Pith. sign in

REVIEW 4 major objections 6 minor 68 references

GenDMR: A dynamic multimodal role-swapping network for identifying risk gene phenotypes

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read GenDMR, a network that swaps teacher-student roles between brain images and SNP data, reaches 94.75% AUC for Alzheimer's and flags 12 potential risk genes including APOE.

desk verdict The role-swapping distillation idea is fresh, but the SNP preselection on the same ADNI labels and the k=12 tuning on the same test data make the reported SOTA and the 12-gene ranking unsupported. read the letter →

arxiv 2506.01456 v1 pith:VGYXH4QF submitted 2025-06-02 q-bio.GN cs.AIcs.LGq-bio.NC

Lina Qin , Cheng Zhu , Chuqi Zhou , Yukun Huang , Jiayi Zhu , Ping Liang , Jinju Wang , Yixing Huang
show 3 more authors
Cheng Luo Dezhong Yao Ying Tan
This is my paper
classification q-bio.GNcs.AIcs.LGq-bio.NC
keywords Alzheimer'sdiseaseimaginggeneticsSNPencodingmulti-instancelearningdynamicknowledgedistillationattentioninterpretabilitysMRIriskgeneidentification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

GenDMR is a deep-learning pipeline that fuses structural MRI (sMRI) with single-nucleotide polymorphism (SNP) data for Alzheimer's disease (AD). Its central proposal is that the weaker modality, genetics, should sometimes act as the teacher for the stronger imaging modality, so the genetic signal is learned rather than suppressed, and that SNPs should be encoded with their chromosomal positions rather than as independent values. The paper reports that on the ADNI cohort this design reaches the best performance among the compared models (AUC 94.75±0.33%, ACC 89.81±0.62%, with paired t-tests at p<0.05), and that the attention scores it produces stabilize into a ranked list of 12 potential high-risk AD genes, headed by VKORC1 and including APOE. A sympathetic reader would care because current imaging-genetics models tend to treat genetic features as a weak auxiliary that only helps classification marginally, whereas GenDMR offers a mechanism for extracting their unique biological value.

What carries the argument

The central mechanism is the dynamic teacher-student role exchange between modalities. The DMSM selects the teacher each training epoch by measuring each modality's posterior entropy (uncertainty) and U-shaped distribution score (discriminability), defaulting to sMRI when the two indicators conflict; the CSM then applies KL-divergence distillation from teacher to student, an uncertainty-aware loss, and a cross-modal contrastive loss, so both modalities co-update. Around that mechanism sit the chromosome-aware SNP encoding (chromosome index plus normalized position plus genotype one-hot), the GRU and ResNet-50 feature generators, and the multi-instance attention module with dual-path local/global attention, gated fusion, and Top-k selection that produces the per-SNP and per-slice interpretable scores.

What would settle it

Refit the full pipeline with 30 SNPs selected without using the diagnosis in the evaluation cohort—for example, from independent genome-wide association summary statistics—and check whether AUC and the 12 attention-prioritized genes are preserved; if AUC collapses toward the SNP-only baseline of 58.50% or the gene list changes entirely, the reported result is an artifact of label-preselecting SNPs.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a dynamic teacher-student role-swapping mechanism, combined with chromosome-aware SNP encoding and multi-instance attention, can push multimodal AD classification beyond models that treat sMRI as the fixed dominant modality. Each SNP is encoded as a concatenation of chromosome number, normalized physical position, and one-hot genotype; a GRU reads the SNP sequence, ResNet-50 extracts sMRI slice features, and a multi-instance attention module (MAM) scores every SNP and slice instance. A dominant modality selection module (DMSM) compares the modalities each epoch using posterior entropy and a U-shaped distribution score, naming one modality teacher and the other student, and a contrastive self-distillation module (CSM) transfers knowledge between them while contrastively aligning same-class features. The full GenDMR model reports AUC 94.75%, ACC 89.81%, PRE 91.63%, and SPE 98.51% on the ADNI cohort, with full-model and module ablation results showing the role-swapping modules contribute the main gains; its SNP attention weights converge on 12 genes the paper interprets as potential AD risk genes, including APOE.

Load-bearing premise

The load-bearing premise is that choosing the 30 input SNPs by how strongly they are associated with Alzheimer's diagnosis in the same cohort used for evaluation does not leak the outcome into the model, so the reported accuracy and the attention-based 12-gene ranking would generalize to independent data.

Editorial extensions

If this is right

  • On the ADNI cohort, the full GenDMR model beats every compared baseline, with the largest margin over the strongest deep-learning baseline IGNet (AUC 91.92% vs 94.75%), so the role-swapping design is the main differentiator rather than the base feature extractors.
  • The SNP encoding comparison attributes a modest but consistent gain to spatial encoding (AUC 93.52% to 94.75%, ACC 88.61% to 89.81%), suggesting that chromosome-position-aware genotype representation is worth adopting in other genetic deep-learning pipelines.
  • The modality ablation shows fusion mainly helps decision boundaries: SNP-alone AUC is 58.50% and sMRI-alone is 93.50%, but the fused model reaches 94.75%, which the paper reads as SNPs reducing misclassification rather than adding independent classification strength.
  • The k-sensitivity analysis fixes the optimal number of retained instances at k=12, with AUC and ACC peaking there and degrading and becoming unstable for k≥20, giving a concrete hyperparameter rule for similar MIL-based fusion models.
  • The 12 attention-prioritized genes—VKORC1, TLE3, NAV3, TMEM132B, NUAK1, RBL2, PCP4, TRAPPC12, APOE, IFNAR2, GALNT18, CKAP4—are proposed by the paper as potential AD risk genes ready for biological follow-up.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 12-gene ranking should be treated as hypothesis-generating: because the 30 input SNPs were screened using AD status in the same ADNI cohort, the attention scores may partly reflect that preselection, and independent-cohort replication is needed before any of these genes is called a confirmed risk gene.
  • The dynamic role-swapping idea transfers to other multimodal problems with imbalanced modality strength, such as pathology slides plus genomics or clinical notes plus lab values, where the standard fixed-teacher distillation would discard the weak modality's unique signal.
  • A natural extension the paper does not test is replacing the linearly normalized SNP position with linkage-disequilibrium-aware distances or recombination maps, which would make the encoding reflect actual inheritance structure rather than raw genomic coordinates.
  • A decisive external check would be to test whether the 12 attention-prioritized SNPs are enriched for genome-wide significant signals from an independent AD consortium; the paper does not report such an enrichment analysis.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes GenDMR, a multimodal deep learning framework that fuses sMRI and SNP data for Alzheimer's disease classification. The method introduces a spatial encoding of SNPs, a multi-instance attention module for interpretable SNP scoring, and a dynamic teacher-student role-swapping mechanism based on a dominant-modality selection module and contrastive self-distillation. On the ADNI dataset, the authors report state-of-the-art performance (AUC 94.75%, ACC 89.81%) and identify 12 potential high-risk AD genes, including APOE, via attention weights. The paper claims that GenDMR improves upon existing multimodal fusion methods and provides interpretable genetic risk assessment.

Significance. If the reported results were valid, GenDMR would be a meaningful contribution to imaging-genetics fusion, particularly in its attempt to encode SNP spatial context and to adaptively balance modality roles during training. The paper also provides a useful design idea in treating SNPs as instances in a multiple-instance learning framework. However, the experimental validation is compromised by two forms of label leakage: the SNP feature set is preselected using AD status from the same cohort used for evaluation, and the hyperparameter k is tuned on the same evaluation splits. These issues directly affect the central claims of state-of-the-art classification and independent risk-gene discovery. The paper does not provide code, an external validation cohort, or a nested cross-validation procedure that would bound the resulting bias. The strengths are the conceptual contributions and the clarity of the proposed architecture, but the empirical evidence as presented does not support the headline conclusions.

major comments (4)
  1. [Section 4.1, Table 2] The 30-SNP input panel is selected using GEMMA mixed linear regression with AD status as the outcome on the same ADNI cohort, combined with AlzData significant SNPs, before the cross-validation splits described in Section 4.2. This means the held-out folds' labels have already influenced which SNPs enter the model, a form of feature-selection leakage. As a result, the AUC and accuracy scores in Table 3 are optimistically biased, and the claim of state-of-the-art performance is not supported by the reported experiments. No external cohort or nested cross-validation is provided to quantify the magnitude of this bias.
  2. [Section 5.1, Fig. 7] The choice of k=12 is made by locating the peak of AUC and ACC on the same evaluation protocol used to produce the final results. This constitutes hyperparameter selection on the test data. Because all models in Table 3 and all ablation conditions use this selected k, the reported comparisons and the paired t-test claim of 'overall optimal performance' in Section 4.3 are not reliable. A proper evaluation would fix k a priori or use an inner validation loop.
  3. [Section 5.2.1] The 12 'potential high-risk genes' are the top-attention SNPs selected from the 30 SNPs that were already screened for AD association in the same cohort. The attention ranking therefore largely recapitulates the screening signal (including AlzData hits and known APOE associations) rather than providing independent discovery. Without replication in an independent cohort or comparison against a genome-wide analysis without same-cohort preselection, the interpretability claim is circular and does not demonstrate new biological insight.
  4. [Table 7, Section 4.4.3] The SNP-only model achieves an AUC of only 58.50±0.50, which is near chance, and the full model's AUC improvement over sMRI-only is 94.75 vs 93.50, a difference of 1.25 percentage points with overlapping standard deviations. The paper does not report a statistical test for this specific comparison, so the claim that SNP features contribute meaningful complementary information by 'optimizing decision boundaries' is not established. This weakens the motivation for the fusion approach as evaluated.
minor comments (6)
  1. [Section 3.2, Eq. (13)] The 'max-softmax normalization' notation is unclear: the max in the denominator should be explicitly defined over the instance index j, and the expression should be written as a proper softmax over the attention logits to avoid ambiguity.
  2. [Section 3.3, Eq. (18)] The U-shaped distribution score is introduced without a formal derivation or citation. Please provide a justification or reference explaining why this variance-based score measures discriminability for binary classification.
  3. [Section 4.3] There are typographical errors: 'Model comprisions' should be 'Model comparisons', and 'DesNet' should be 'DenseNet' throughout the text and tables.
  4. [Section 4.4.1, Table 5] The phrase 'dynamic faculty-student interchange mechanism' appears to be a typo for 'teacher-student interchange mechanism'; please correct it.
  5. [Section 5.2.1] In the list of the 12 genes, 'VKO RC1' should be 'VKORC1' (the space appears to be a typographical artifact).
  6. [Section 4.2] The description 'five 50% discounted cross-validations' is ambiguous. Please specify whether these are five random 50/50 splits with retraining, and clarify how the reported mean and standard deviation are computed across splits.

Circularity Check

3 steps flagged · score 7.0 of 10

SNP preselection on the same ADNI labels and k=12 tuning on the same evaluation folds make the reported SOTA metrics and the 12-gene attention 'discovery' partially circular.

  1. fitted input called prediction [Section 4.1, SNP quality check and screening; Section 4.2, Implementation details; Table 2]
    "After QC this study did gemma mixed linear regression with age (age) and gender (gender) and the Brief Mental State Examination Scale (MMSE) as covariates (cov) to determine the association of each SNP with AD. (6) Using SNPs highly associated with AD and significant SNPs analyzed by gemma mixed linear regression in AlzData (AD database), 30 SNPs were screened as shown in Table 2 and used as the input data set for the study."

    The 30-SNP input set is selected by regressing SNP genotypes on AD status in the same ADNI cohort used for all five evaluation folds (Section 4.2). Every downstream model and the attention module therefore see features that are already a function of the labels used to compute the reported AUC/ACC and to define 'risk'. The model cannot discover AD associations independently; it can only re-rank variants that were preselected for AD association in the same data. The reported SOTA metrics inherit this label information, so the feature-selection step is effectively part of the fitted model rather than an external prior.

  2. fitted input called prediction [Section 5.2.1, Discussion on the Interpretability of SNPs; Abstract; Contribution 3]
    "we considered the most important 12 of all SNPs to be potentially high-risk SNPs associated with AD. As shown in Fig 8, we use the SNP attention weight scores saved on the validation set for heat mapping."

    The '12 potential high-risk genes' are not a genome-wide discovery: they are the top-12 attention weights over the 30 SNPs that Section 4.1 already filtered for AD association via GEMMA on the same cohort plus AlzData. The list includes the canonical APOE (rs439401, an AlzData-flagged SNP) and other variants already significant in Table 2. Attention to a label-preselected panel cannot provide independent evidence of risk-gene status; the interpretability result reduces to re-emitting the association signal that was used to build the input set.

1 more flagged steps
  1. fitted input called prediction [Section 5.1, Sensitivity analysis; Figure 7; Table 3]
    "The model achieves peak performance at k = 12, after which both metrics decline and become more unstable as k increases. ... In summary, we choose k = 12 as the optimal number of features to ensure the best classification performance and the smallest fluctuation."

    The sensitivity analysis scans k using the same five-fold evaluation protocol that produces the headline numbers in Table 3, and k=12 is selected because it maximizes AUC/ACC on those folds. Reporting the k=12 row (AUC 94.75%, ACC 89.81%) as the model's state-of-the-art performance is therefore the result of test-data hyperparameter selection, not an unbiased evaluation of a fixed model. The same k also fixes the number of 'high-risk' SNPs in Section 5.2.1, coupling the discovery claim to label-informed tuning.

full rationale

The architecture of GenDMR itself is not circular: the GRU/ResNet feature generation, multi-instance attention, dominant-modality selection, and contrastive self-distillation are substantive methodological proposals, and the module ablations are internally consistent. No load-bearing self-citation or imported uniqueness theorem was found. However, the paper's two headline outcomes — state-of-the-art classification and the identification of 12 potential AD risk genes — rest on a label-contaminated input pipeline. Section 4.1 uses GEMMA mixed linear regression with AD status as the outcome on the same ADNI cohort to screen 30 SNPs before any split, so the feature set is a function of the labels used later for evaluation and ranking. The k=12 hyperparameter is then selected as the peak AUC/ACC on the same five-fold protocol, making the reported 94.75% AUC / 89.81% ACC a tuned maximum rather than an unbiased estimate. Because the attention weights rank only the preselected 30-SNP panel, the gene list in Section 5.2.1 largely recapitulates known GWAS/AlzData associations (including APOE) and cannot support independent risk-gene discovery. The relative comparison in Table 3 may still be fair among methods, since all models share the same input panel, but the absolute SOTA claim and the interpretability claim are circular in the sense defined above. No external validation or released code is provided, so the magnitude of the leakage cannot be bounded.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central empirical claims rest on the 30-SNP preselection (Section 4.1), the MIL bag assumption (Section 2.2), the U-shape score heuristic (Section 3.3), and several hand-set hyperparameters (Section 3.2-3.4). The 30-SNP list and the k=12 choice are the most consequential because both are derived from the same ADNI labels used for evaluation. No new physical or biological entities are introduced.

free parameters (4)
  • k (number of top instances/SNPs selected by MAM) = 12
    Chosen from sensitivity analysis on the same ADNI data; AUC peaks at k=12 (Section 5.1, Fig. 7), so the reported performance is optimized on the evaluation data.
  • learnable attention temperatures tau_loc, tau_glob = learned during training
    Equations 7 and 9 introduce learnable temperature coefficients controlling softmax smoothness in the Multi-Instance Attention Module.
  • adaptive scaling factors gamma_loc, gamma_glob = constrained to (0.3, 3.0)
    Equations 8 and 10 use adaptive scaling factors to amplify local and global features; these are learned per the constraint.
  • preset hyperparameters (beta_g=0.2, beta_f=0.2, beta_com=10, lambda_a=0.2, lambda_b=0.1, lambda_c=0.1, lambda_d=0.1) = as listed in Section 3.2-3.4
    Hand-chosen constants for gating, feature scaling, and loss weighting; no sensitivity analysis is provided for them.
assumptions (5)
  • domain assumption A subject is a bag; it is AD-positive if at least one instance (SNP locus or sMRI slice) is positive (Eq. 4, Section 2.2).
    This is the multiple-instance learning assumption applied to genetics and brain slices; there is no instance-level ground truth, so the model is trained only on bag labels.
  • domain assumption SNPs physically close on a chromosome are non-randomly associated (linkage disequilibrium), so encoding chromosome number and normalized position captures functional genetic context.
    Section 2.1 invokes LD to justify the spatial encoding; this is a background genetics assumption, not proven in this dataset.
  • ad hoc to paper The 30 SNPs selected by GEMMA mixed linear regression and AlzData on this ADNI cohort are the appropriate genetic features for AD classification.
    Section 4.1 screens from the full genotype array using the same cohort's AD labels; this is a paper-specific selection that may leak label information and is not externally validated.
  • ad hoc to paper Attention weights from the validation set reflect biological disease risk of the corresponding SNPs.
    Section 5.2.1 interprets the top-12 attended SNPs as 'potential high-risk genes' without any external validation or statistical significance testing against known GWAS.
  • ad hoc to paper The U-shaped distribution score (Eq. 18) measures modality discriminability for binary classification.
    Section 3.3 introduces this scoring heuristic without prior validation or comparison to other calibration measures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GenDMR: A dynamic multimodal role-swapping network for identifying risk gene phenotypes." pith.science (2026). https://pith.science/paper/VGYXH4QF

@misc{pith2026250601456,
  author       = {Pith},
  title        = {Pith review of: GenDMR: A dynamic multimodal role-swapping network for identifying risk gene phenotypes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VGYXH4QF}},
  note         = {Machine review of arXiv:2506.01456}
}
read the original abstract

Recent studies have shown that integrating multimodal data fusion techniques for imaging and genetic features is beneficial for the etiological analysis and predictive diagnosis of Alzheimer's disease (AD). However, there are several critical flaws in current deep learning methods. Firstly, there has been insufficient discussion and exploration regarding the selection and encoding of genetic information. Secondly, due to the significantly superior classification value of AD imaging features compared to genetic features, many studies in multimodal fusion emphasize the strengths of imaging features, actively mitigating the influence of weaker features, thereby diminishing the learning of the unique value of genetic features. To address this issue, this study proposes the dynamic multimodal role-swapping network (GenDMR). In GenDMR, we develop a novel approach to encode the spatial organization of single nucleotide polymorphisms (SNPs), enhancing the representation of their genomic context. Additionally, to adaptively quantify the disease risk of SNPs and brain region, we propose a multi-instance attention module to enhance model interpretability. Furthermore, we introduce a dominant modality selection module and a contrastive self-distillation module, combining them to achieve a dynamic teacher-student role exchange mechanism based on dominant and auxiliary modalities for bidirectional co-updating of different modal data. Finally, GenDMR achieves state-of-the-art performance on the ADNI public dataset and visualizes attention to different SNPs, focusing on confirming 12 potential high-risk genes related to AD, including the most classic APOE and recently highlighted significant risk genes. This demonstrates GenDMR's interpretable analytical capability in exploring AD genetic features, providing new insights and perspectives for the development of multimodal data fusion techniques.

Figures

Figures reproduced from arXiv: 2506.01456 by the authors.

Figure 4
Figure 4. Framework of our method. (A) SNP data is processed using a Gate Recurrent Unit (GRU) [28], and sMRI features are extracted using ResNet-50 [29]. The features from both modalities serve as inputs to the subsequent model. (B) A Multi-Instance Attention Module (MAM) is used to calculate the weights of SNP instances to capture key SNPs. (C) A Dominant Modality Selection Module (DMSM) determines the dominant and auxiliar… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 67 canonical work pages

  1. [1]

    Alzheimer's & dementia : the journal of the Alzheimer's Association, 2023, 19(4): 1598-695

    2023 Alzheimer's disease facts and figures [J]. Alzheimer's & dementia : the journal of the Alzheimer's Association, 2023, 19(4): 1598-695

    work page 2023

  2. [2]

    Disentangling the heterogeneity of brain atrophy in Alzheimer’s Disease [J]

    LIU H, TOLEDO J B, GROTHE M J, et al. Disentangling the heterogeneity of brain atrophy in Alzheimer’s Disease [J]. Alzheimer's & Dementia, 2023, 19(S3): e067447

    work page 2023

  3. [3]

    Alzheimer's Disease: Epidemiology and Clinical Progression [J]

    TAHAMI MONFARED A A, BYRNES M J, WHITE L A, et al. Alzheimer's Disease: Epidemiology and Clinical Progression [J]. Neurology and therapy, 2022, 11(2): 553 -69

    work page 2022

  4. [4]

    Hypergraph -regularized multimodal learning by graph diffusion for imaging genetics based Alzheimer's Disease diagnosis [J]

    WANG M, SHAO W, HUANG S, et al. Hypergraph -regularized multimodal learning by graph diffusion for imaging genetics based Alzheimer's Disease diagnosis [J]. Medical image analysis, 2023, 89: 102883

    work page 2023

  5. [5]

    A Deep Generative -Discriminative Learning for Multimodal Representation in Imaging Genetics [J]

    KO W, JUNG W, JEON E, et al. A Deep Generative -Discriminative Learning for Multimodal Representation in Imaging Genetics [J]. IEEE transactions on medical imaging, 2022, 41(9): 2348 -59

    work page 2022

  6. [6]

    Linkage disequilibrium --understanding the evolutionary past and mapping the medical future [J]

    SLATKIN M. Linkage disequilibrium --understanding the evolutionary past and mapping the medical future [J]. Nature reviews Genetics, 2008, 9(6): 477-85

    work page 2008

  7. [7]

    Opening the Black Box: Interpretable Machine Learning for Geneticists [J]

    AZODI C B, TANG J, SHIU S H. Opening the Black Box: Interpretable Machine Learning for Geneticists [J]. Trends in genetics : TIG, 2020, 36(6): 442-55

    work page 2020

  8. [8]

    Attention -based Deep Multiple Instance Learning [Z]//JENNIFER D, ANDREAS K

    ILSE M, TOMCZAK J, WELLING M. Attention -based Deep Multiple Instance Learning [Z]//JENNIFER D, ANDREAS K. Proceedings of the 35th International Conference on Machine Learning. Proceedings of Machine Learning Research; PMLR. 2018: 2127 --36

    work page 2018

Show all 68 references
  1. [9]

    Multimodal deep learning models for early detection of Alzheimer’s disease stage [J]

    VENUGOPALAN J, TONG L, HASSANZADEH H R, et al. Multimodal deep learning models for early detection of Alzheimer’s disease stage [J]. Scientific reports, 2021, 11(1): 3254

    2021

  2. [10]

    A Transformer -Based Knowledge Distillation Network for Cortical Cataract Grading [J]

    WANG J, XU Z, ZHENG W, et al. A Transformer -Based Knowledge Distillation Network for Cortical Cataract Grading [J]. IEEE transactions on medical imaging, 2024, 43(3): 1089 -101

    2024

  3. [11]

    Comprehensive learning and adaptive teaching: Distilling multi- modal knowledge for pathological glioma grading [J]

    XING X, ZHU M, CHEN Z, et al. Comprehensive learning and adaptive teaching: Distilling multi- modal knowledge for pathological glioma grading [J]. Medical image analysis, 2024, 91: 102990

    2024

  4. [12]

    Predictive classification of Alzheimer's disease using brain imaging and genetic data [J]

    SHENG J, XIN Y, ZHANG Q, et al. Predictive classification of Alzheimer's disease using brain imaging and genetic data [J]. Scientific reports, 2022, 12(1): 2405

    2022

  5. [13]

    Distilling the Knowledge in a Neural Network [J]

    HINTON G, VINYALS O, DEAN J. Distilling the Knowledge in a Neural Network [J]. Computer Science, 2015, 14(7): 38-9

    2015

  6. [14]

    A Self-organizing Deep Auto-Encoder approach for Classification of Complex Diseases using SNP Genomics Data [J]

    PIRMORADI S, TESHNEHLAB M, ZARGHAMI N, et al. A Self-organizing Deep Auto-Encoder approach for Classification of Complex Diseases using SNP Genomics Data [J]. Applied Soft Computing, 2020, 97: 106718

    2020

  7. [15]

    A novel representation of genomic sequences for taxonomic clustering and visualization by means of self -organizing maps [J]

    DELGADO S, MOR áN F, MORA A, et al. A novel representation of genomic sequences for taxonomic clustering and visualization by means of self -organizing maps [J]. Bioinformatics (Oxford, England), 2015, 31(5): 736-44

    2015

  8. [16]

    BOOST: A fast approach to detecting gene -gene interactions in genome-wide case-control studies [J]

    WAN X, YANG C, YANG Q, et al. BOOST: A fast approach to detecting gene -gene interactions in genome-wide case-control studies [J]. American journal of human genetics, 2010, 87(3): 325 -40

    2010

  9. [17]

    Informative SNP selection methods based on SNP prediction [J]

    HE J, ZELIKOVSKY A. Informative SNP selection methods based on SNP prediction [J]. IEEE transactions on nanobioscience, 2007, 6(1): 60-7

    2007

  10. [18]

    Machine learning -based models for genomic predicting neoadjuvant chemotherapeutic sensitivity in cervical cancer [J]

    GUO L, WANG W, XIE X, et al. Machine learning -based models for genomic predicting neoadjuvant chemotherapeutic sensitivity in cervical cancer [J]. Biomedicine & pharmacotherapy = Biomedecine & pharmacotherapie, 2023, 159: 114256

    2023

  11. [19]

    Landmark-based deep multi-instance learning for brain disease diagnosis [J]

    LIU M, ZHANG J, ADELI E, et al. Landmark-based deep multi-instance learning for brain disease diagnosis [J]. Medical image analysis, 2018, 43: 157-68

    2018

  12. [20]

    MRI -based Alzheimer's disease prediction via distilling the knowledge in multi-modal data [J]

    GUAN H, WANG C, TAO D. MRI -based Alzheimer's disease prediction via distilling the knowledge in multi-modal data [J]. NeuroImage, 2021, 244: 118586

    2021

  13. [21]

    Deep Multi -Instance Learning Using Multi-Modal Data for Diagnosis of Lymphocytosis [J]

    SAHASRABUDHE M, SUJOBERT P, ZACHARAKI E I, et al. Deep Multi -Instance Learning Using Multi-Modal Data for Diagnosis of Lymphocytosis [J]. IEEE journal of biomedical and health informatics, 2021, 25(6): 2125-36

    2021

  14. [22]

    Solving the multiple instance problem with axis-parallel rectangles [J]

    DIETTERICH T G, LATHROP R H, LOZANO-PéREZ T. Solving the multiple instance problem with axis-parallel rectangles [J]. Artificial Intelligence, 1997, 89(1): 31-71

    1997

  15. [23]

    Attention -based Deep Multiple Instance Learning [J]

    ILSE M, TOMCZAK J M, WELLING M. Attention -based Deep Multiple Instance Learning [J]. 2018

    2018

  16. [24]

    Regional cerebral atrophy correlates with amyloid accumulation in preclinical Alzheimer's disease [J]

    IMABAYASHI E, ISHII K, TOYOHARA J, et al. Regional cerebral atrophy correlates with amyloid accumulation in preclinical Alzheimer's disease [J]. Alzheimer's & Dementia, 2023, 19(Sup3): 1

    2023

  17. [25]

    Enhanced polygenic risk score incorporating gene -environment interaction suggests the association of major depressive disorder with cardiac and lung function [J]

    PAN C, CHENG B, QIN X, et al. Enhanced polygenic risk score incorporating gene -environment interaction suggests the association of major depressive disorder with cardiac and lung function [J]. Briefings in bioinformatics, 2024, 25(2)

    2024

  18. [26]

    CXCR6 orchestrates brain CD8(+) T cell residency and limits mouse Alzheimer's disease pathology [J]

    SU W, SARAVIA J, RISCH I, et al. CXCR6 orchestrates brain CD8(+) T cell residency and limits mouse Alzheimer's disease pathology [J]. Nature immunology, 2023, 24(10): 1735 -47

    2023

  19. [27]

    Be Your Own Teacher: Improve the Performance of Convolutional Neural Networks via Self Distillation [J]

    ZHANG L, SONG J, GAO A, et al. Be Your Own Teacher: Improve the Performance of Convolutional Neural Networks via Self Distillation [J]. IEEE, 2019

    2019

  20. [28]

    Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling [J]

    CHUNG J, GULCEHRE C, CHO K H, et al. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling [J]. Eprint Arxiv, 2014

    2014

  21. [29]

    Deep Residual Learning for Image Recognition [J]

    HE K, ZHANG X, REN S, et al. Deep Residual Learning for Image Recognition [J]. IEEE, 2016

    2016

  22. [30]

    Adam: A Method for Stochastic Optimization [J]

    KINGMA D, BA J. Adam: A Method for Stochastic Optimization [J]. Computer Science, 2014

    2014

  23. [31]

    XGBoost: A Scalable Tree Boosting System [J]

    CHEN T, GUESTRIN C. XGBoost: A Scalable Tree Boosting System [J]. ACM, 2016

    2016

  24. [32]

    Random Forests [J]

    BREIMAN L. Random Forests [J]. Machine Learning, 2001, 45(1): 5-32

    2001

  25. [33]

    Support-vector networks [J]

    CORTES C, VAPNIK V. Support-vector networks [J]. Machine Learning, 1995, 20(3): 273-97

    1995

  26. [34]

    Multimodal integration of neuroimaging and genetic data for the diagnosis of mood disorders based on computer vision models [J]

    LEE S, CHO Y, JI Y, et al. Multimodal integration of neuroimaging and genetic data for the diagnosis of mood disorders based on computer vision models [J]. Journal of psychiatric research, 2024, 172: 144-55

    2024

  27. [35]

    Multi -modal deep learning from imaging genomic data for schizophrenia classification [J]

    KANYAL A, MAZUMDER B, CALHOUN V D, et al. Multi -modal deep learning from imaging genomic data for schizophrenia classification [J]. Frontiers in psychiatry, 2024, 15: 1384842

    2024

  28. [36]

    LEE S, LEE J, LEE M, et al. Integrating Genetic Information for Early Alzheimer ’s Diagnosis through MRI Interpretation; proceedings of the 2023 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI), F 15-18 Oct. 2023, 2023 [C]

    2023

  29. [37]

    Alzheimer's Disease Classification Through Imaging Genetic Data With IGnet [J]

    WANG J X, LI Y, LI X, et al. Alzheimer's Disease Classification Through Imaging Genetic Data With IGnet [J]. Frontiers in neuroscience, 2022, 16: 846638

    2022

  30. [38]

    BrainGNN: Interpretable Brain Graph Neural Network for fMRI Analysis [J]

    LI X, ZHOU Y, DVORNEK N, et al. BrainGNN: Interpretable Brain Graph Neural Network for fMRI Analysis [J]. Medical image analysis, 2021, 74: 102233

    2021

  31. [39]

    Robustness in machine learning explanations: does it matter? [Z]

    HANCOX-LI L. Robustness in machine learning explanations: does it matter? [Z]. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency. Barcelona, Spain; Association for Computing Machinery. 2020: 640–7.10.1145/3351095.3372836

    2020

  32. [40]

    Sanity checks for saliency maps [Z]

    ADEBAYO J, GILMER J, MUELLY M, et al. Sanity checks for saliency maps [Z]. Proceedings of the 32nd International Conference on Neural Information Processing Systems. Montr éal, Canada; Curran Associates Inc. 2018: 9525–36

    2018

  33. [41]

    Vitamin K and the nervous system: an overview of its actions [J]

    FERLAND G. Vitamin K and the nervous system: an overview of its actions [J]. Advances in nutrition (Bethesda, Md), 2012, 3(2): 204-12

    2012

  34. [42]

    Decrease in naturally occurring antibodies against epitopes of Alzheimer's disease (AD) risk gene products is associated with cognitive decline in AD [J]

    GU D, WANG L, ZHANG N, et al. Decrease in naturally occurring antibodies against epitopes of Alzheimer's disease (AD) risk gene products is associated with cognitive decline in AD [J]. Journal of neuroinflammation, 2023, 20(1): 74

    2023

  35. [43]

    The transcriptional cofactor Tle3 reciprocally controls effector and central memory CD8+ T cell fates [J]

    ZHAO X, HU W, PARK S R, et al. The transcriptional cofactor Tle3 reciprocally controls effector and central memory CD8+ T cell fates [J]. Nature immunology, 2024, 25(2): 294-306

    2024

  36. [44]

    Transducin-like enhancer of split 3 (TLE3) in adipose tissue is increased in situations characterized by decreased PPAR γ gene expression [J]

    ORTEGA F J, SERRANO M, RODRIGUEZ -CUENCA S, et al. Transducin-like enhancer of split 3 (TLE3) in adipose tissue is increased in situations characterized by decreased PPAR γ gene expression [J]. Journal of molecular medicine (Berlin, Germany), 2015, 93(1): 83 -92

    2015

  37. [45]

    Aberrant microRNA expression in the brains of neurodegenerative diseases: miR-29a decreased in Alzheimer disease brains targets neurone navigator 3 [J]

    SHIOYA M, OBAYASHI S, TABUNOKI H, et al. Aberrant microRNA expression in the brains of neurodegenerative diseases: miR-29a decreased in Alzheimer disease brains targets neurone navigator 3 [J]. Neuropathology and applied neurobiology, 2010, 36(4): 320 -30

    2010

  38. [46]

    The impact of blood MCP-1 levels on Alzheimer's disease with genetic variation of UNC5C and NAV3 loci [J]

    HUANG J, WANG Y, STEIN T D, et al. The impact of blood MCP-1 levels on Alzheimer's disease with genetic variation of UNC5C and NAV3 loci [J]. Research square, 2023

    2023

  39. [47]

    Defining Alzheimer's Disease through Proteomic CSF Profiling [J]

    PEñ A-BAUTISTA C, ÁLVAREZ -SáNCHEZ L, BALAGUER Á, et al. Defining Alzheimer's Disease through Proteomic CSF Profiling [J]. Journal of proteome research, 2024, 23(11): 5096 -106

    2024

  40. [48]

    Association of CSF proteins with tau and amyloid β levels in asymptomatic 70-year-olds [J]

    REMNESTå L J, BERGSTRö M S, OLOFSSON J, et al. Association of CSF proteins with tau and amyloid β levels in asymptomatic 70-year-olds [J]. Alzheimer's research & therapy, 2021, 13(1): 54

    2021

  41. [49]

    Clinical and genetic characterization of a progressive RBL2 -associated neurodevelopmental disorder [J]

    AUGHEY G N, CALI E, MAROOFIAN R, et al. Clinical and genetic characterization of a progressive RBL2 -associated neurodevelopmental disorder [J]. Brain : a journal of neurology, 2025, 148(4): 1194-211

    2025

  42. [50]

    Exploration and validation of biomarkers for Alzheimer's disease based on GEO database [J]

    GAN Y, SUN J, YANG D, et al. Exploration and validation of biomarkers for Alzheimer's disease based on GEO database [J]. IBRO Neuroscience Reports, 2025

    2025

  43. [51]

    Reduction of Nuak1 Decreases Tau and Reverses Phenotypes in a Tauopathy Mouse Model [J]

    LASAGNA-REEVES C A, DE HARO M, HAO S, et al. Reduction of Nuak1 Decreases Tau and Reverses Phenotypes in a Tauopathy Mouse Model [J]. Neuron, 2016, 92(2): 407 -18

    2016

  44. [52]

    PCP4 Promotes Alzheimer's Disease Pathogenesis by Affecting Amyloid-β Protein Precursor Processing [J]

    HU D, DONG X, WANG Q, et al. PCP4 Promotes Alzheimer's Disease Pathogenesis by Affecting Amyloid-β Protein Precursor Processing [J]. Journal of Alzheimer's disease : JAD, 2023, 94(2): 737 - 50

    2023

  45. [53]

    GNG13 Is a Potential Marker of the State of Health of Alzheimer's Disease Patients' Cerebellum [J]

    SANFILIPPO C, MUSUMECI G, KAZAKOVA M, et al. GNG13 Is a Potential Marker of the State of Health of Alzheimer's Disease Patients' Cerebellum [J]. Journal of molecular neuroscience : MN, 2021, 71(5): 1046-60

    2021

  46. [54]

    Genome -wide pleiotropy analysis of neuropathological traits related to Alzheimer's disease [J]

    CHUNG J, ZHANG X, ALLEN M, et al. Genome -wide pleiotropy analysis of neuropathological traits related to Alzheimer's disease [J]. Alzheimer's research & therapy, 2018, 10(1): 22

    2018

  47. [55]

    Comparison of methods for multivariate gene -based association tests for complex diseases using common variants [J]

    CHUNG J, JUN G R, DUPUIS J, et al. Comparison of methods for multivariate gene -based association tests for complex diseases using common variants [J]. European journal of human genetics : EJHG, 2019, 27(5): 811-23

    2019

  48. [56]

    Apolipoprotein E and Alzheimer disease: risk, mechanisms and therapy [J]

    LIU C C, LIU C C, KANEKIYO T, et al. Apolipoprotein E and Alzheimer disease: risk, mechanisms and therapy [J]. Nature reviews Neurology, 2013, 9(2): 106 -18

    2013

  49. [57]

    Type -I-interferon signaling drives microglial dysfunction and senescence in human iPSC models of Down syndrome and Alzheimer's disease [J]

    JIN M, XU R, WANG L, et al. Type -I-interferon signaling drives microglial dysfunction and senescence in human iPSC models of Down syndrome and Alzheimer's disease [J]. Cell stem cell, 2022, 29(7): 1135-53.e8

    2022

  50. [58]

    An environment -dependent transcriptional network specifies human microglia identity [J]

    GOSSELIN D, SKOLA D, COUFAL N G, et al. An environment -dependent transcriptional network specifies human microglia identity [J]. Science (New York, NY), 2017, 356(6344)

    2017

  51. [59]

    Genome-wide association study identifies SIAH3 locus influencing the rate of ventricular enlargement in non -demented elders [J]

    LI X, CHU S G, SHEN X N, et al. Genome-wide association study identifies SIAH3 locus influencing the rate of ventricular enlargement in non -demented elders [J]. Aging, 2019, 11(21): 9862 - 74

    2019

  52. [60]

    Polypeptide N -acetylgalactosaminyltransferase-Associated Phenotypes in Mammals [J]

    KATO K, HANSEN L, CLAUSEN H. Polypeptide N -acetylgalactosaminyltransferase-Associated Phenotypes in Mammals [J]. Molecules (Basel, Switzerland), 2021, 26(18)

    2021

  53. [61]

    Transcriptomic and glycomic analyses highlight pathway-specific glycosylation alterations unique to Alzheimer's disease [J]

    TANG X, TENA J, DI LUCENTE J, et al. Transcriptomic and glycomic analyses highlight pathway-specific glycosylation alterations unique to Alzheimer's disease [J]. Scientific reports, 2023, 13(1): 7816

    2023

  54. [62]

    p63 Regulates adult neural precursor and newly born neuron survival to control hippocampal -dependent Behavior [J]

    CANCINO G I, YIU A P, FATT M P, et al. p63 Regulates adult neural precursor and newly born neuron survival to control hippocampal -dependent Behavior [J]. The Journal of neuroscience : the official journal of the Society for Neuroscience, 2013, 33(31): 12569-85

    2013

  55. [63]

    Predicting Alzheimer's disease subtypes and understanding their molecular characteristics in living patients with transcriptomic trajectory profiling [J]

    HUANG X, JANNU A J, SONG Z, et al. Predicting Alzheimer's disease subtypes and understanding their molecular characteristics in living patients with transcriptomic trajectory profiling [J]. Alzheimer's & dementia : the journal of the Alzheimer's Association, 2025, 21(1): e14241

    2025

  56. [64]

    Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single - subject brain [J]

    TZOURIO-MAZOYER N, LANDEAU B, PAPATHANASSIOU D, et al. Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single - subject brain [J]. NeuroImage, 2002, 15(1): 273-89

    2002

  57. [65]

    Distinct regional vulnerability to A β and iron accumulation in post mortem AD brains [J]

    YAO J, LI Z, ZHOU Z, et al. Distinct regional vulnerability to A β and iron accumulation in post mortem AD brains [J]. Alzheimer's & dementia : the journal of the Alzheimer's Association, 2024, 20(10): 6984-97

    2024

  58. [66]

    Therapeutic potential and mechanisms of repetitive transcranial magnetic stimulation in Alzheimer's disease: a literature review [J]

    ZHANG X, ZHU L, LI Y, et al. Therapeutic potential and mechanisms of repetitive transcranial magnetic stimulation in Alzheimer's disease: a literature review [J]. European journal of medical research, 2025, 30(1): 233

    2025

  59. [67]

    β2-Microglobulin coaggregates with Aβ and contributes to amyloid pathology and cognitive deficits in Alzheimer’s disease model mice [J]

    ZHAO Y, ZHENG Q, HONG Y, et al. β2-Microglobulin coaggregates with Aβ and contributes to amyloid pathology and cognitive deficits in Alzheimer’s disease model mice [J]. Nature Neuroscience, 2023, 26(7): 1170-84

    2023

  60. [68]

    APOE4/4 is linked to damaging lipid droplets in Alzheimer’s disease microglia [J]

    HANEY M S, PáLOVICS R, MUNSON C N, et al. APOE4/4 is linked to damaging lipid droplets in Alzheimer’s disease microglia [J]. Nature, 2024, 628(8006): 154-61

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.