CoChat
Sign Up
Cited by 14Alzheimer's disease research and treatmentsDementia and Cognitive Impairment ResearchGenetic Associations and Epidemiology

A simulative deep learning model of SNP interactions on chromosome 19 for predicting Alzheimer's disease risk and rates of disease progression

Open access
Bae, Jinhyeong, Logan, Paige E., Acri, Dominic J., Bharthur, Apoorva, Nho, Kwangsik, Saykin, Andrew J., Risacher, Shannon L., Nudelman, Kelly +6 more·2023·Alzheimer's & Dementia
DOI 10.1002/alz.13319PMC PMC10770299
Open this paper in CoChat — free accountOpen PDF

Abstract

BACKGROUND: Identifying genetic patterns that contribute to Alzheimer's disease (AD) is important not only for pre‐symptomatic risk assessment but also for building personalized therapeutic strategies. METHODS: We implemented a novel simulative deep learning model to chromosome 19 genetic data from the Alzheimer's Disease Neuroimaging Initiative and the Imaging and Genetic Biomarkers of Alzheimer's Disease datasets. The model quantified the contribution of each single nucleotide polymorphism (SNP) and their epistatic impact on the likelihood of AD using the occlusion method. The top 35 AD‐risk SNPs in chromosome 19 were identified, and their ability to predict the rate of AD progression was analyzed. RESULTS: Rs561311966 ( APOC1 ) and rs2229918 ( ERCC1/CD3EAP ) were recognized as the most powerful factors influencing AD risk. The top 35 chromosome 19 AD‐risk SNPs were significant predictors of AD progression. DISCUSSION: The model successfully estimated the contribution of AD‐risk SNPs that account for AD progression at the individual level. This can help in building preventive precision medicine.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper develops a Capsule Network deep-learning model using all 266,161 chromosome 19 SNPs from ADNI to predict AD versus CU and quantify individual and epistatic SNP contributions via occlusion. It identifies top chromosome 19 AD-risk SNPs, led by rs56131196 (APOC1) and rs2229918 (ERCC1/CD3EAP), and shows their chromosomal risk impact scores predict rates of cognitive and CSF biomarker progression.

Key findings

  • The deep-learning model achieved 68.18% accuracy (specificity = 72.04%, sensitivity = 62.30%, AUC = 0.67, equal error rate = 0.37), slightly above random guess (60.39%) and APOE-information-only classification (67.53%).
  • Among the top 35 AD SNPs, 7 belonged to APOC1, 7 to TOMM40, 5 to ZNF473, and 3 to VRK3. Rs56131196 (APOC1) was the most significant SNP for predicted AD participants, with occlusion producing a CRIS of −0.023 (−2.3%); rs2229918 (ERCC1/CD3EAP) was the most powerful for predicted CU participants. Co-occlusion produced 574 and 358 amplification and 21 and 170 attenuation interactions for pADp and pCUp, respectively. The APOE SNPs rs429358 and rs7412 were not detected as AD-risk SNPs.
  • The top 35 AD SNPs predicted progression: CRIS of 3 SNPs correlated with rate of cognitive decline (P < .05, r²-adjusted = 0.19–0.43) and with CSF biomarker changes over time (P < .05, r²-adjusted = 0.89–0.99). Replacing rs56131196 A/G with G/G in pADp decreased psAD by 0.079 (7.9%) and led to ≈36% of pADp being predicted as pCUp.
  • In external ImaGENE validation, 26 of the top 35 ADNI pADp SNPs were also identified as strong AD-contributing SNPs; rs56131196 (APOC1) and rs2229918 (ERCC1/CD3EAP) were the most and second most significant AD SNPs with respect to pADp and pCUp.

Methods

  • Used chromosome 19 GWAS data from ADNI: 457 AD and 313 CU participants (N = 770), randomly divided into 60% train (N = 462), 20% validation (N = 154), and 20% test (N = 154) sets; all 266,161 SNPs were used. The model was a Capsule Network with primary capsule layers and dynamic routing; occlusion maps quantified individual SNP and SNP-SNP interaction contributions, and genotype replacement simulated editing effects.
  • Multiple linear regression tested every possible combination of 3 top-35 AD SNPs plus their four interaction terms to predict rate of cognitive decline (memory, language, executive, visuospatial composites) and rate of CSF Aβ, t-tau, p-tau, and ratio changes; covariates were sex, age, education, and APOE ε4 genotype.
  • External validation used ImaGENE: 50 participants (22 who converted to AD dementia and 28 CU who remained CU over 5 years), with 206,756 of the 266,161 ADNI SNPs available.

Limitations

  • Stated: the proof-of-concept model used only chromosome 19, which the authors believed limited predictive performance; future work will use multiple/all chromosomes, and inter-chromosomal epistatic interactions were beyond scope.
  • Stated: no single dataset is fully representative of the population, so biological interpretation from the weight matrix at lowest validation loss might not optimally align with nature; a different model, dataset, or preprocessing might identify different features.
  • Evident: the external validation sample was small (50 participants), and the ADNI sample was 98.7% White, limiting generalizability; the APOE ε4 haplotype SNPs rs429358 and rs7412 were not detected as predictive despite being established AD risk factors, a surprising result the authors report consistently across models and datasets.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1002/alz.13319

Cites

55 of 55 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 23 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Rs56131196 (APOC1) showed the strongest interactions with other SNPs.

    “Rs56131196 (APOC1) also showed the strongest interactions with other SNPs.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The change in psAD essentially indicates the individual SNP contribution to the CRIS.
  • Findingown data

    Co-occlusion of two AD SNPs at a time resulted in 358 amplification and 170 attenuation interactions of pCUp.

    “Co-occlusion of two AD SNPs at a time resulted in 574 and 358 amplification and 21 and 170 attenuation interactions of pADp and pCUp, respectively (Table S9 in supporting information).”
    Quote in the paperNot a substantial claim
  • Findingreasoning

    Top AD SNPs determined by our model did not include genome-wide significant SNPs from the three latest GWAS research projects.

    “Top AD SNPs determined by our model, however, did not include genome-wide significant SNPs from the three latest GWAS research projects.[13][14][15]”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [13][14][15]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Findingnone given

    Sixty-four test set participants with higher psAD than psCU were classified as pADp.

    “Sixty-four test set participants with higher psAD than psCU were classified as pADp.”
    Quote in the paperSubstantial claim
  • Findingown data

    Replacing rs56131196 A/G with G/G led to approximately 36% of pADp now being predicted as pCUp.

    “This led to ≈36% of pADp now being predicted as pCUp (Figure S12 in supporting information).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The change in psAD essentially indicates the individual SNP contribution to the CRIS.
  • May combine two claims. 27
  • Classified as probably not a substantial claim. 18

  • May combine two claims.

    Findingown data

    The deep learning prediction model achieved an accuracy of 68.18%, specificity 72.04%, sensitivity 62.30%, AUC 0.67, and equal error rate 0.37 on the test set.

    “Our deep learning prediction model achieved an accuracy of 68.18% (specificity = 72.04%,sensitivity = 62.30%, area under the curve [AUC] = 0.67, equal error rate = 0.37 [Figure S1]).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The weakest association between CRIS and rate of cognitive decline was observed for the memory domain and the highest for the executive function domain.

    “The weakest association was observed for the memory domain and the highest for the executive function domain (Table 3 and Tables S16 and S17 in supporting information).”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The change in psAD essentially indicates the individual SNP contribution to the CRIS.

  • May combine two claims.

    Findingown data

    The individual CRIS of rs147510483 (VRK3) and rs149633759 (VRK3) were -0.002 (-0.2%) and -0.004 (-0.4%), respectively, but when removed simultaneously, a -0.010 (-1%) CRIS was observed.

    “Their individual CRIS were -0.002 (-0.2%) and -0.004 (-0.4%), respectively, but when they were removed simultaneously, a -0.010 (i.e., -1%) CRIS was observed.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: If the prediction score change from co-occlusion is higher or lower than the sum of prediction changes of each individual SNP, there is an interaction (amplification or attenuation) between them.

  • May combine two claims.

    Findingown data

    In the test set, 63.93% of AD participants and 30.11% of CU participants were APOE ε4 carriers.

    “Also, in the test set, 63.93% of AD participants as opposed to 30.11% of CU participants were APOE ε4 carriers.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Rs56131196 (APOC1) was recognized as the most powerful individual pADp SNP and the fourth most powerful pCUp SNP.

    “Rs56131196 (APOC1) was recognized as the most powerful individual pADp SNP and the fourth most powerful pCUp SNP;”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    Our model captured the variability of not only the individual SNPs but also their positional relations.

    “Our model captured the variability of not only the individual SNPs but also their positional relations.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Among the top 35 AD SNPs, 7 belonged to APOC1 and another 7 to TOMM40.

    “Among the 35 AD SNPs, 7 belonged to APOC1 and another 7 to TOMM40.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    For the attenuation effect, VRK3 and ERCC1 showed the highest attenuated impact for pADp and pCUp, respectively.

    “For the attenuation effect, VRK3 and ERCC1 showed the highest attenuated impact for pADp and pCUp, respectively.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    79.69% of pADp had an adenine/guanine (A/G) genotype while approximately 88.89% of pCUp had the reference G/G genotype at rs56131196.

    “79.69% of pADp had an adenine/guanine (A/G) genotype while ≈88.89% of pCUp had the reference, G/G, genotype at this locus.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    A small number of SNPs (in this case 3) out of the 35 were powerfully associated with CSF biomarker changes over time and also showed modest association with decline across all cognitive domains.

    “We were able to demonstrate that a small number of SNPs (in this case 3) out of the 35 powerfully associated with CSF biomarker changes over time and also show modest association with decline across all cognitive domains.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Among the top 35 AD SNPs, 5 were from ZNF473 and 3 from VRK3.

    “Additionally, we observed 5 SNPs from zinc finger protein 473 (ZNF473) and 3 SNPs from vaccina-related kinase serine/threonine kinase 3 (VRK3; Table 2, left panel).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Each of rs60229698 (VRK3) and rs149633759 (VRK3) resulted in -0.004 (0.4%) CRIS, yet when removed together, only a -0.004 (-0.4%) CRIS was observed.

    “Each SNP resulted in -0.004 (0.4%) CRIS, yet when removed together, we observed still only a -0.004 (-0.4%)CRIS (Figures S6 and S7 in supporting information).”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: If the prediction score change from co-occlusion is higher or lower than the sum of prediction changes of each individual SNP, there is an interaction (amplification or attenuation) between them.

  • May combine two claims.

    Findingown data

    The APOE ε4 haplotype did not contribute to the model's predictive performance, as it did not affect prediction changes when occluded nor show epistatic effects with any other chromosome 19 SNPs when co-occluded.

    “APOE ε4 haplotype did not affect prediction changes when occluded nor epistatic effects with any other SNPs in chromosome 19 when co-occluded, indicating that it did not contribute to model's predictive performance (see Data S7).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Findingnone given

    CapsNet is able to analyze positional relations between AD risk SNPs, leading to the use of spatial and positional information in our predictive model.

    “CapsNet is able to analyze positional relations between AD risk SNPs leading to the use of spatial and positional information in our predictive model.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Replacement of rs429358 and/or rs7412 showed no change in CRIS and no amplification or attenuation effects.

    “Replacement of rs429358 and/or rs7412 showed no change in CRIS and no amplification or attenuation effects.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The strongest prediction was seen for Aβ/p-tau and the weakest for t-tau.

    “The strongest prediction was seen for Aβ/p-tau and the weakest for t-tau (Table 4).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    All regression models were significant (P < .05, adjusted r2 = 0.89-0.99).

    “All regression models were significant (P < .05,r 2adjusted = 0.89-0.99).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The genetic code is determined by SNP order and location, their spatial relation to each other, and their epistatic interactions with other SNPs.

    “The genetic code is determined by single nucleotide polymorphisms (SNPs) order and location, their spatial relation to each other, and their epistatic interactions with other SNPs.[8][9][10]”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [8]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [9]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [10]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Interpretationown data

    The deep learning model performed slightly better than both the random guess and the APOE information.

    “Therefore, our model performed slightly better than both the random guess and the APOE information.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    A previous study reported that deficient ERCC1 protein leads to imbalance between DNA repair and oxidative stress and neurodegeneration.

    “A previous study reported that deficient ERCC1 protein leads to imbalance between DNA repair and oxidative stress and neurodegeneration. 44”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [44]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    VRK3 is known to progress AD via oxidative stress-induced cyclin-dependent kinase 5 (CDK5).

    “VRK3 was also known to progress AD via oxidative stress-induced cyclin-dependent kinase 5 (CDK5).53”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [53]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    ERCC1 Δ/- mice showed neuronal apoptosis and synaptic plasticity deficits in the hippocampus and accelerated age-dependent cognitive decline.

    “ERCC1 Δ/-mice showed neuronal apoptosis and synaptic plasticity deficits in the hippocampus and accelerated age-dependent cognitive decline.[45][46][47][48][49][50]”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [45]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [46]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [47]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    The difference between top AD SNPs from our model and genome-wide significant SNPs from the three latest GWAS projects is due to differences in diversity of race, population size, and AD risk factor identification process.

    “This is due to the differences in diversity of race, population size, and AD risk factor identification process.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Data-driven methods (machine and deep learning models) have identified new AD-linked SNPs but still fall short in correctly estimating the impact of AD risk and protective variants at the individual level.

    “Data-driven methods (e.g., machine and deep learning models) are cutting-edge tools for pattern recognition and have been applied to GWAS data.[20][21][22][23][24] These methods have identified new AD-linked SNPs, but so far these methods still fall short in correctly estimating the impact of AD risk and protective variants at the individual level.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [20]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [21]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [22]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Using a single chromosome significantly reduces the computational burden to explore the feasibility and effectiveness of this type of model.

    “Using a single chromosome significantly reduces the computational burden to explore the feasibility and effectiveness of this type of model.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationnone given

    Our approach shows promise for clinical implementation as an AD risk assessment tool and for pre-symptomatic clinical trial enrichment with patients at high likelihood for AD.

    “Our approach shows promise for clinical implementation as an AD risk assessment tool and for pre-symptomatic clinical trial enrichment with patients at high likelihood for AD.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Our hypothesis-free deep learning approach identified potential AD risk SNPs that might bring us closer to a full understanding of AD's heritability and personalized genome-level risk assessment.

    “In conclusion, our hypothesis-free deep learning approach identified potential AD risk SNPs that might bring us closer to a full understanding of AD's heritability and personalized genome-level risk assessment.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Many research groups have spotlighted the association between APOC1 and cognitive performance in AD.

    “Many research groups have spotlighted the association between APOC1 and cognitive performance in AD. [37][38][39][40]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [37][38][39][40]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    The biological interpretation originating from the weight matrix at the lowest validation loss in our dataset might not optimally align with the full breadth of observations in nature.

    “Hence, the biological interpretation originating from the weight matrix at the lowest validation loss in our dataset might not optimally align with the full breadth of observations in nature.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: No single dataset is fully representative of the population.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The University of California Los Angeles Institutional Review Boards (UCLA IRBs) approved the study.

    “The University of California Los Angeles Institutional Review Boards (UCLA IRBs) approved the study.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    rs439401 has been previously associated with AD.

    “This SNP has been previously associated with AD. 42,43”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [42]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [43]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The restrictive information of input data (chromosome 19) limited the model's predictive performance.

    “We believed that the restrictive information of input data, that is, chromosome 19, limited the model's predictive performance.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    The model was modified to have a smaller complexity and was trained with multiple regularization methods to manage overfitting.

    “The model was modified to have a smaller complexity and was trained with multiple regularization methods (see Data S4 in supporting information) to manage overfitting.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    All ImaGENE participants or their legally authorized representatives provided informed consent for data collection and publication according to the Declaration of Helsinki, US federal regulations, local state laws and regulations, and policies of the UCLA IRBs.

    “All ImaGENE participants or their legally authorized representatives provided informed consent for data collection and publication according to the Declaration of Helsinki, US federal regulations, local state laws and regulations, and policies of the UCLA IRBs.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductown data

    We validated the findings through three different model architectures, five internal train/validation/test sets, and in an external dataset and produced similar outcomes.

    “To minimize this risk and increase our confidence in the results, we validated the findings through three different model architectures, five internal train/validation/test sets, and in an external dataset and produced similar outcomes.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Prendecki et al. reported that defects in APOC1, in addition to the TOMM40 gene, increase oxidative stress and have an APOE ε4-independent effect on AD progression.

    “In terms of the interactive effects between SNPs in APOC1 and TOMM40 (Figure S6), Prendecki et al. reported that defects in APOC1, in addition to the TOMM40 gene, increase oxidative stress and have an APOE ε4-independent effect on AD progression.41”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [41]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    The positional information and variability in spatial relationships between SNPs and/or SNP sequences is a major component of the missing heritability.

    “It is highly likely that the positional information and variability in spatial relationships between SNPs and/or SNP sequences is a major component of the missing heritability.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Statistical GWAS approaches disregard SNP-SNP interactions.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    This CRISPR simulation technique could provide useful insight into the primary prevention or disease modification opportunities that could be achieved through gene editing.

    “This CRISPR simulation technique could provide useful insight into the primary prevention or disease modification opportunities that could be achieved through gene editing.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Rs56131196 from APOC1 has been previously identified as an APOE ε4-independent AD risk factor associated with hippocampal atrophy.

    “Rs56131196 from APOC1 has been previously identified as an APOE ε4-independent AD risk factor associated with hippocampal atrophy.36”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [36]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

39 of 39

  • The paper relies on this without support.

    Only genetic factors can identify an individual's AD risk prior to any disease activity, rendering primary prevention possible.

    “Yet, only genetic factors can identify one's AD risk prior to any disease activity rendering primary prevention possible.”

  • The paper relies on this without support.

    GWAS does not take epistatic interactions into account.

    “However, GWAS does not take epistatic interactions into account.”

  • The paper relies on this without support.

    Most machine/deep learning models rely on pre-determined feature selection methods for SNP data reduction based on assumptions and/or prior knowledge, which restricts scientific discovery by eliminating data linked to the outcome of interest that is yet to be discovered.

    “First, to avoid overfitting, 25 most machine/deep learning models rely on pre-determined feature selection methods for SNP data reduction based on assumptions and/or prior knowledge.This approach restricts scientific discovery by eliminating data linked to the outcome of interest that is yet to be discovered.”

  • The paper relies on this without support.

    Deep learning methods in genetics have considered only individual variants but not their potential interactions or positional information.

    “Second, deep learning methods in genetics have considered only individual variants but not their potential interactions or positional information.”

  • The paper relies on this without support.

    The change in psAD essentially indicates the individual SNP contribution to the CRIS.

    “The change in psAD essentially indicates the individual SNP contribution to the CRIS.”

  • The paper relies on this without support.

    If the prediction score change from co-occlusion is higher or lower than the sum of prediction changes of each individual SNP, there is an interaction (amplification or attenuation) between them.

    “If the prediction score change was higher or lower than the sum of prediction changes of each individual SNP, we concluded that there was an interaction (amplification or attenuation) between them (Figure S4 in supporting information).”

  • The paper relies on this without support.

    Statistical GWAS approaches disregard SNP-SNP interactions.

    “Statistical GWAS approaches disregard SNP-SNP interactions.”

  • The paper relies on this without support.

    No single dataset is fully representative of the population.

    “No single dataset is fully representative of the population.”

  • The paper relies on this without support.

    The CapsNet approach is the first model that allows for the exploration of the epistatic interactions between AD risk or protective SNPs.

    “To our knowledge, the CapsNet approach is the first model that allows for the exploration of the epistatic interactions between AD risk or protective SNPs.”
9 of 9

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Aβ/p-tau change means cannot describe the same participant partitions

    Major

    Table 1 reports mean rates of −1.1 and −1.7 for the AD and CU groups, but −0.1 for each of the train, validation, and test sets. Because each set mean is −0.1, their combined mean cannot also fall between the two diagnostic-group means.

  • Abstract gives a different identifier for the leading APOC1 SNP

    Minor

    The abstract names rs561311966, whereas the results repeatedly identify the leading APOC1 SNP as rs56131196. These are different SNP identifiers.

  • Training-set size differs between the methods and Table 1

    Minor

    The methods give a training-set size of 462, but Table 1 gives 452. The table's three set sizes total 760 rather than the stated 770 participants.

  • AD and CU participant counts are reversed in Table 1

    Minor

    The methods report 457 CU participants and 313 AD participants, while Table 1 assigns those counts to AD and CU, respectively. This reverses the reported diagnostic group sizes.

4 of 4

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. XiE Capsule network performance on complex data.arXiv preprint arXiv:171203480.2017.

Cited by

Sorted by citations

Related

Authors
Bae, J., Logan, P. E., Acri, D. J., Bharthur, A., Nho, K., Saykin, A. J., Risacher, S. L., Nudelman, K., Polsinelli, A. J., Pentchev, V., Kim, J., Hammers, D. B., Apostolova, L. G., & Initiative, F. T. A. D. N.
Affiliations
  • Indiana University Indianapolis (US)
  • Department of Neurology, School of Medicine Indiana University School of Medicine Indianapolis Indiana USA
  • Indiana University Bloomington (US)
Corresponding author
Apostolova, Liana G.
ORCID
Venue
Alzheimer's & Dementia
Publisher
Wiley
Published
2023-07-06
Volume
19
Issue
12
Pages
5690-5699
ISSN
  • 1552-5260 (ISSN-L)
  • 1552-5279
DOI
10.1002/alz.13319
PMC
PMC10770299
OpenAlex
W4383302036
Semantic Scholar
760f7989b52d7bc0c74e15b2ed57f580ecb9d596
Cited by
14
References
51
Citations by year
2026: 2 · 2025: 7 · 2024: 5
FWCI
2.02
Type
journal article
Language
en
Open access
Licence
Locations
  • Alzheimer s & Dementia (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • IUScholarWorks (Indiana University) (submittedVersion)
  • PubMed Central (submittedVersion)
Funders
National Institutes of Health (U01 AG057195)
Topics
  • Health Sciences › Medicine › Physiology › Alzheimer's disease research and treatments
  • Health Sciences › Medicine › Psychiatry and Mental health › Dementia and Cognitive Impairment Research
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Genetic Associations and Epidemiology
Keywords
Alzheimer's disease, chromosome 19
MeSH terms
Deep Learning, Alzheimer Disease, Chromosomes, Human, Pair 19, Humans, Magnetic Resonance Imaging, Disease Progression, Polymorphism, Single Nucleotide, Neuroimaging

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-17.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-17.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review