Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Accurately predicting drug-target interactions (DTI) is central to accelerating computational drug discovery; however, existing methods still face challenges in effectively constructing multimodal features, fusing heterogeneous features, and capturing complex non-linear relationships. This paper proposes a novel dual-modal neural network framework, MGK-DTI, designed to deeply integrate the sequence and structural information of drugs and proteins to address these challenges. First, we designed Mambaformer in the sequence branch, utilizing Mamba to capture long-range dependencies while leveraging the advantages of the multi-head attention mechanism in global context; subsequently, we designed a hybrid Gated Graph Convolutional Network in the structural branch, which uses an adaptive gating mechanism to dynamically balance intra-modal topological updates with cross-modal feature injection, achieving bidirectional fine-grained interaction between protein structural features and drug molecular graph structures; finally, after unifying the dual-path feature space through the Latent Alignment Module, the Kolmogorov-Arnold Network is introduced to strengthen non-linear interaction modeling. Experimental results on four benchmark datasets demonstrate that MGK-DTI outperforms state-of-the-art methods. Ablation studies validate the contribution of each core module, and the case study on Histamine H1 receptor further confirms the immense potential of the model in virtual screening. Code is available at https://github.com/HellojiePro/MGK-DTI.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4913_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/HellojiePro/MGK-DTI

Link to the Dataset(s)

N/A

BibTex

@InProceedings{CheDon_DualModality_MICCAI2026,
        author = { Chen, Dongjie AND Liu, Yuemeng AND Liu, Binfeng AND Wang, Zican AND Lin, Hangxing AND Yang, Zhihui},
        title = { { Dual-Modality Neural Network Integrating Sequence and Structural Information for Drug-Target Interaction Prediction } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16896},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces a novel drug-target interaction prediction framework called MGK-DTI. The framework uses a dual-modality neural network, integrating sequence and structural information of drugs and proteins, enabling more accurate and efficient computational drug discovery. Highlighted innovations includes: (1) the method leverages deep learning architectures (Mambaformer) combining Mamba for capturing long-range dependencies and multi-head attention for global context, improving sequence feature extraction; (2) it employs Hybrid Gated Graph Convolutional Network (HGGCN) with an adaptive gating mechanism, dynamically balancing intra-modal topological updates and cross-modal feature injection to allow for fine-grained interaction between protein structural features and drug molecular graphs, and (3) Latent Alignment Module (LAM) unifies the dual-path feature space, followed by the Kolmogorov-Arnold Network (KAN) to enhance non-linear interaction modeling. Experiments are conducted on benchmark DTI datasets where MGK-DTI outperforms SOTA methods, supported by ablation study to confirm the effectiveness of each module. Validation also includes a case study on the Histamine H1 receptor demonstrating the model’s potential in virtual screening.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The dual-modality formulation integrating sequence and structural information from drugs and proteins is novel. This formulation enables the model to capture complex, non-linear relationships and synergistic features that single-modality methods often miss
    • The paper demonstrates strong empirical results, outperforming state-of-the-art methods on four benchmark DTI datasets. The case study on Histamine H1 receptor also highlights both robustness and real-world applicability
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The evaluation is limited to benchmark dataset and one case study and does not include any wet-lab validation. There is still a gap to practical impact and translational relevance limiting the clinical feasibility for real-world deployment.
    • The authors are encouraged to include a discussion of the limitations of the approach such as potential overfitting or whether there are any challenges with scaling to larger or more diverse dataset.
    • Can the authors also comment on how well the model generalizes to unseen drug-target pairs for readers to have some idea regarding the robustness and broader applicablibility of the model.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper is well-written introducing a novel dual-modality neural network framework. The work is supported by comprehensive experimental validation across 4 benchmark datasets and ablation studies. A real-world case study was also introduce to demonstrate the robustness and practical impact of the approach. The work is certainly relevant for the computational drug discovery community.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper proposes MGK-DTI, a dual-modality neural framework for drug-target interaction prediction. It integrates a sequence branch (Mambaformer combining Mamba state-space layers with multi-head attention) and a structural branch (Hybrid Gated GCN, HGGCN) to capture sequence dependencies and cross-modal topological features. A Latent Alignment Module unifies dual-modality features, and a Kolmogorov-Arnold Network (KAN) is used for final non-linear interaction modeling. Experiments on four benchmark datasets achieve state-of-the-art AUC-ROC and AUC-PR, with ablation studies verifying module effectiveness and a case study on the H1 histamine receptor demonstrating application potential.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.Technical novelty: The architecture innovatively combines Mamba (long-range sequence modeling) and multi-head attention, and designs a gated fusion mechanism for intra/inter-modal signals in the structural branch; applying KAN for late fusion is a timely and valuable attempt. 2.Experimental integrity: Ablation studies confirm the contribution of each core module, and a qualitative case study on a clinically relevant GPCR target with latent space visualization and molecular docking is provided. 3.Presentation clarity: The overall system diagram and modular decomposition help readers quickly grasp the pipeline, and mathematical descriptions of key modules are concise and clear. 4.Research significance: DTI prediction is a high-impact task for computational drug discovery, and the multimodal fusion design aligns with the frontier research trends of the community.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.Protein “structural” modeling only uses sequential adjacency via DSConv instead of real 3D structures, overstating biological fidelity. 2.Cross-modal attention has high computational complexity with no scalability discussion. 3.Core components (LAM, gating) are standard, and novelty mainly comes from architectural combination. 4.Missing critical details of dataset splitting (random/scaffold/cold-start), negative sample construction, data deduplication and tokenization 5.No multi-seed experiments, variance analysis or significance testing 6.Missing comparison with recent state-of-the-art baselines. (MMDG-DTI, UnitedDTA, etc.), all comparison methods were developed before 2025.7.Insufficient motivation for LAM/KAN design.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
    • Fig. 1 is too colorful.
    • There is occlusion (word “Gate Generation”) at the bottom-right of Fig.1.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper targets a high-impact DTI prediction task and proposes a novel architecture integrating Mambaformer, HGGCN, LAM and KAN, with promising preliminary experimental results. However, it has some flaws: missing key experimental details ; incomplete baseline comparisons; insufficient discussion on biological fidelity. These problems make the paper unable to fully prove its core claims, so it is recommended as weak reject. I’d like to change my rating if the authors would have fixed these issues.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors have submitted a responsive rebuttal that addresses my concerns. I recommend acceptance.



Review #3

  • Please describe the contribution of the paper

    The authors propose MGK-DTI, a dual-model neural network framework for predicting drug–target interactions using raw drug molecule and target protein information. The main contributions include the introduction of a Mambaformer-based sequence branch to capture multi-scale dependencies, the development of an HGGCN-based structural branch for topological feature extraction, and the incorporation of KAN to enhance nonlinear interaction modeling. Comprehensive experiments demonstrate that MGK-DTI outperforms state-of-the-art methods in DTI prediction.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper has several notable strengths. First, it presents a task-oriented, end-to-end framework for DTI prediction that explicitly integrates complementary sequence and graph-based information. On the sequence side, the proposed Mambaformer is designed to capture longer-range dependencies in protein/drug sequences more efficiently than purely attention-based designs, which is a meaningful design choice for DTI where relevant motifs may be far apart. Second, the structural branch introduces a Hybrid Gated GCN (HGGCN) that performs both intra-modality message passing and cross-modality feature injection between drug atoms and protein residues, offering an explicit mechanism to model drug–target interplay rather than treating the two encoders as independent. Third, the Latent Alignment Module (LAM) provides a principled way to project the two modality-specific representations into a shared latent space, which helps stabilize fusion and makes the downstream predictor operate on unified features. From an evaluation perspective, the paper includes ablation studies evaluated with both AUC-ROC and AUC-PR, which is appropriate for imbalanced interaction prediction and helps isolate the contribution of key components (sequence encoder, HGGCN fusion, LAM, and predictor). Finally, the case study on the target protein P35367, reporting a ranked list of top candidate drugs, provides an application-oriented demonstration that the method can be used in a practical screening scenario and not only as a benchmark classifier.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The biggest concern of this paper is that it does not specify a sufficiently rigorous and reproducible evaluation protocol (e.g., exact train/validation/test split strategy, whether splits are random vs. scaffold/cold-drug/cold-target, number of seeds, and overlap control). For DTI benchmarks, random splits can lead to overly optimistic performance due to shared scaffolds across drugs or high sequence similarity across proteins between training and test sets. Without explicit split definitions and repeated runs (mean±std), the reported improvements—often small in absolute terms—are difficult to interpret and may not reflect real generalization. Also, the method combines multiple sophisticated components (Mambaformer, HGGCN, and KAN), but the experiments do not convincingly attribute the performance gains to each module under a fair, capacity-controlled setting. In particular, it is unclear whether the improvements come from the proposed architectural ideas or simply from increased model scale/parameter count and compute. The paper should report parameter counts/FLOPs and include matched-capacity baselines (e.g., KAN vs. MLP with similar parameters; Mambaformer vs. a Transformer of comparable size) to support the claim that each module contributes meaningfully at a comparable scale.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents an innovative approach to predicting drug–target interactions using raw drug molecule and target protein information, and the rationale for each module is clearly described. However, the data splitting protocol and evaluation strategy need to be better specified and, potentially, revised to properly validate the author’s hypotheses.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We sincerely thank all reviewers for their careful review and constructive comments! Rigor of experimental evaluation (R1,R2,R3): Our evaluation method and dataset partitioning follow the research work [10] cited in the corresponding part of the paper. BindingDB is split into train/val/test sets, where the validation set is only for hyperparameter tuning. Other datasets adopt 5-fold cross-validation, and all reported results are averaged with standard deviation below 0.004.We agree that multi-seed experiments and significance tests are crucial for the robustness evaluation of the model and will be added in future work. Key details of the dataset will be updated in the revised version. Open source (R2): We promise to open-source code and dataset for reproducibility. Limitations and generalization (R1,R3): Our model achieves SOTA results on four benchmark datasets of small, medium and large scales, verifying its effectiveness. The two branches are complementary and do not rely on domain-specific prior substructure features. They capture the underlying topological structure and sequence features, possessing the ability to generalize to unknown substructures. In future work, we should also add drug/target/drug-target cold-start and single-source/multi-source cross-domain experiments (larger or more diverse) to further validate the model’s capabilities. Baseline comparison and fairness of ablation experiments (R2,R3): Thanks for pointing out this deficiency. At the reviewers’ request, we added comparisons for MMDG-DTI. We also benchmarked KAN vs MLP with identical parameters, and Mambaformer vs Transformer under the same scale. All results meet our expectations. Due to the length and rules of this rebuttal, we omit detailed analysis here. Relevant comparison details will be presented on GitHub or in future work. Module innovation and insufficient motivation (R2): While our main innovation lies in the overall architecture, the core components are equally crucial. The initial design purpose of LAM is to align the heterogeneous representation spaces of two modalities, resolve modal differences, and provide a unified representation foundation for subsequent cross-modal deep fusion. The gating mechanism adaptively balances raw features, intra-modal topology and cross-modal interactions at each layer, which is also an indispensable design. Ablation results validate the necessity of both modules. Most existing DTI models adopt MLP as fusion and classification backbone. To overcome MLP’s inherent limitations, we notice the application potential of the emerging KAN. With learnable spline activation, KAN better approximates complex continuous functions, achieves higher parameter efficiency and faster convergence than MLP. We therefore replace MLP with KAN as the prediction layer to better capture high-order nonlinear DTI patterns. Ablation also verifies that model performance drops notably when KAN is replaced by MLP. Biological fidelity (R2): The structural branching design aims to explore the structural relationships between drug atoms and protein amino acids, not the actual 3D structure. Therefore, adjacent amino acids are defined as adjacent nodes according to sequence order. As mentioned in the paper, the reason for the limitations is that protein sequences are difficult to directly construct realistic amino acid diagrams, and predicting 3D structures is computationally expensive and prone to introducing prediction errors, thus restricting the model. Complexity (R2): Although the cross-modal attention has some computational complexity, the biomolecules are all short sequence inputs, so the actual overhead is controllable. It aims to build correlation matrices between proteins and drugs. Wet-lab Validation (R1): We acknowledge that this work lacks wet-lab validation due to research constraints. Occlusions (R2): “Gate Generation” is not an occlusion; it merely describes gate G’s generation in HGGCN, causing reviewer misunderstanding.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    Several critical concerns have been raised by all the reviewers about the limited evaluation and protocol, generlization analysis, missing critical details and comparison. The authors are invited to submit a rebuttal to carefully respond and clarify all the concerns raised by the reviewers.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This paper was invited for rebuttal to reviewers’ concerns. The reviewer has changed a negative rating into a positive. After carefully evaluating the authors’ rebuttal, the reviewers’ revised feedback, and re-examining the paper, I conclude that the authors have effectively addressed the key concerns raised, and the work now meets MICCAI’s publication standards.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    After rebuttal, all reviewers suggest accepting this paper with one reviewer not submitting the final score.

    It is suggested that the authors sufficiently consider the comments raised by the reviewers and further improve the paper in the final version.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The reviewers appreciate the novelty and experimental validation presented in the paper, highlighting its relevance to the computational drug discovery community. Authors are suggested to address concerns such as incomplete experimental details, baseline comparisons, evaluation strategies, and biological fidelity in the next version.



back to top