acceptodds
Withdrawn from ICLR 2027

From Cellular Responses to Pharmacological Domains: Multimodal Zero-Shot Drug Representation Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal drug discovery enables drug representation learning beyond chemical structure by incorporating cellular responses such as gene expression and cell morphology. However, direct fusion and instance-level contrastive alignment may mix mechanism-related signals with modality-specific noise and incorrectly separate structurally dissimilar but biologically related compounds. This limitation can obscure transferable mechanism patterns required for predicting the properties of unseen compounds. We introduce PMRD, a pharmacological response domain guided framework for multimodal zero-shot drug property prediction. PMRD separates mechanism-consistent factors from modality-specific information and constructs a consensus response domain across three modalities. Mechanism candidate augmentation identifies locally stable factors, while retrieval-geometry attribution dynamically reweights the alignment and augmentation objectives according to whether their updates preserve inter-drug discriminability. This feedback suppresses training signals that conflict with mechanism-discriminative retrieval. PMRD further combines complementary representations through reliability-aware multiview retrieval. Experiments on public datasets show improved zero-shot property prediction and more biologically coherent drug neighborhoods. Hard negative analysis further indicates fewer conflicts between structurally dissimilar but response-related compounds. These results support PMRD as an effective framework for mechanism-aware multimodal drug representation learning

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multimodal Drug Discovery, Zero-Shot Drug Property Prediction, Drug Representation Learning, Pharmacological Response, Mechanism-Aware Learning, Multiview Retrieval, Contrastive Alignment, Reliability-Aware Fusion.

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_55jzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PMRD, a multimodal framework for zero-shot drug property prediction using molecular structures, gene expression profiles, and cellular morphology. The method separates mechanism-consistent and modality-specific representations through cross-modal response modeling, and introduces stability-guided optimization to improve representation robustness. A reliability-aware retrieval strategy is further developed to combine predictions from multiple representation views. Experiments on ChEMBL2K and Broad6K demonstrate promising results compared with existing multimodal and zero-shot learning methods.

Critical strengths

The paper addresses an important problem in drug discovery. Integrating molecular structures with cellular responses is a meaningful direction for improving drug representation learning and property prediction.

Critical weaknesses

Top concern: Separating shared information from modality-specific variations during multimodal alignment has been explored in prior research. In particular, DECODE (Empowering Chemical Structures with Biological Insights for Scalable Phenotypic Virtual Screening, Bioinformatics, 2026) also integrates molecular structures, transcriptomic profiles, and cellular morphology, and explicitly separates shared and modality-specific representations through orthogonality constraints and contrastive alignment. The proposed MCR-MSR decomposition appears closely related to these existing techniques. However, the paper does not sufficiently discuss or compare with such related approaches. others:

  1. The framework uses molecular structures, gene expression profiles, and cellular morphology for drug representation learning and prediction. However, it is unclear whether all three modalities are required for unseen drugs during inference. If transcriptomic and morphological measurements are necessary for every test compound, the method may have limited applicability to newly designed compounds without experimental profiling.

  2. The proposed mechanism-consistent representations are mainly learned through representation separation, cross-modal alignment, contrastive learning, and stability regularization. However, these objectives do not necessarily ensure that the learned shared representations correspond to actual pharmacological mechanisms, since cross-modal correlations may also arise from experimental conditions or other confounding factors.

  3. In Eq. (4), the proposed geometric reference penalizes high representation similarity between different drugs. However, drugs with different chemical structures may share targets or mechanisms of action and should potentially remain close in the learned representation space.

Questions for authors

Please see the critical weaknesses above

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

This paper proposes PMRD, a pharmacological response domain-guided framework for multimodal zero-shot drug property prediction. It introduces a multimodal framework that incorporates molecular structure, gene expression, and cellular morphology. The proposed method separates shared information in different modalities from modality-specific information, and learns a shared representation of drug responses. It also incorporates Stability-Guided Optimization (SGO) module for robust representations and a a Reliability-Aware Retrieval (RAR) process for zero-shot prediction without task-specific fine-tuning. The authors evaluated PMRD on ChEMBL2K and Broad6K with scaffold splits. The proposed method outperforms existing baselines and the ablation studies demonstrate the effectiveness of 5 different components.

Critical strengths

This method separates Mechanism-Consistent Representation (MCR) and Modality-Specific Representation (MSR). I find this idea and application especially interesting and compelling, as they do reflect how data from different modalities work in nature: they have the information that is specific/unique to each of them, meanwhile they share some hidden/foundational common knowledge as different forms of representation of the same entity. A model being able to learn good representation in these two different aspects would be very valuable.

Critical weaknesses

The paper focuses a lot on learning mechanism-consistent drug representations/MCR, but I am not too convinced that the experiments provide significant or sufficient evidence on the actual quality of the learned representations and their capability to capture pharmacological mechanisms. My two main concerns being:

(1) The downstream predictions seem to rely heavily on RAR, which is a downstream property prediction component. As shown in Table 4, removing RAR reduces the average AUROC from 77.8% to 67.8%. That is a 10% drop and it is a much larger drop than removing any of the other components. While this does show that RAR is important, RAR does not seem to be a part of the representation learning itself.

(2) I can see that Figure 4 shows the learned representations form clusters associated with biological or assay-related themes. My concern is that these themes might not necessarily correspond to specific "pharmacological mechanisms". The response that the assays measure can be caused by different mechanisms, that is, compounds associated with the same assay may act through very different targets or pathways. I do agree that this visualization provides some evidence of biological relevance, but again, biological relevance does not necessarily imply a shared pharmacological mechanism, so I think it shows more of the ability to group compounds based on related biological or assay responses more than the ability to identify the potential common mechanism-of-actions.

Questions for authors
  1. Can the authors do some analysis to address my concerns in the critical weakness section?

  2. How are DAM's dynamic weights computed? I see the retrieval-geometry objective and the dynamically weighted loss in the paper, but the mathematical procedure for calculating the weights based on directional support is not fully specified.

  3. It seem like the models are only evaluated by AUROC. For a more thorough evaluation, can the authors provide some other metrics? For example AUPRC

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

This paper proposes PMRD, a multimodal framework for drug property prediction using molecular structures, gene expression, and cellular morphology. Cross-Modal Response Modeling separates each modality’s representation into mechanism-consistent and modality-specific branches and aligns the shared representations through a consensus center. Stability-Guided Optimization combines locally perturbed candidate representations with loss reweighting based on a retrieval-geometry reference. At inference, Reliability-Aware Retrieval aggregates predictions from five representation views using labeled reference compounds.

The intended setting excludes target-property labels from representation learning while permitting label transfer from the reference set during prediction. Experiments on ChEMBL2K and Broad6K use scaffold splits and report results over five runs. The paper also presents component ablations, a low-structural-similarity hard-negative analysis, and embedding visualizations to support its claims about predictive performance and mechanism-consistent representations.

Critical strengths
  1. In Table 1, PMRD achieves an average AUROC of 77.8±0.3, compared with 71.1±0.8 for CLIP and 71.5±0.8 for CCL. These gains are substantial within the reported evaluation and provide a meaningful reason to investigate the proposed framework further. The use of scaffold splits and five independent runs strengthens the empirical evaluation.

  2. Learning transferable drug representations from chemical structure and cellular responses is a valuable problem. The framework explicitly addresses the coexistence of shared and modality-specific information, and its property-dependent aggregation of retrieval views is reasonably motivated. The ablations in Table 4 suggest that the complete system benefits from its constituent components, with reliability-aware retrieval having a particularly large effect.

Critical weaknesses
  1. The central algorithm and evaluation protocol are insufficiently specified to assess reproducibility and attribute the reported gains. Sections 3.2–3.4 describe several important quantities primarily through their intended effects. The manuscript does not provide the complete candidate acceptance criteria for CAM, the computation of the attribution weights in Eq. (5), or the mapping from confidence, agreement, neighborhood overlap, label coverage, and attribution priors to the reliability weights in Eq. (6). The fusion representation’s construction and training objective are also unclear. These choices are central to the proposed method and could materially affect its behavior.

2.Table 3 reports the same-target rate among top-K candidates with Morgan Tanimoto similarity below 0.1, treats a lower rate as better, and interprets this as evidence that PMRD is less likely to separate biologically related compounds. However, the candidate set and ranking function are not sufficiently defined. If these candidates are nearest neighbors in the learned representation, a lower same-target rate could indicate that same-target compounds have been moved farther apart. If they are the negatives actually selected for training, the corresponding selection procedure must be specified; Section 3.2 instead states that other drugs in the mini-batch serve as negatives.

Questions for authors
  1. Which target-property labels are accessible during representation learning, model selection, hyperparameter tuning, reliability estimation, and final retrieval? Are the same property tasks shared across the training, validation, and test sets? Please clarify whether “zero-shot” refers to held-out compounds with no task-specific fine-tuning or to genuinely unseen property tasks.
  2. What are the exact candidate pool and ranking function used to construct the top-K sets? Are these nearest neighbors in the learned space or negatives that actually receive repulsive training gradients? Please explain why a lower same-target rate in the defined set supports the conclusion that fewer biologically related compounds are incorrectly separated.
  3. Please provide the CAM candidate acceptance criteria, the calculation of the DAM weights, the RAR reliability formula, and the construction and training objective of the fusion view. These definitions are necessary to determine what algorithm produced the reported results. 4.What prediction rule replaces RAR in Table 4? Under a common retrieval procedure, reference-label set, and data split, does PMRD retain its advantage over the relevant baselines? Conversely, how much do compatible baseline representations improve when given the same RAR procedure? These comparisons would materially affect my assessment of the contribution of the learned representations.

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

Related papers

Discussion (0)

Sign in to comment.