09, 0.12, and 0.06. An ARI near 0.9 requires high agreement with the labels and is hard to reconcile with an NMI near 0.1.
    • Deep baselines frequently score below the Baseline row, which the paper does not define and which I assume to be the uncorrected preprocessed features as in Arevalo et al. (2024). In Scenario 2 with DeepProfiler features, cpDistiller reaches Overall 0.40 against Baseline 0.50 (Table 19); in Scenario 4 with CellProfiler features, scBCN, DESC, and cpDistiller are 0.49, 0.41, and 0.48 against 0.51 (Table 16), and on BBBC022, all four deep baselines are near or below Baseline (Table 22). cpDistiller is the one baseline designed for exactly the triple-effect structure this paper targets. A method that is worse than doing nothing is concerning, and the paper gives no tuning protocol, hyperparameters, or sanity checks for any baseline.
    • The ablation tables disagree about what the full model is, and the appendix text reads as if CLUB were present where the numbers say it is not. The top rows of Table 2(A) and Table 2(B) are identical (0.64, 0.58, 0.60) and equal the no-disentanglement row of Table 2(C) and the full-hierarchy row of Table 1, while the full model with CLUB is (0.68, 0.60, 0.63) in Table 2(C), Table 10, and Table 11, and Table 10 lists the no-CLUB model at (0.64, 0.58, 0.61) rather than 0.60. So Tables 1, 2(A), and 2(B) seem to have been run without CLUB. Appendix M.4 nevertheless states that in the technical-branch ablation the disentanglement loss is kept unchanged. No reason is given for ablating on a reduced model, and the choice is not neutral, since point 2 shows the pair weights matter only when Z^{bio} still depends on Z^{tech}, which is the regime a model without CLUB is in. Table 10 re-checks the biological-branch ordering with CLUB, but the hierarchy (Table 1) and technical-branch (Table 2(B)) conclusions are carried over to the full model without such a check.
    • In Scenarios 3 and 5 with CellProfiler features, HICON is at 0.41 vs. 0.40 and 0.38 vs. 0.36 Overall, with label LISI saturated at 1.00 and ARI at or below 0.07 for every method (Tables 15, 17) and no error bars. The text presents this as HICON remaining among the strongest, but it tends to be within noise on saturated metrics.
  5. The soft target seems equivalent to sample-wise weighting. The other half of the interpolated target in Eq. 16 is the model's own stop-gradient output, which contributes no gradient, so the soft target reduces to scaling the hard-label cross-entropy by w_i. The confidence mechanism therefore consists of weighting two things by the same w_i, the loss and the prototype update in Eq. 3. Read this way, Table 2(B) attributes 0.05 to weighting the prototype update and 0.04 to weighting the loss, with no separate contribution from feature-inferred assignments, which is not how the method is presented.

Questions for authors
  1. The replicate-based contrastive variant of Table 2(A) is the natural reference for separating the effect of supervision from the effect of the technical branch, but it appears only in the ablation setting. Could the authors report this variant, with the same training configuration as HICON, on the benchmark settings of Fig. 3, or at least on the five JUMP-CP scenarios? The size of HICON's margin over this variant, rather than over the unsupervised baselines, is what I would use to judge the contribution of the proposed components.

  2. Which objective is the paper proposing, maximizing I(Z^{bio}; P | T), which by itself does not constrain I(Z^{bio}; T), or making Z^{bio} independent of T through CLUB? If the former, what motivates CLUB beyond the -I(U; C) term of the particular bound in Eq. 34 and the informal statement in Sec. 3.2 that it promotes a clean conditioning proxy? If the latter, how is the method different in principle from the alignment approaches criticized in Sec. 1?

  3. On the ablation setting with clean metadata, at the end of training, what fraction of wells receive a confidence weight below 0.5 at each level, how often does the feature-inferred assignment argmax of gamma_i agree with the metadata label, and how does the pair weight correlate with the same-batch indicator and with the same-perturbation indicator?

  4. Were Tables 1, 2(A), and 2(B) run without the CLUB term? The identical rows across these tables and the 0.63 of the full model in Tables 2(C), 10, and 11 suggest so, while Appendix M.4 states the disentanglement loss is kept.

  5. Table 16 lists Leiden ARI of 0.92, 0.89, and 0.89 for Scanorama, DESC, and cpDistiller next to NMI of 0.09, 0.12, and 0.06, which seem inconsistent with each other. Could the authors comment on this? In the same spirit, could the authors describe how cpDistiller was configured, given that it falls below the Baseline row in Tables 16, 19, and 22, which metadata columns each baseline received at training time, and what the Baseline row in Fig. 3, Fig. 5, and Tables 13-34 denotes?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank the Area Chair and reviewers for their careful assessment and constructive feedback. We are encouraged that the reviewers recognize the practical importance, hierarchical modeling, and extensive evaluation of HICON. We have addressed their main concerns through additional experiments, methodological clarifications, and manuscript revisions:

  • Supervision fairness: We add supervision-matched comparisons with SupCon and scANVI across seven benchmark settings. HICON consistently outperforms both methods. Additional comparisons with SupCon + Harmony and SupCon + Seurat CCA further demonstrate the benefits of joint technical conditioning and biological representation learning (Reviewers eCMc and d4bU).

  • Technical proxy validation: We provide compound-leakage analyses, pair-weight correlations, and representation-similarity evaluations. Confidence diagnostics further clarify the role of feature–metadata consistency (Reviewers eCMc, d4bU, and Ms6h).

  • Information-theoretic formulation: We clarify CMI as the primary objective and CLUB as an auxiliary regularizer, including their interaction under perturbation–batch dependence. We also revise the multi-positive theoretical derivation and provide CLUB implementation details (Reviewers eCMc and 2JrM).

  • Evaluation consistency and reproducibility: We clarify the Overall score aggregation, correct reporting inconsistencies, and verify that HICON remains ranked first across all 20 benchmark settings under both aggregation protocols. We further document baseline configurations, hyperparameter sensitivity, and confidence/momentum scheduling (Reviewers eCMc, Ms6h, and 2JrM).

We appreciate Reviewer eCMc's positive feedback on the additional comparisons and theoretical clarifications. We also appreciate Reviewer 2JrM's recognition that our revisions addressed their main concerns and their subsequent decision to increase the rating.

We thank the Area Chair and reviewers again for their constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.