acceptodds
Under review as a conference paper at ICLR 2027

CellMINT: Disentangled Multimodal Learning with Biologically Informed Task Interaction for Assay Prediction in Drug Discovery

Abstract

Cell Painting phenotypes are increasingly integrated with chemical perturbation information to support diverse assay prediction tasks in drug discovery. However, existing multimodal approaches often emphasize modality-shared information while underutilizing task-relevant modality-specific signals. Moreover, most multi-task methods for joint assay prediction do not explicitly model interactions among biologically related assay endpoints, limiting effective knowledge sharing across them. In this paper, we propose CellMINT, a multimodal multi-task framework that enhances Cell Painting-based prediction through disentangled multimodal learning and biologically informed task interaction. To identify task-relevant multimodal evidence, CellMINT disentangles multimodal features into modality-shared and modality-specific components, enabling each assay endpoint to adaptively aggregate the information most relevant to its prediction. To further exploit complementary information across related tasks, CellMINT constructs a biological prior from assay context to guide selective task interaction among informative endpoints. Experiments on three public Cell Painting benchmarks demonstrate that CellMINT improves average predictive performance and high-performing task coverage over unimodal and multimodal baselines, highlighting its potential for comprehensive assay profiling in drug discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
multimodal learning, multi-task learning, Cell Painting, biological assay prediction

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

This paper proposes CellMINT, a multimodal multi-task framework for predicting biological assays from Cell Painting images in drug discovery. CellMINT decomposes image and chemical features under the guidance of controls. Perturbed CP image embeddings are calibrated against plate-matched negative controls. Each feature is then split into three parts: phenotype-specific, shared across modalities, and chemistry-specific. The framework also uses a task-adaptive interaction mechanism for cross-task transfer. The authors state that it is the first such mechanism for heterogeneous assay collections across biological domains. CellMINT is then validated on three public CP benchmarks.

Critical strengths

The paper studies assay prediction on heterogeneous, sparsely labeled data, where different endpoints depend on evidence from different modalities. This problem has received little attention, and the proposed two-stage solution is a reasonable response to it. The design is simpler than the uniform fusion and shared-space alignment used in earlier work, and its motivation is clear. The empirical support is broad. On three public CP benchmarks the method beats both unimodal and multimodal baselines, and the gains appear not only in mean AUROC but also in the fraction of tasks that reach a high threshold.

Critical weaknesses

The biological prior is produced by an LLM reasoning over self-defined structured task profiles. But the authors didn't report the validation of that prior's accuracy or reliability. There is no agreement with known relations, no measure of prompt/rubric sensitivity, no quantification of the noise injected by misinferred or missed helper relations. So a key claimed novelty relay on an unverified and non-reproducible component. The evaluation is also narrow and potentially overstated. The results come only from three CP benchmarks with three indication and no external validation etc.

The biological prior comes from an LLM that reasons over hand-defined structured task profiles. The authors do not report how accurate or reliable this prior is. There is no comparison with known relations among tasks. There is no analysis of how sensitive the output is to prompt and rubric design. There is also no estimate of the noise introduced when a helper relation is inferred wrongly or missed. The claimed novelty therefore depends on a component whose quality has not been checked. The evaluation is also narrowed. All results come from three public CP benchmarks covering three indications, with no further or external validation.

Questions for authors

How accurate is the LLM-generated task-relation prior? Please provide a quantitative validation of it. Are the reported gains significant? Report results over multiple seeds. Do the same for the main table and for the ablations. The baselines were adapted to the multi-task setting by the authors. Were they tuned with a comparable budget? Finally, why are modern multi-task and transfer approaches absent from the comparison?

1 comment

Thanks for these additional validations. These new experiments strengthen the empirical support. But the core limitations I think, the unverified LLM prior and narrow evaluation scope, still remain fundamental constraints on this work. I therefore maintain my rating of 3.

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

CellMINT predicts sparsely observed biological assay activities from Cell Painting images and compound descriptions. It calibrates image features with plate-matched controls, separates shared and modality-specific representations, uses assay queries to select features, and restricts task attention with a directed prior inferred from assay metadata and refined using training-label correlations. Its main methodological contribution is extending biologically guided task interaction to heterogeneous assays. Evaluation covers three public benchmarks, with predictive metrics, task coverage, component ablations, and task-level analyses.

Critical strengths
  • The empirical gains are consistent across heterogeneous benchmarks. mean AUROC improves over the best reported baselines, including the structure-cluster evaluation on cpg0016.
  • The task-interaction controls are informative, especially on cpg0012, where no interaction, language-model-only, label-correlation-only, and combined priors achieve superior AUROC than random and unrestricted interaction. This supports the usefulness of metadata-informed task selection beyond simply adding attention.
Critical weaknesses
  • The central disentanglement regularizer is not sufficiently specified. Equations (12)--(13) use learned conditional estimators, but Equation (6) contains no estimator-fitting objective, and lines 840–841 only state that these estimators are learned from matched pairs during joint training. CLUB's Algorithm 1 distinguishes conditional-likelihood maximization for the estimator from dependence minimization for the representation. The missing update and gradient-flow details matter because minimizing Equation (13) over both parameter groups would be a different procedure. This makes fully understanding the pipeline difficult.
  • Relatedly, the evidence does not isolate the role of this regularizer in producing the claimed shared/specific separation. The disentanglement step in Table 2 improves AUROC while introducing the projected representations and their objectives together. Figure 3 measures predictive sensitivity to feature removal, which establishes feature utility but does not itself establish reduced shared--specific dependence.
  • Table 7 varies nonzero CLUB weights but does not include a zero-penalty control. A comparison with the same projections and querying architecture but no CLUB penalty would materially strengthen the mechanism claim.
  • The task-interaction mechanism closely follows BiGMINT, Section 3.3.1, which already combines biological-prior selection, correlation pruning, and helper-task cross-attention. CellMINT's main extension is a metadata-derived prior supporting heterogeneous assays. This is useful, but limits the novelty of the interaction mechanism itself. Comparing against a simpler prior constructed from the same metadata would clarify what language-model reasoning adds beyond the inherited approach.
Questions for authors
  • Clarification: What objective and update schedule fit each conditional estimator, and which parameters are frozen or detached during estimator fitting versus representation optimization?
  • Evidence: With the projection and querying architecture held fixed, what happens to predictive performance and shared–specific dependence when the CLUB weight is zero? This would help determine how much of the Table 2 improvement supports the proposed disentanglement mechanism.
  • Novelty clarification: Beyond the source of the task prior and the attention-output combination, what methodological advance does the interaction module make over BiGMINT? With the downstream architecture and helper-task budget held fixed, how does the language-model prior compare with a simpler prior based on shared assay metadata, such as target, cell type, or readout?
Reviewer_L6HKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CellMINT, a multimodal model for predicting many biological assay outcomes from Cell Painting images and chemical information. It separates information into phenotype-specific, shared, and chemistry-specific parts, lets each task use the parts that matter most, and builds an LLM-based task graph so related tasks can share information. Across three public datasets, CellMINT reports AUROC scores of 0.783, 0.769, and 0.681, beating the listed baselines.

Critical strengths

The paper gets strong results on three different datasets. The ablations are also useful: random or unrestricted task interaction performs worse than the proposed task graph, which suggests that choosing the right task links matters.

Critical weaknesses

My main concern is attribution of the LLM-based prior. BiGMINT (Pati et al., 2026), the closest prior work and the source of the correlation-based sparsification step, is not included as a baseline. The GPT-5.5 prior only gets 0.777 AUROC versus 0.771 for label correlation alone on cpg0012, and Table 2 gives no variance or significance test for that difference. The final graph also filters LLM edges using label correlation, so it is hard to tell how much of the gain comes from the LLM. Because these assays are public, the paper should also explain how it rules out GPT-5.5 using assay knowledge seen during pretraining.

A second concern is the chemical input and evaluation setup. CellMINT gives general BERT both the compound name and SMILES, so a SMILES-only or ECFP-style ablation would help show that names are not adding extra semantic information. The paper should also clarify what observations the paired t-test is run over.

Minor: the main text says g > 0, while the appendix constrains 0 < g < 1.

Questions for authors
  1. Why are BiGMINT and MoMIL not included as baselines?
  2. Can you report variance or a significance test for the GPT-5.5 prior (0.777) versus the label-correlation prior (0.771)?
  3. Can you compare the LLM prior with simple metadata-based priors at the same edge count?
  4. How do you rule out GPT-5.5 using knowledge of these public assays from pretraining rather than only the supplied metadata?
  5. What happens with SMILES only, or an ECFP/ChemBERTa-style chemical encoder, instead of compound name + SMILES through BERT?
  6. What exactly is paired in the paired t-test?
  7. Please clarify the g constraint: g > 0 in the main text versus 0 < g < 1 in the appendix.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.