acceptodds
Under review as a conference paper at ICLR 2027

What to Learn and What to Trust: Multimodal Learning with Reliable Supervision for Bioactivity Prediction

Abstract

Cell Painting phenotypes and molecular structures provide complementary views of compound activity, but learning from them is complicated by replicate variation, non-equivalent structure–phenotype relationships, and sparse assay supervision. Existing multimodal approaches typically align paired molecular and morphological observations, implicitly encouraging agreement between the two modality-specific representations. We propose BioCoReT, a two-stage framework that addresses what to learn from rich multimodal observations and what to trust under sparse assay supervision. First, BioCoReT performs contrastive learning across fused views of the same compound. It constructs biology-aware multimodal views from matched controls and Cell Painting replicates, and learns consistency across these compound-level views while preserving modality-specific information. Building on these representations, assay-specific specialists learn task-specific decision boundaries from the available labels and generate candidate predictions for unlabeled compound-task pairs. BioCoReT then evaluates each candidate using complementary evidence from teacher confidence, molecular and morphological topology, and replicate consistency, transferring only reliable task-specific knowledge to a unified multi-task student. BioCoReT consistently outperforms existing methods across three public Cell Painting benchmarks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Phenotypic drug discovery, Cell Painting, Multimodal representation learning, Reliable pseudo-labeling, Multi-task learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_s8ReOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BioCoReT, a two-stage framework for bioactivity prediction from molecular structure and Cell Painting data under sparse assay labels. Stage 1 learns a task-agnostic compound representation through contrastive learning over fused multimodal views built from different replicates and matched with controls, rather than aligning the two modalities directly. Stage 2 trains assay-specific teachers, filters their predictions into pseudo-labels using teacher confidence, molecular and morphological topology, and replicate consistency, and uses them to train a multitask student. The authors report consistent gains over existing methods on three public Cell Painting benchmarks, along with an interesting reduced-label ablation.

(Edit: Most concerns addressed during rebuttal, score updated to a 3 accordingly).

Critical strengths

The paper addresses an important and challenging problem: improving bioactivity prediction with limited and sparse data. Progress here has clear practical value.

The design is adapted to real properties of the data: replicate structure, imperfect alignment between modalities, and sparse, heterogeneous labels. Its assumptions are generally reasonable, which makes the approach promising despite its complexity.

Critical weaknesses

The main results and ablations report no standard deviations across seeds or cross-validation folds, and hyperparameter tuning is not described. Without these, the improvements over the baselines cannot be properly judged.

Lack of details in the main text and no indication of code publication. The main text lacks key method details: architectures, whether the pretrained encoders are frozen, what the teacher model is, and which components are trained. This makes the setup hard to understand and reproduce. There is no indication that code will be released. Given the complexity of the pipeline, a code release is necessary for the work to be reproducible and useful to the community.

Addressing these two would make me inclined to increase my score to weak accept.

Questions for authors

Can you report standard deviations across seeds or folds for the main results and ablations? Is cross-validation used, and how were hyperparameters tuned?

Are the datasets split randomly by compound or by structural similarity, i.e., do the splits require scaffold hopping?

Which components are trained versus frozen? What are the architectures and training data of the pretrained image and molecular encoders, and are they built for Cell Painting data?

What model is used for the assay-specific teachers? How is overfitting avoided with a few hundred or fewer labels per assay? How are the reliability weights and pseudo-label thresholds chosen?

Which representations are used for the topology and replicate consistency scores? Are they control-adjusted or normalized, and do they come from trained or frozen extractors?

Does the reliability filtering mainly select very easy examples as pseudo-labels?

How does the unimodal-to-multimodal alignment loss avoid reducing the representation to the information shared by both modalities, the problem raised with activity cliffs and scaffold hopping?

Are the hidden-label accuracies in Section 4.5 balanced or raw, given the class imbalance?

In Section 3.1, does “matched untreated control image for replicate R” mean the DMSO control on the same plate?

Is “JUMP 29” in Figure 4 the same dataset as CPG0016? If so, please use consistent naming.

Reviewer_E3RnOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents BioCoRet, which is a two-stage framework designed for compound representation learning from multimodal data and makes use of pseudo-labels when there are only limited amounts of assay labels.

The first stage uses matched controls and experimental replicates to learn compound representations. Contrastive learning brings different views of the same compound closer together, and additional objectives help preserve information from individual modalities. In the second stage, assay-specific specialists predict pseudo-labels for missing compound-task pairs, and the predictions are filtered based on confidence, molecular and morphological topology, and replicate consistency before being used to train a multi-task student model.

BioCoRet is evaluated on 3 benchmarks and the model performs well across all three. The ablation studies also suggest that both stages contribute to the results.

Critical strengths

There are two critical strengths to this paper/model. One in strategy design and one in evaluation and performance.

  1. The multimodal representation learning takes into account the nature and characteristics of cell painting data. Instead of directly matching molecular and morphological representations, this model combines information from both modalities with experimental replicates and matched controls to construct different views for contrastive learning. This design makes sense and better reflect the nature of such data: two modalities are complementary, and they not necessarily share the same similarity structure. And the cell painting data might contain experimental variation. In Figure 3(b) (the table), it shows that the full stage 1 representation achieves an AUROC of 0.716 while simply concatenating two feature vectors achieves 0.671 (which is a solid improvement in this case). And the performance also drops when CCRE or unimodal-to-multimodal alignment is removed. This suggests that the representation learning strategy brings benefits beyond simply combining features from pretrained models.

  2. The evaluation is thorough and clearly presented. It considers different levels of label availability, and the quality of pseudo-labels. The 3 benchmarks the authors have evaluated on have different number of assay tasks and level of label availability. The results listed in Table 1 shows that the proposed model is compared against 11 different baseline models and it shows SOTA performance across three benchmarks. Table 2 shows that the model is able to maintain the best across 10 models that are compared under different fractions of observed assay labels, even when fewer labels are available.

I also really appreciate that the authors evaluate the reliability of the pseudo-labels rather than relying only on the final prediction results, and experiments on how changing the selection threshold affects pseudo-label quality and coverage. These experiments help support the proposed pseudo-label strategy.

Critical weaknesses

The downstream contribution of the proposed reliability filtering is not fully isolated.

In Figure 4(b), the authors compares different pseudo-label filtering strategies (confidence-only, confidence + molecular topology, and full CTPL). It supports that the full CTPL selects better pseudo-labels than confidence-only filtering. And Figure 4(c) shows that the unified student generally outperforms the teacher. There is a gap in between: there is no evaluation/evidence of how each filtering strategy affects the final student model.

I think it would be interesting and useful to compare a student trained with confidence-only pseudo-labels, confidence combined with molecular and morphological topology, and with the full CTPL, while keeping the training setup, architecture, and pseudo-label coverage the same.

Questions for authors
  1. Can you evaluate how much of stage 2's improvement comes from CTPL compared to confidence-based and confidence+topology-based pseudo-labeling? You can find more details in the weakness section.

  2. Can you report the pseudo-label accuracy in positive and negative predictions separately? As bioactivity datasets can be highly imbalanced, often with more negative/inactive compounds, it would be interesting to see whether the high overall accuracy is (partly) due to a larger ratio of negative labels.

  3. Can you evaluate how much of the improvement comes from selecting better pseudo-labels versus weighting them by their reliability scores? For example, considering evaluating the student's performance if all selected pseudo-labels were assigned an equal weight. Also, how much do the reliability scores actually vary among the selected pseudo-labels (given that the threshold is set to 0.8)?

Reviewer_mZy2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BioCoReT, a two-stage multimodal learning framework for bioactivity prediction under sparse and heterogeneous assay supervision. The framework aims to address two challenges: learning informative compound representations from molecular structures and Cell Painting phenotypes, and identifying reliable supervision from incomplete assay annotations. Experiments on three Cell Painting benchmarks demonstrate improvements over several existing molecular, morphological, multimodal, and semi-supervised learning approaches. Additional ablation studies investigate the contributions of the two stages and the reliability of selected pseudo-labels.

Critical strengths
  1. The paper addresses an important challenge in phenotypic drug discovery: integrating molecular structures and cellular morphology under sparse and incomplete bioactivity supervision.

  2. BioCoReT achieves the best reported performance across three public Cell Painting benchmarks in the evaluated metrics. The results under reduced label supervision further suggest that the framework can benefit from unlabeled compound–task pairs.

Critical weaknesses
  1. The proposed method does not directly address the Activity Cliffs and Scaffold Hopping highlighted in the introduction. In particular, it still treats different compounds as negative samples, potentially pushing apart structurally distinct compounds with similar cellular phenotypes, which contradicts the authors' motivation.

  2. Although the authors introduce control-conditioned representation encoding (CCRE) and contrastive learning across experimental replicates, these mechanisms do not necessarily eliminate systematic batch effects. It remains unclear whether the model can distinguish technical noise from meaningful biological variation.

  3. The proposed CTPL reliability score combines teacher confidence, molecular and morphological neighborhood agreement, and replicate consistency through a weighted average. However, the rationale behind this design is insufficiently justified, particularly since these signals may be correlated and neighborhood-based agreement may be unreliable for Activity Cliffs or highly sparse assay tasks.

  4. Several closely related studies, including InfoAlign (ICLR 2025) and ACANet (Nature Communications, 2026), have investigated overlapping challenges. Although MINER is included as a baseline, its methodological relationship to BioCoReT is insufficiently discussed, while InfoAlign and ACANet are not included in the experimental comparisons. More comprehensive comparisons and discussions are needed to clarify the novelty of the proposed approach.

  5. The paper lacks targeted evaluations for Activity Cliffs and Scaffold Hopping, making it difficult to validate its central motivation. Additionally, since pseudo-labels are generated at the replicate level, the authors should clarify how potentially conflicting pseudo-labels from different replicates of the same compound are handled.

Questions for authors
Questions For Authors
  1. How does the proposed method explicitly address Activity Cliffs and Scaffold Hopping, given that different compounds are still treated as negative samples in contrastive learning?

  2. How does the model distinguish technical noise from meaningful biological variation across experimental replicates? Can the authors provide evidence that CCRE and replicate consistency effectively mitigate batch effects?

  3. What justifies the weighted aggregation of the four reliability signals in CTPL? How robust is this strategy when the signals conflict or neighborhood labels are highly sparse?

  4. Could the authors further clarify the methodological differences from closely related studies, such as InfoAlign, MINER, and ACANet, and provide additional comparisons where applicable?

  5. Could the authors provide targeted evaluations for Activity Cliffs and Scaffold Hopping? Additionally, how does the method handle potentially conflicting pseudo-labels across different replicates of the same compound?

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 of bioactivity prediction under sparse supervision (mZy2, s8Re), the biology-aware multimodal representation learning design (E3Rn, s8Re), and the strong performance across three benchmarks (mZy2, E3Rn). We have addressed their main concerns through additional experiments, methodological clarifications, and manuscript revisions:

  • Activity cliffs and scaffold hopping: We add targeted scaffold-hopping retrieval and activity-cliff evaluations, where BioCoReT outperforms the compared baselines. We further clarify how fused-view contrastive learning preserves complementary molecular and morphological information and add comparisons with InfoAlign and ACANet (Reviewers mZy2 and s8Re).

  • Technical noise and biological preservation: We provide cross-plate replicate retrieval experiments and an additional CCRE ablation, demonstrating improvements beyond both raw DINO and Stage 1 without CCRE. These results further support CCRE's contribution to mitigating technical variation (Reviewer mZy2).

  • Pseudo-label reliability and downstream contribution: We add student-level CTPL ablations, class-wise pseudo-label precision, and reliability-weighting comparisons. Additional evaluations under conflicting reliability signals and sparse neighborhoods further support CTPL's robustness (Reviewers E3Rn, mZy2, and s8Re).

  • Reproducibility and methodological clarity: We clarify compound-level splits, pretrained encoders, trainable components, teacher architectures, and hyperparameter tuning. We also provide variability results for the main experiments and ablations, alongside additional implementation and evaluation details (Reviewers s8Re and E3Rn).

We appreciate Reviewer mZy2's acknowledgment that the additional experiments address several original concerns and strengthen the empirical evidence for BioCoReT. We also appreciate Reviewer s8Re's recognition that most concerns have been addressed and their subsequent decision to increase the rating. We further appreciate Reviewer E3Rn's positive assessment of our multimodal learning design and comprehensive evaluation.

We sincerely thank the Area Chair and all reviewers 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.