Abstract

During the process of brain aging, the changes of white matter structural connectivity are closely correlated with the cognitive traits and brain function. Genes have strong controls over this transition of structural connectivity-altering, which influences brain health and may lead to severe dementia disease, e.g., Alzheimer’s disease. In this work, we introduce a novel deep-learning diagram, an oblique genomics mixture of experts(OG-MoE), designed to address the prediction of brain disease diagnosis, with awareness of the structural connectivity changes over time, and coupled with the genomics influences. By integrating genomics features into the dynamic gating router system of MoE layers, the model specializes in representing the structural connectivity components in separate parameter spaces. We pretrained the model on the self-regression task of brain connectivity predictions and then implemented multi-task supervised learning on brain disorder predictions and brain aging prediction. Compared to traditional associations analysis, this work provided a new way of discovering the soft but intricate inter-play between brain connectome phenotypes and genomic traits. It revealed the significant divergence of this correlation between the normal brain aging process and neurodegeneration.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/5275_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/autumnbreed/OG-MoE

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LyuYan_Oblique_MICCAI2025,
        author = { Lyu, Yanjun and Zhang, Jing and Zhang, Lu and Ruan, Wei and Liu, Tianming and Zhu, Dajiang},
        title = { { Oblique Genomics Mixture of Experts: Prediction of Brain Disorder With Aging-Related Changes of Brain’s Structural Connectivity Under Genomic Influences } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15963},
        month = {September},

}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces a novel deep learning framework called OG-MoE (Oblique Genomics Mixture of Experts), which integrates genomic features into a transformer-based mixture-of-experts (MoE) architecture for predicting brain disorders (CN/MCI/AD) based on structural connectivity (SC) changes over time. The unique contribution is the oblique (non-direct) fusion of genomics into the expert-routing mechanism, allowing the model to dynamically modulate the computation path for SC features without directly merging multimodal inputs. The model is pretrained using masked self-regression on SC sequences and fine-tuned with multi-task learning for disease classification and aging-related SC change 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.
    1. The use of genomics as a bias term in the MoE routing mechanism offers an elegant and indirect method of integrating genetic data into neuroimaging models.

    2. By constructing paired SC sequences across visits, the model learns age-related connectivity changes, distinguishing between normal and pathological aging.

    3. The model’s contrastive loss separates genomic-driven and environment-driven representations, aligning well with known neurobiological mechanisms.

  • 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. The ADNI dataset includes participants from diverse ethnic backgrounds, and prior research has shown that genetic traits can vary significantly across populations. Therefore, population stratification is essential, or appropriate correction methods must be applied to mitigate potential confounding effects. Additionally, ADNI data have been collected over approximately two decades across various research phases, requiring harmonization of the reference panel. The paper does not clearly address how these issues were handled, which raises concerns about the reliability of the identified SNVs.

    2. The study employs fiber count, which refers to the number of reconstructed tracts, but fiber density is generally considered a more accurate metric. While fiber count measures quantity, fiber density captures the concentration or packing of fibers within a region, providing a more nuanced understanding of microstructural connectivity. Therefore, it would be more appropriate to use fiber density in the analysis.

    3. In structural connectivity (SC) studies, changes in brain connectome topology or message-passing efficiency are often more informative than raw SC values. Since fiber counts are highly sensitive to MRI quality and tractography algorithms, which can introduce substantial noise, It would be better capture aging-related changes in brain connectivity by focusing on topological feature (i.e., BC, DC, CC, Gradients), rather than connectivity patterns.

  • Please rate the clarity and organization of this paper

    Poor

  • 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.

  • 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/2025/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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
    1. The ADNI dataset includes participants from diverse ethnic backgrounds, and prior research has shown that genetic traits can vary significantly across populations. Therefore, population stratification is essential, or appropriate correction methods must be applied to mitigate potential confounding effects. Additionally, ADNI data have been collected over approximately two decades across various research phases, requiring harmonization of the reference panel. The paper does not clearly address how these issues were handled, which raises concerns about the reliability of the identified SNVs.

    2. The study employs fiber count, which refers to the number of reconstructed tracts, but fiber density is generally considered a more accurate metric. While fiber count measures quantity, fiber density captures the concentration or packing of fibers within a region, providing a more nuanced understanding of microstructural connectivity. Therefore, it would be more appropriate to use fiber density in the analysis.

    3. In structural connectivity (SC) studies, changes in brain connectome topology or message-passing efficiency are often more informative than raw SC values. Since fiber counts are highly sensitive to MRI quality and tractography algorithms, which can introduce substantial noise, It would be better capture aging-related changes in brain connectivity by focusing on topological feature (i.e., BC, DC, CC, Gradients), rather than connectivity patterns.

  • Reviewer confidence

    Very confident (4)

  • [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.

    It appears that the authors performed minimal processing for the harmonization of genetic data.

    Additionally, the referenced papers [2–4] rely on outdated methods. The authors should consider adopting more advanced approaches, such as MRtrix3 or other multi-shell tractography algorithms, to improve the robustness and contemporary relevance of their analyses.



Review #2

  • Please describe the contribution of the paper

    This paper proposed a novel deep-learning diagram, an oblique genomics mixture of experts(OG-MoE), designed to address the prediction of brain disease diagnosis, with awareness of the structural connectivity changes over time, and coupled with the genomics influences. By a novel deep-learning diagram, an oblique genomics mixture of experts(OG-MoE), designed to address the prediction of brain disease diagnosis, with awareness of the structural connectivity changes over time, and coupled with the genomics influences. By integrating genomics features into the dynamic gating router system of MoE layers, the model specializes in representing the genomics structural connectivity components in separate parameter spaces. To validate the proposed model, we pretrained the model on the self-regression task of brain connectivity predictions and then implemented multi-task supervised learning on brain disorder predictions and brain aging prediction. To validate the proposed model, the authors pretrained the model on the self-regression task of brain connectivity predictions and then implemented multi-task supervised learning on brain disorder predictions and brain aging 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.
    1. This paper integrated genomics features into the dynamic gating router system of MoE layers, the model specializes in representing the genomics structural connectivity components in separate parameter spaces.
    2. The self-regression task of brain connectivity predictions and multi-task supervised learning on brain disorder predictions and brain aging prediction are used to validate the proposed model.
  • 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. Pay attention to the completeness of certain sentences. In the abstract: to validate the proposed model. 2.The number of samples in the test dataset is too small, making the experimental results less convincing. It is recommended to extract a subset from the validation set to re-evaluate the model performance.
    2. To provide a more comprehensive evaluation of the classification performance in Table 2, it is recommended to report additional metrics including sensitivity, specificity, and the area under the ROC curve (AUC).
    3. The x-axis in Figure 3 lacks a clear definition, which hinders the interpretation of the presented results.
  • 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.

  • 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/2025/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 — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
    1. This paper takes the imaging modality (SCs) as the main feature and uses genomic data as an auxiliary feature to control the computation course of image modality.
    2. This paper presents a method that integrates genetic and structural connectivity data to investigate how significant genes influence changes in structural connectivity between AD/MCI and MCI/CN groups.
  • 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.

    Well done



Review #3

  • Please describe the contribution of the paper

    The paper presents an MoE model for AD prediction that leverages genetic information as a gating mechanism to enable dynamic perception of imaging tokens. The proposed method was evaluated on the ADNI cohort.

  • 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.

    S1: The motivation is well-justified, leveraging genomic knowledge to guide finer-grained utilization of imaging data. S2: The proposed method is overall reasonable. S3: The paper is well written.

  • 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. Including both disease classification and positive/negative SC tasks may be problematic. The underlying genetic mechanisms and structural brain changes associated with normal aging versus pathological changes could be fundamentally distinct. Ablation studies are needed to validate whether adding the positive/negative classification task actually improves model performance.
    2. While constructing two atlases might help capture both fine- and coarse-grained brain connectivity, I think fusing information from both atlases at the beginning is more likely to introduce noise rather than provide complementary/additional information. However, convincing ablation results could address this concern.
    3. Minor issues: For SNVs (single nucleotide variants), there are typically three possible states {0,1,2}, so why is the feature vector dimension 2×216 rather than 3×216?
    4. There are some typographical errors, such as: lacking spacing between text and citations; Duplicate notation of V_g in Section 2.4.
  • 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.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    Extensive experiments and ablation analyses can help demonstrate the effectiveness of the proposed modules.

  • 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?

    Utilizing genetic information as an expert to provide underlying genetic mechanism insights for imaging-based disease analysis is ingenious.

  • Reviewer confidence

    Very confident (4)

  • [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.

    Utilizing genetic information as an expert to provide underlying genetic mechanism insights for imaging-based disease analysis is ingenious.




Author Feedback

We appreciate the valuable comments and the recognition of the strengths in our paper: (1)The unique integrating methods of genomics features into the dynamic gating router system of MoE.(R1, R2, R3) (2)The construction of paired SC sequences and self-regression learning.(R1, R3) Code and data will be released upon acceptance.

We summarize and address the concerns: Reviewer 1: Q1, Concern about heterogeneity of ADNI data and reliability of the SNVs features. We agree that the data from ADNI was collected over decades, from diverse ethnic backgrounds, and processed by different sites. However, the ADNI project employs a standard workflow to ensure data quality and comparability across subjects [1]. Due to its design, our model does not estimate the association between genetic variants and disease, which allows the model to be robust to ancestry-related confounding, in contrast to standard association testing methods. Since we have carefully extracted our genomic profile from a pre-defined candidate gene pool and processed data with statistical correction and quality control (Sec 2.1, Table 3). We processed the SNVs data from the ADNI dataset across multiple phases, by: a) using the IDs to relocate the SNP to the reference genome GRCh38.p14 through NCBI’s database. b) retrieving the original SNP location of the microarray chip and mapping the original reference genome to GRCh38.p14 to relocate the variant loci. We manually double-checked the results from these two approaches to ensure the consistency of SNVs data. All processing codes and PLINK logs are available upon request for further review. Q2&Q3, Concern about the SC features It appears there’s a misunderstanding of our work. OG-MoE is an end-to-end self-regression model of embedding and reconstruction of brain structural connectivity (SC) topology. The brain structural connectivity (SC) is a suitable feature for the model design, free from rigid statistical graph metrics. We applied not only fiber counting upon tractography reconstructions, but a comprehensive workflow including normalizations and corrections to the computation of the SC (Sec 2.1 Imaging data process). Our construction and utilization of SC exactly follow the mainstream of the research [2-4].

[1]“Alzheimer’s Disease Neuroimaging Initiative (ADNI): clinical characterization”. Neurology. 2010. [2]“Structural and Functional Brain Networks: From Connections to Cognition.” Science. 2013 [3]“The Human Connectome: A Structural Description of the Human Brain.” PLOS Computational Biology. 2005. [4]“Brain Connectivity and Novel Network Measures for Alzheimer’s Disease Classification.” Neurobiology of Aging. 2015.

Reviewer 2: Q3: We take SNVs directly from the nuclide type of the variation marked as ‘Alter: 1’ and ‘primary: 0’, the dimension 2 is two chromosomes. We treat the genomic profile more like sequence data than use it as pre-estimated additive effects {0,1,2}.

Reviewer 3: Q2: We found that the pre-training stage of our model is data-hungry, which requires that we put all the available subjects with paired data into the training. For now, we use subjects with a solo visit as test data to prevent potential data leakage. We will extract a subset from the validation set as the test for improvements of this work. Q3: We totally agree with this point. In the experiment, we actually have the results of the ROC AUC, and it shows a consistent trend with the F1 score. We keep only the F1 score due to the limited space of the chart. Q4: The x-axis in Figure 3 corresponds to the ID of the 8 different experts employed in the model. The figure emphasizes the different activating patterns of these experts across different tasks, and the absence of the experts does not hinder overall comprehension. We will improve the figure in the further extension of this work.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top