acceptodds
Under review as a conference paper at ICLR 2027

Multimodal Learning with Missing Modalities via Multi-Level Alignment for Drug–Target Affinity Prediction

Abstract

Drug–target affinity (DTA) prediction estimates the binding strength between a small-molecule drug and a target protein. Existing methods typically rely on either sequence or structure information, limiting the joint exploitation of complementary supervision from heterogeneous datasets. Label heterogeneity is often overlooked, leading to the conflation of distinct experimental measurements into a single prediction target. To address these issues, we propose M3A-DTA, a missing-modality, multimodal, and multi-assay framework that learns from any available subset of drug sequence, protein sequence, ligand structure, and protein structure. Modality-specific pretrained encoders provide chemically and biologically informed representations. Through multimodal alignment, adaptive fusion, and task-specific supervision, M3A-DTA learns robust representations that accommodate incomplete modalities. The design enables joint training on sequence-centric data together with structure-centric complexes. We specify evaluation under complete, naturally missing, and deliberately masked modalities, with cold-drug, cold-target, and cold-pair splits. M3A-DTA achieves strong performance across multiple evaluation settings while demonstrating broader applicability than most existing methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Drug–Target Affinity Prediction, Multimodal Learning, Missing Modalities, Multi-Level Alignment, Deep Learning, Drug Discovery, Bioinformatics

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_6LATOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes to improve DTI prediction by leveraging both structure-based features and sequence-based features. However, the paper is incomplete in its current state: missing experimental results and ablation studies.

Critical strengths

NA

Critical weaknesses
  • Missing experimental results
  • Missing ablation studies
  • Table 3 is incomplete
Questions for authors

See above

Reviewer_kmhpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes M3A-DTA, a unified framework for drug–target affinity prediction using incomplete and heterogeneous molecular information. It integrates four different (optional) modalities: drug sequence, protein sequence, ligand structure, and protein structure. The model adapts modality-specific pretrained encoders to provide chemically and biologically informed representations. Through multimodal alignment, adaptive fusion, and task-specific supervision, M3A-DTA learns representations that accommodate incomplete modalities. The model is evaluated on under-complete, naturally missing, and deliberately masked modalities. The proposed model achieves improved performance across multiple evaluation settings.

Critical strengths

This framework addresses a practical and important limitation of multimodal DTA prediction, that is, in real life application, while drug and protein sequence data tend to have a higher availability, high-quality protein–ligand complex structure data are often much more limited. As a result, models that require all of these modalities for prediction might face significant limitation if they are to be applied to real-world drug discovery scenarios.

I think M3A-DTA addresses this issue in a meaningful way by allowing certain modalities to be optional. Overall, I think the motivation is well-grounded in reality, and the methodology provides a reasonable way to address it.

Critical weaknesses

The experimental analysis provided in the paper is not complete enough to fully evaluate or prove the effectiveness of the proposed model.

  1. The evaluation of missing modalities is insufficient. Although Section 4.5 briefly discusses missing-modality evaluation, no performance results specific to each modality availability settings have been reported.

  2. In Section 4.2, the authors describe the cold-drug, cold-target, and cold-pair data-splitting strategies, but the result table (Table 1) only reports results on a random split of PDBbind v2020 full set.

  3. One of the contributions claimed is "We align sequence and structure through known atom- and residue-level correspondence", but there no ablation results have been provided to specifically evaluate the contribution of this design choice.

  4. The paper introduces separate prediction heads for different affinity measurements and a loss for handling censored observations. However, there is no comparison between separate heads and a single joint/shared head to validate this design choice.

  5. I do not understand what Table 3 is intended to show. It lists six different methods but only reports a numerical result from one of them. Without any number from the other models and the complete M3A-DTA model (which is the proposed model), it is difficult to make any meaningful comparison or understand whether and how the proposed strategy actually improves censor data handling.

Questions for authors

Can the authors provide:

  1. results under different modality availability settings?
  2. results from different splits as described in Section 4.2?
  3. ablation results which validate the design choice of each module?
  4. the complete results in Table 3 and perhaps also some analysis related to it?
Reviewer_nvEiOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper addresses the insufficient utilization of complementary multimodal information in drug–target affinity prediction by proposing a unified framework that integrates sequence and structural representations while supporting missing modalities. Its main designs include cross-modal alignment that preserves element-level correspondences and measurement-type-aware prediction heads. However, the manuscript places excessive emphasis on label heterogeneity without providing experimental evidence to demonstrate its actual impact. Although the potential benefits of joint training on datasets such as Davis and KIBA are discussed earlier in the paper, they are not reflected in the subsequent experimental results. Overall, the contributions are primarily limited to mechanistic extensions of existing frameworks. The lack of targeted ablation studies for the proposed designs makes it difficult to establish their effectiveness and substantive contributions.

Critical strengths
  1. Researching the problem of drug-target affinity in scenarios where certain modalities are missing helps facilitate the practical implementation of methods in the field of drug discovery.

  2. The author has made some initial explorations into the problem, though the methodology remains a combinatorial innovation based on existing approaches to multimodal tasks.

Critical weaknesses
  1. Severely insufficient experimental validation. The research questions raised in Section 4.1, including cold-start generalization and robustness to missing modalities, lack corresponding experimental results. The joint utilization of datasets such as Davis and KIBA is also not demonstrated in the experiments, and Table 3 does not even report the results of the full model. These shortcomings prevent me from verifying whether the paper's core claims are supported and constitute my primary concern in the overall evaluation.

  2. Insufficiently demonstrated novelty and contributions. The main designs focus on element-level cross-modal alignment and measurement-type-aware prediction heads, but the paper lacks in-depth comparisons with existing methods and targeted ablation studies. Consequently, I cannot determine whether the reported performance improvements genuinely result from these newly introduced mechanisms, nor can I recognize substantial contributions beyond the combination and incremental extension of existing techniques.

Questions for authors
  1. Section 4.1 presents six research questions, but the current experimental results mainly focus on the random split of PDBbind. Have the cold-start generalization and missing-modality experiments been completed? If so, please provide the corresponding results to support the core claims regarding the model's generalization ability and robustness to missing modalities.

  2. How much performance improvement is independently attributable to element-level cross-modal alignment and measurement-type-aware prediction heads?

  3. Were Davis, KIBA, BindingDB, and PDBbind, as illustrated in Figure 1, actually used for joint training of the model evaluated in Table 1? Please clarify the role of each dataset, the cross-dataset deduplication strategy, and whether the proposed method and the baselines were trained using the same training data.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.