Interpretable deep survival analysis of Alzheimer's disease via metabolic genetic variants.

Interpretable deep survival analysis of Alzheimer’s disease via metabolic genetic variants - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Bioinformatics . 2026 Apr 30;42(6):btag213. doi: 10.1093/bioinformatics/btag213 Interpretable deep survival analysis of Alzheimer’s disease via metabolic genetic variants Sungwoo Goo 1, 2 , Soyoung Lee 3, 4, ✉ , Jung-Woo Chae 5, 6, 7, ✉ , Sangkeun Jung 8, 9, ✉ , Hwi-yeol Yun 10, 11, ✉ Editor: Björn Grüning Author information Article notes Copyright and License information 1 College of Pharmacy, Chungnam National University, Daejeon, 34134, Republic of Korea 2 Department of Bio-AI convergence, Chungnam National University, Daejeon, 34134, Republic of Korea 3 College of Pharmacy, Chungnam National University, Daejeon, 34134, Republic of Korea 4 Institute of Drug Research and Development, Chungnam National University, Daejeon, 34134, Republic of Korea 5 College of Pharmacy, Chungnam National University, Daejeon, 34134, Republic of Korea 6 Department of Bio-AI convergence, Chungnam National University, Daejeon, 34134, Republic of Korea 7 Institute of Drug Research and Development, Chungnam National University, Daejeon, 34134, Republic of Korea 8 Department of Bio-AI convergence, Chungnam National University, Daejeon, 34134, Republic of Korea 9 Department of Computer Science and Engineering, Chungnam National University, Daejeon, 34134, Republic of Korea 10 College of Pharmacy, Chungnam National University, Daejeon, 34134, Republic of Korea 11 Department of Bio-AI convergence, Chungnam National University, Daejeon, 34134, Republic of Korea ✉ Corresponding authors. Soyoung Lee, College of Pharmacy, Chungnam National University, Daejeon, 34134, Republic of Korea. E-mail: sy.lee@cnu.ac.kr ; Jung-Woo CHAE, College of Pharmacy, Chungnam National University, Daejeon, 34134, Republic of Korea. E-mail: jwchae@cnu.ac.kr ; Sangkeun Jung, Department of Computer Science and Engineering, Chungnam National University, Daejeon, 34134, Republic of Korea. Email: hugman@cnu.ac.kr ; Hwi-yeol Yun, College of Pharmacy, Chungnam National University, Daejeon, 34134, Republic of Korea. E-mail: hyyun@cnu.ac.kr . Roles Sungwoo Goo : Conceptualization, Data curation, Formal analysis, Methodology, Writing - original draft Soyoung Lee : Supervision Jung-Woo Chae : Supervision Sangkeun Jung : Supervision Hwi-yeol Yun : Project administration, Supervision Björn Grüning : Associate Editor Received 2025 Dec 10; Revised 2026 Apr 8; Accepted 2026 Apr 22; Collection date 2026 Jun. © The Author(s) 2026. Published by Oxford University Press. This is an Open Access article distributed under the terms of the Creative Commons Attribution License ( https://creativecommons.org/licenses/by/4.0/ ), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited. PMC Copyright notice PMCID: PMC13224968  PMID: 42063212 Abstract Background Alzheimer’s disease (AD) is a progressive neurodegenerative disease. Traditional models for estimating AD onset cannot capture nonlinear interactions (epistasis) among the numerous genetic variables that contribute to AD risk. Methods We developed a feedforward neural network (FFN)–Weibull survival model to predict AD onset using large-scale single-nucleotide polymorphism (SNP) data. We integrated an XAI technique, Shapley additive explanations (SHAP), to address the black-box nature of deep learning, interpret model predictions, and quantify the contribution of each genetic factor to AD. Results The FFN model achieved a mean concordance index of 0.647, demonstrating an approximately 3.6% improvement over the traditional linear baseline (0.625). The FFN-SHAP model validated established findings, identifying APOE E4 as a primary AD risk factor. APOE E2 strongly protected against AD. Metabolic-disorder-related SNPs had conflicting effects, suggesting gene–environment interactions influence AD onset. Conclusions By effectively bypassing the combinatorial explosion of interaction terms, the predictive power of an FFN combined with XAI provides a robust methodological tool for identifying the genetic basis of complex diseases, even in cohorts with limited sample sizes. Our model generated novel testable hypotheses regarding the intricate roles of gene–gene and gene–environment interactions in AD pathogenesis. 1 Introduction Alzheimer’s disease (AD) is a neurodegenerative disorder characterized by a gradual and continuous long-term decline in cognitive function. AD onset occurs over time and is not instantaneous; as such, survival analysis provides a more appropriate framework than simple classification models for assessing AD risk. The pathogenesis of AD is not yet fully understood; however, AD is a multifactorial disease arising from the complex interplay among genetic and environmental factors. AD risk prediction has primarily relied on traditional statistical models such as the Cox proportional hazards model. These statistical models are valuable for providing intuitive interpretations of hazard ratios; however, their ability to capture nonlinear relationships and complex epistatic interactions among multiple genetic variables is limited. Specifically, the interaction terms of dozens of genetic variables cannot be simultaneously and directly included in a model to account for epistasis due to practical limitations. Simultaneously including all interaction terms leads to combinatorial explosion, in which the number of combinations considered exponentially grows. Recent studies have successfully applied deep learning survival models to AD risk prediction to overcome these barriers. For instance, approaches utilizing massive cohorts (e.g., UK Biobank, n > 41 000) and integrating both genetic and extensive clinical factors have achieved high predictive accuracy using models like DeepSurv ( Yuan et al . 2024 ). While these large-scale frameworks, along with advanced methods such as Graph Convolutional Networks (GCNs) ( Hamed et al 2025 ), demonstrate excellent performance, acquiring such comprehensive, multi-modal datasets is often restricted by practical and privacy constraints. In contrast, our study addresses a different challenge: extracting meaningful, nonlinear genetic interactions from a much smaller, strictly genetic cohort ( n = 956 ). While the absence of acquired clinical predictors and the limited sample size inherently cap the maximum achievable predictive accuracy compared to those massive multi-modal models, they highlight the critical need for robust methodological frameworks. In this context, our study demonstrates that a compact neural network—utilizing a carefully curated subset of targeted SNPs—can effectively bypass combinatorial explosion and perform robust survival analysis while providing deep biological interpretability. We aimed to overcome these limitations of prior models by developing an approach that integrates deep learning techniques into AD survival analysis. We applied a feedforward neural network (FFN) to address this challenge. FFNs are data-driven methods that flexibly model nonlinear patterns and interactions from high-dimensional genomic data. FFNs automatically learn patterns without requiring manual feature selection or predefined interaction terms, even for large numbers of covariates. FFNs thus identify strong signals and discount noise or less impactful features, which is a key limitation of traditional models that require manual intervention. We combined an FFN with the Weibull survival model to overcome the limitations of traditional linear models and develop a powerful tool for predicting AD onset risk. However, the black-box nature of deep learning models poses a new challenge: difficulty in interpreting the basis of predictions in a clinical context. Medical predictions must be explainable to ensure the reliability and clinical utility of the model. We considered this requirement by using Shapley additive extensions (SHAP), an explainable artificial intelligence (XAI) technique to quantitatively analyze the contribution of each genetic variable to the model predictions. Our goal was to not just build a model but also validate the ability of the FFN-SHAP approach to find a strong signal (APOE) within a high-dimensional noisy dataset of metabolic single-nucleotide polymorphisms (SNPs). Therefore, the two goals of this study were as follows: First, we aimed to develop an FFN-based Weibull survival model that predicts age at AD onset using a high-dimensional genetic dataset. The dataset included the APOE genotype, which is a powerful risk factor for AD, along with a large number of SNPs related to metabolic disorders. These SNPs are also risk factors for AD; however, their effects are weaker than those of APOE and vary in strength. Second, we aimed to verify the methodological utility of this approach. We used SHAP to test whether our model could autonomously identify the dominant contribution of APOE to numerous, less-potent metabolic SNPs. Successfully isolating the strong APOE signal would validate the ability of the FFN to handle complex, high-dimensional genomic data and confirm the biological plausibility of the model. 2 Methods 2.1 Data collection and preprocessing We created a dataset that realistically challenges the ability of a model to learn and shift through high-dimensional, mixed-signal genetic data ( Table 1 ). We selected two types of genetic features. First, we included the APOE genotype, which is the most robust and potent genetic marker of AD. This served as the ground truth and positive control for the interpretability of our model. Second, we collected the 27 SNPs associated with metabolic disorders, specifically dyslipidemia and type 2 diabetes mellitus. This group was chosen for its known biological link to AD pathogenesis. However, the effects on these SNPs on AD risk are likely weaker, more complex, and noisy than those of APOE. This dataset composition was intentional. The dataset embeds powerful signals within a large collection of relevant but less potent features. This design was specifically adopted to test the primary strength of the proposed FFN: the ability of the FN to autonomously identify strong signals and model complex interactions from a high-dimensional input space. Data were obtained from the Clinical & Omics Data Archive(CODA) of the Korea National Institute of Health(KNIH; distribution No.CODA_S2400001-01). The dataset originated from the “Genetic Studies of AD in Korea” project, which was initially approved by the Institutional Review Board of Chosun University (IRB No. 2–1041055-AB-N-01–2022-31). The collected data included demographic and clinical information as well as whole-genome sequencing data in Genome Variant Call Format (gVCF). All genotype position information was based on the human reference genome GRCh38. SNPs were denoted using reference IDs from the NCBI dbSNP database ( Sherry et al 2001 ). We used the gene name (e.g., APOE) for readability in the main text. The coordinate-based representation is provided in the tables and figures to ensure the reproducibility of our methods. The format ‘chromosome: position: reference allele>alternative allele’ is used when necessary to specify the variant. The subjects were 956 adults aged 55 year older, divided into 473 patients diagnosed with AD and a control group of 483 individuals without AD. The study cohort consisted of subjects who were administered the Mini-Mental State Examination and provided sex information. An event was defined as the initial diagnosis of AD during the follow-up period. The 27 SNPs selected are summarized in Table 2 , which included SNPs related to dyslipidemia and T2D. Each SNP genotype was converted into a binary variable. Furthermore, genotype information for APOE was integrated into the model. We split the dataset into 5 folds for cross-validation. Each fold was used as a test set once and the remaining four folds served as the training set ( Figure 1 ). Table 1. Summary of dataset characteristics and feature composition. Category Item Value Subjects Total number of subjects 956 Alzheimer’s Disease (AD) patients 473 Control subjects 483 Features Total input features used in FFN 55 APOE genotypes (E2, E3, E4) 3 Metabolic SNPs (Binary encoded) 27 Non-genetic features (MMSE, Sex) 2 Validation Cross-validation scheme 5-fold Data split ratio (Train/Test) 80:20 Weibull AFT Model Shape parameter ( k shared ) Shared across subjects Scale parameter ( λ individual ) Output of FFN for each subject Open in a new tab Table 2. List of genetic variants associated with metabolic disorders. Gene rsID Variant Ref. DL APOA5 3135506 G > C ( Lai et al. 2016 ) DL APOA5 662799 G > A ( Xu et al. 2013 ) DL CELSR2 12740374 G > T ( Nuinoon et al. 2023 ) DL CETP 708272 G > A ( Karam et al. 2024 ) DL EGLN1 2808607 G > A ( De Castro-Orós et al. 2014 ) DL LPL 328 C > G ( Al-Bustan et al. 2019 ) DL LPL 7007797 T > G ( De Castro-Orós et al. 2014 ) DL MC4R 17782313 T > C ( Huang et al. 2011 ) DL MC4R 571312 C > A ( Yanasegaran et al. 2024 ) DL MLXIPL 17145738 C > G ( Wang et al. 2008 ) DL NNMT 1941404 A > G ( Zhu et al. 2016 ) DL NNMT 694539 C > T ( Hasan et al. 2018 ) DL PCDH15 10825269 C > T ( Huertas-Vazquez et al. 2010 ) DL TBL2 17145738 C > T ( Akimoto et al. 2019 ) T2D CDKAL1 7756992 A > G ( Amin et al. 2022 ) T2D FTO 1421085 T > C ( Grzeszczak et al. 2017 ) T2D FTO 1558902 T > A ( Zheng et al. 2015 ) T2D FTO 17817449 T > G ( Amine Ikhanjal et al. 2023 ) T2D FTO 9939609 T > A ( Quevedo Alves et al. 2022 ) T2D GCKR 1260326 T > C ( Petit et al. 2016 ) T2D IL6-AS1 1800795 C > G ( Vaxillaire et al. 2008 ) T2D KCNQ1 2283228 A > C ( Xu et al. 2021 ) T2D PAX4 3824004 G > A ( Gao et al. 2021 ) T2D SLC30A8 11558471 A > G ( Rees et al. 2011 ) T2D SLC30A8 13266634 C > T ( Faghih et al. 2014 ) T2D TCF7L2 12255372 G > T ( Wang et al. 2013 ) T2D TCF7L2 7903146 C > T ( Podboi et al. 2021 ) Open in a new tab DL: Dyslipidemia; T2D: Type 2 Diabetes. Figure 1. Open in a new tab Flowchart of the study design. 2.2 FFN-Weibull accelerated failure time model We used a parametric survival model, specifically the Weibull accelerated failure time (AFT) model, to predict the time to AD onset ( Weibull 1951 ). The Weibull distribution was selected based on the premise that the risk of developing AD increases over time, corresponding to a shape parameter of k > 1 . We used an approach that integrates an FFN into the Weibull AFT framework to capture the complex nonlinear relationships between the covariates and survival time ( Figure 2 ). Figure 2. Open in a new tab Architecture of the proposed FFN-based Weibull AFT model. The model consists of an input layer that receives the covariates, two hidden layers with SoftPlus activation functions to capture complex nonlinear relationships, and a single-node output layer that predicts the log of the scale parameter λ individual for each subject. The shape parameter k shared is shared across all subjects and is optimized during training. This architecture allows the model to learn intricate interactions between genetic variants and clinical features while maintaining interpretability through the Weibull AFT framework. In our model, the shape parameter k is a single shared parameter across all subjects. The scale parameter λ i for each subject i is determined by the output of the FFN, allowing λ i to vary as a function of the individual covariates x i . The relationship is defined as log ( λ i ) = FFN ( x i ; θ ) (1) Where θ represents the complete set of learnable parameters within the FFN. The network architecture consists of an input layer, two hidden layers with a SoftPlus activation function, and a single-node output layer ( Zheng et al 2015 ). This design was adopted to effectively model complex interactions while mitigating the risk of overfitting given the limited number of variables and samples. The FFN architecture is more flexible than traditional linear models, enabling the automatic learning of feature combinations and gene-gene interactions (epistasis). Furthermore, the weights associated with less-influential features tend toward zero, resulting in an implicit feature selection effect. The entire set of model parameters ( θ , k ) was simultaneously optimized using a backpropagation algorithm to maximize the likelihood function. The number of parameters in the model is 1968. 2.3 Evaluating and comparing model with baseline models Due to the practical challenges in acquiring an independent external dataset comprising both high-dimensional genomic and clinical data, we rigorously evaluated the performance and generalization capability of the proposed FFN-based Weibull model using 5-fold cross-validation. To comprehensively assess the models in the context of survival analysis, we calculated standard metrics required for time-to-event data, focusing on both discrimination and calibration. Specifically, the primary metric for assessing discriminative accuracy was the concordance index (C-index), which evaluates the model’s ability to correctly rank survival times. Furthermore, we calculated the integrated Brier score to measure the absolute calibration and accuracy of the predicted survival probabilities over time. We established a traditional linear Weibull AFT model as a baseline for comparison to demonstrate the ability of the FFN to learn nonlinear relationships compared with that of the baseline. The linear model is interpretable and computationally efficient but is constrained by the assumption of a linear association between the covariates and survival time. The evaluation metrics of both models were compared to quantitatively validate the stronger predictive performance of the FFN model in capturing complex data patterns, which a simpler linear approach cannot. 2.4 Interpretation of model predictions We interpreted the FFN model predictions and identify genetic factors that substantially contribute to the risk of AD using the SHAP technique ( Nohara et al 2022 ). The SHAP values provide a unified measure of feature importance by quantifying the contribution of each feature to the model output for individual predictions. We calculated the SHAP values for all features across the entire dataset after training the FFN model on the full dataset. The absolute mean SHAP value for each feature was computed to rank the overall importance of the feature in predicting AD onset. The feature importance of the linear Weibull model was determined using the absolute values of the estimated regression coefficients ( β ) to provide a comparative context. The importance rankings from both models were comparatively analyzed to identify the nonlinear interactions and epistatic effects captured by the FFN model but missed by the linear counterpart. Furthermore, an iterative modeling procedure was implemented to assess the impact of feature selection and validate the effectiveness of SHAP-derived rankings. We trained a series of FFN models using incrementally larger sets of top-ranked features based on the SHAP importance. Specifically, the models were trained with the top 10, 20, 30, 40, and overall features. The change in the C index was evaluated for each model to determine how predictive performance strengthened as more-important features were added. Finally, we visualized a beeswarm plot of the distribution of the SHAP values for the top 20 features. This plot illustrates not only the importance of each feature, but also the nature of its influence on the model’s predictions, providing deeper insights into how specific genetic variants affect the risk of earlier AD onset. 2.5 Software and libraries The following Python libraries were used to perform the analysis: lifelines (v0.30.0) ( Davidson-Pilon 2019 ), torchsurv(v0.1.5) ( Monod et al 2024 ), shap (v0.49.1) ( Lundberg and Lee 2017 ), pytorch (v2.8.0) ( Paszke et al 2019 ), numpy (v2.3.0) ( Harris et al 2020 ), pandas (v2.3.2) ( McKinney 2011 ), and scikit-learn (v1.7) ( Pedregosa et al 2011 ). 3 Results 3.1 5-Fold cross-validation and performance evaluation The performance of the FFN and linear regression models was evaluated using a five-fold cross-validation scheme. The results are summarized in Table 3 . The low standard deviation observed in the C-index values across the folds indicated a high degree of model stability. This stability demonstrates the ability of the models to generalize new data while mitigating overfitting. The mean C-index of the FFN model was consistently higher than that of the linear regression model, yielding an approximate 3.6% improvement in the mean test C-index (0.6472 for FFN versus 0.6247 for the linear baseline). Both models were retrained on the entire dataset following this confirmation of robust generalization to maximize the use of the available information. Regarding computational efficiency, the lightweight FFN model (comprising approximately 2,000 parameters) required about 30 minutes for complete training and evaluation on a single NVIDIA RTX 4090 GPU. Table 3. Concordance index(C-index) from 5-Fold Cross-Validation, LR: Linear Regression; FFN: Feedforward Neural Network; Std: Standard Deviation. Fold FFN Train FFN Test LR Train LR Test 0 0.7253 0.6481 0.6869 0.6116 1 0.7808 0.6514 0.6733 0.6446 2 0.7571 0.6736 0.6760 0.6562 3 0.7723 0.5933 0.6876 0.5930 4 0.7590 0.6698 0.6827 0.6181 Mean 0.7589 0.6472 0.6813 0.6247 Std 0.0212 0.0322 0.0064 0.0255 Open in a new tab 3.2 Feature importance analysis using SHAP The mean absolute SHAP values were computed for the impact of each factor on the prediction of AD onset to quantify the importance of each feature. Figure 3 illustrates the top 15 features ranked using these values. The FFN model assigned high importance scores to key genetic markers such as APOE E4, and APOE E2, which aligns with the established literature on AD risk factors ( Verghese et al 2011 ). Specifically, APOE E4 was identified as the most important feature, consistent with its association with increased AD risk. APOE E2 was the fifth most important feature, reflecting its protective effects against AD. In contrast, the linear regression model ranked APOE E4 and APOE E2 as the seventh- and ninth-most important features, respectively. Figure 3. Open in a new tab Top 15 features from the SHAP analysis. Sorted by mean absolute SHAP value. We further assessed the effect of feature selection by iteratively training an FFN model by incrementally adding features based on their SHAP importance ranking. The C index increased and the integrated Brier score decreased as more important features were included, indicating increasing predictive accuracy ( Figure 4 ). Figure 4. Open in a new tab Performance metrics by number of selected features. The five-feature model serves as the baseline (MMSE, APOE E2, APOE E3, APOE E4, Sex). Additional features were added sequentially in descending order of importance based on the mean absolute SHAP values from the FFN model. (a) Integrated Brier Score: A lower score indicates better absolute calibration of the survival curves, reflecting the model’s reliability in predicting the true probability of AD onset over time in a clinical diagnostic setting. (b) Concordance index (C-index): A higher score demonstrates the model’s improved discriminative ability to correctly rank patients according to their relative risk of AD onset. 3.3 Model interpretation Figure 5 provides a detailed visualization of the SHAP values for the top 20 features across all samples. The presence of the top-ranked feature, APOE E4 (indicated in red), correlated with negative SHAP values, signifying a higher risk of earlier AD onset. Conversely, the absence of APOE E4 (blue) corresponded to positive SHAP values, suggesting a lower risk of earlier AD onset. This pattern underscores the strong impact of APOE E4 on AD risk captured with the FFN model. Figure 5. Open in a new tab Beeswarm plot of SHAP values for the top 20 features. Each point represents a SHAP value for a feature and an individual sample. The color indicates the feature value (red for high, blue for low). The horizontal position shows the impact on the model output, with points further to the right indicating a lower risk of earlier AD onset. This visualization highlights both the importance of each feature and the nature of its influence on the model’s predictions. A similar pattern was observed for the fifth-ranked feature, APOE E2. The presence of APOE E2 (red) was associated with strongly positive SHAP values, reflecting protective effects against AD. Conversely, the absence of APOE E2 (blue) corresponded to a small absolute magnitude of negative SHAP values, representing a subtle, nonlinear relationship between APOE E2 and AD risk: the absence of APOE E2 is a minor risk factor. The effect of APOE E2 absence is substantially weaker than the strong protective impact of its presence. The traditional linear model failed to capture this complex pattern. The TBL2 (rs17145738) T and G genotypes were identified as the third- and fourth-most important features, respectively. The presence of the T genotype (red) was linked to negative SHAP values, indicating an increased risk of early AD onset. T genotype absence (blue) corresponded to positive SHAP values, suggesting a lower risk. The G genotype exhibited the opposite pattern. In a study on a Chinese population, the T allele of the rs17145738 genetic variant was associated with lower blood lipid levels, including triglycerides and cholesterol. This suggests that the genetic variant has negative health implications for those of East Asian descent ( Zeng et al 2013 ). The second-ranked feature, the Mini-Mental State Examination score, strongly and positively correlated with the risk of earlier-onset AD. Higher scores (red) were associated with positive SHAP values (lower risk), whereas lower scores (blue) corresponded to negative SHAP values (increased risk). The model identified metabolic disorder-related risk genotypes as protective factors against AD. This conflicts with the findings of previous studies ( Table 2 ) reporting these genotypes as risk factors. This discrepancy in results may be attributed to the complex interactions between multiple genetic and environmental factors that influence AD onset. Furthermore, to investigate potential nonlinear interactions between the top-ranked metabolic SNPs and APOE SNPs, we generated SHAP dependence plots. However, no distinct interaction patterns were observed, suggesting that the effects of these metabolic SNPs are largely independent of the APOE status ( Supplementary Figure 1 , available as supplementary data at Bioinformatics online for detailed plots). 4 Discussion 4.1 Model validation and methodological advantages We validated our proposed method, demonstrating that our FFN-XAI framework is a powerful tool for evaluating complex genetic architectures. The model not only substantiated existing biological knowledge but was also proved to robustly generate novel, testable hypotheses, particularly regarding the counterintuitive effects of metabolic SNPs. The reliability of the FFN-SHAP framework was validated, demonstrating the ability to identify known high-impact genetic factors in AD. The model correctly identified APOE E4 as the primary risk factor and APOE E2 as the main protective factor despite simultaneously evaluating numerous genetic covariates. These findings agree with the established epidemiological literature, confirming the capacity of the model to identify critical features from complex genetic data. This capacity to capture complex and nonlinear patterns represents a fundamental advantage of our approach over traditional linear models. Analyzing epistatic (gene-gene) interactions with a linear framework requires the explicit inclusion of all potential interaction terms. However, this approach quickly becomes computationally infeasible and suffers from combinatorial explosion as the included number of genetic factors increases. This practical barrier indicates the theoretical limitation of a single-layer perceptron, which cannot solve the nonlinear separable exclusive or (XOR) problem ( Rumelhart et al 1986 ). In contrast, the multilayer perceptron architecture of the FFN enables the inherent modeling of these high-order nonlinear relationships from the data during training without requiring them to be explicitly defined. This capability explains the higher-accuracy predictive performance of the FFN. while the limited availability of independent external genomic survival cohorts restricted cross-dataset validation in this study, the low variance observed in our 5-fold cross-validation confirms the internal stability and generalization capability of the model. 4.2 Limitations and further study The model additionally provided insights into more complex genetic relationships, particularly concerning SNPs related to metabolic disorders. Our analysis revealed that the risk genotypes for metabolic disorders did not uniformly act to increase AD risk. In some instances, these genotypes appeared to confer a protective effect. To investigate whether this counterintuitive observation was driven by epistatic interactions with major risk factors, we evaluated SHAP dependence plots for the top-ranked metabolic SNPs ( Supplementary Figure 1 , available as supplementary data at Bioinformatics online). The plots revealed no distinct interaction patterns, indicating that the protective effects of these metabolic SNPs act independently of the APOE genotypes. This lack of strong epistasis supports our hypothesis that this counterintuitive observation was due to the limitations of our dataset, which lacked information on acquired environmental factors. Ultimately, the FFN may have learned the net outcome of unobserved exogenous factors—such as active pharmacological interventions or lifestyle modifications typically prescribed to patients with metabolic risk genotypes—leading to a specific risk genotype being interpreted as protective in the context of these latent influences. These specific findings, hypothesized to stem from unobserved gene–environment interactions, underscore the critical next step. Future research should apply this validated methodology on more comprehensive datasets that integrate genetic information with acquired environmental factors, to directly investigated these complex interactions, moving from hypothesis generation to a more comprehensive etiological understanding. Supplementary Material btag213_Supplementary_Data btag213_supplementary_data.pdf (1.2MB, pdf) Contributor Information Sungwoo Goo, College of Pharmacy, Chungnam National University, Daejeon, 34134, Republic of Korea; Department of Bio-AI convergence, Chungnam National University, Daejeon, 34134, Republic of Korea. Soyoung Lee, College of Pharmacy, Chungnam National University, Daejeon, 34134, Republic of Korea; Institute of Drug Research and Development, Chungnam National University, Daejeon, 34134, Republic of Korea. Jung-Woo Chae, College of Pharmacy, Chungnam National University, Daejeon, 34134, Republic of Korea; Department of Bio-AI convergence, Chungnam National University, Daejeon, 34134, Republic of Korea; Institute of Drug Research and Development, Chungnam National University, Daejeon, 34134, Republic of Korea. Sangkeun Jung, Department of Bio-AI convergence, Chungnam National University, Daejeon, 34134, Republic of Korea; Department of Computer Science and Engineering, Chungnam National University, Daejeon, 34134, Republic of Korea. Hwi-yeol Yun, College of Pharmacy, Chungnam National University, Daejeon, 34134, Republic of Korea; Department of Bio-AI convergence, Chungnam National University, Daejeon, 34134, Republic of Korea. Author contributions Sungwoo Goo (Conceptualization [Equal], Data curation [Equal], Formal analysis [Equal], Methodology [Equal], Writing—original draft [Equal]), Soyoung Lee (Supervision [Equal]), Jung-woo CHAE (Supervision [Equal]), Sangkeun Jung (Supervision [Equal]), and Hwi-yeol Yun (Project administration [Equal], Supervision [Equal]) Supplementary material Supplementary material is available at Bioinformatics online. Conflict of interests None declared. Funding This study was supported by Chungnam National University, Institute of Information & Communications Technology Planning Evaluation (IITP) grant funded by the Korea government (MSIT) [No.RS-2022–00155857, Artificial Intelligence Convergence Innovation Human Resources Development (Chungnam National University)]; the National Research Foundation of Korea (NRF) grant funded by the Korea government [MSIT; No. RS-2023–00278597, RS-2022-NR070856 [Senior Health Convergence Research Center based on Life Cycle (Chungnam National University)]]; the Basic Science Research Program through the National Research Foundation of Korea (NRF) founded by the Ministry of Education [RS-2025–25397599]; the Korea Environmental Industry & Technology Institute (KEITI) through Core Technology Development Project for Environmental Diseases Prevention and Management funded by the Korea Ministry of Environment (MOE) [RS-2021-KE001333]; a grant from the Korea Machine Learning Ledger Orchestration for Drug Discovery Project (K-MELLODDY) funded by the Ministry of Health & Welfare and Ministry of Science and ICT, Republic of Korea [RS-2024–00460694]; the Korea Institute of Toxicology (KIT) Research Program [No. 2710008763, KK-2401–01]; the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) grant funded by the Ministry of Health & Welfare, Republic of Korea [grant number: RS-2024–00336984, RS-2025–02306055], support from the Ministry of Trade, Industry, and Energy (MOTIE), Korea, under the “Infrastructure program for industrial innovation” supervised by the Korea Institute for Advancement of Technology (KIAT) [RS-2024–00434342]; and a grant [No. 26212MFDS008] from the Ministry of Food and Drug Safety (MFDS), Republic of Korea. Code availability The source codes for the proposed FFN-Weibull model and SHAP analysis are provided at git repository ( https://github.com/swgoo/dementia_xai_sa ). The dataset and trained model weights cannot be shared due to Institutional Review Board (IRB) and genetic privacy restrictions. References Akimoto S, Goto C, Kuriki K.  Relationship between ethanol consumption and TBL2 rs17145738 on LDL-c concentration in Japanese adults: a four season 3-day weighed diet record study. BMC Nutr  2019;5:61. 10.1186/s40795-019-0315-6 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Al-Bustan SA, Al-Serri A, Alnaqeeb MA  et al.  Genetic association of LPL rs1121923 and rs258 with plasma TG and VLDL levels. Sci Rep  2019;9:5572. 10.1038/s41598-019-42021-3 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Amin USM, Parvez N, Rahman TA  et al.  CDKAL1 gene rs7756992 A/G and rs7754840 G/C polymorphisms are associated with gestational diabetes mellitus in a sample of Bangladeshi population: implication for future T2DM prophylaxis. Diabetol Metab Syndr  2022;14:18. 10.1186/s13098-021-00782-w [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Amine Ikhanjal M, Ali Elouarid M, Zouine C  et al.  FTO gene variants (rs9939609, rs8050136 and rs17817449) and type 2 diabetes mellitus risk: a meta-analysis. Gene  2023;887:147791. 10.1016/j.gene.2023.147791 [ DOI ] [ PubMed ] [ Google Scholar ] Davidson-Pilon C.  Lifelines: survival analysis in Python. JOSS  2019;4:1317. [ Google Scholar ] De Castro-Orós I, Cenarro A, Tejedor MT  et al.  Common genetic variants contribute to primary hypertriglyceridemia without differences between familial combined hyperlipidemia and isolated hypertriglyceridemia. Circ Cardiovasc Genet  2014;7:814–21. Dec. ISSN 1942-3268. 10.1161/circgenetics.114.000522 [ DOI ] [ PubMed ] [ Google Scholar ] Faghih H, Khatami S-R, Azarpira N  et al.  SLC30a8 gene polymorphism (rs13266634 C/T) and type 2 diabetes mellitus in South Iranian population. Mol Biol Rep  2014;41:2709–15. 10.1007/s11033-014-3158-x [ DOI ] [ PubMed ] [ Google Scholar ] Gao A, Gu B, Zhang J  et al.  Missense variants in PAX4 are associated with early-onset diabetes in Chinese. Diabetes Ther  2021;12:289–300. 10.1007/s13300-020-00960-5 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Grzeszczak W, Molsa M, Tłuczykont M  et al.  The age of developing diabetes and FTO polymorphisms (rs9939609, rs1421085, and rs9930506). Endokrynol Pol  2017;68:402–6. 10.5603/ep.a2017.0032 [ DOI ] [ PubMed ] [ Google Scholar ] Hamed BA, Farghaly HM, Omar A  et al.  Identifying key genetic variants in Alzheimer’s disease progression using graph convolutional networks (GCN) and biological impact analysis. J Big Data  2025;12:171. [ Google Scholar ] Harris CR, Millman KJ, van der Walt SJ  et al.  Array programming with numpy. Nature  2020;585:357–62. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Hasan EM, Abd Al Aziz RA, Sabry D  et al.  Genetic variants in nicotinamide-n-methyltransferase (NNMT) gene are related to the stage of non-alcoholic fatty liver disease diagnosed by controlled attenuation parameter (CAP)-fibroscan. J Gastrointestin Liver Dis  2018;27:265–72. 10.15403/jgld.2014.1121.273.wsh [ DOI ] [ PubMed ] [ Google Scholar ] Huang W, Sun Y, Sun J.  Combined effects of FTO rs9939609 and MC4R rs17782313 on obesity and BMI in CHINESE Han populations. Endocrine  2011;39:69–74. 10.1007/s12020-010-9413-6 [ DOI ] [ PubMed ] [ Google Scholar ] Huertas-Vazquez A, Plaisier CL, Geng R  et al.  A nonsynonymous SNP within PCDH15 is associated with lipid traits in familial combined hyperlipidemia. Hum Genet  2010;127:83–9. 10.1007/s00439-009-0749-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Karam ZM, Yari A, Najmadini A  et al.  Association of the ESR1 (rs9340799), OLR1 (rs3736234), LIPC (rs2070895), VDR (rs2228570), and CETP (rs708272) polymorphisms with risk of coronary artery disease in Iranian patients. J Clin Lab Anal  2024;38:e25026. 10.1002/jcla.25026 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Lai C-Q, Wojczynski MK, Parnell LD  et al.  Epigenome-wide association study of triglyceride postprandial responses to a high-fat dietary challenge. J Lipid Res  2016;57:2200–7. 10.1194/jlr.m069948 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. Curran Associates, Inc., 2017. http://papers.nips.cc/paper/7062-a-unified-approach-to-interpreting-model-predictions.pdf McKinney W. Pandas: a foundational python library for data analysis and statistics. Python High Perform Sci Comput  2011;14:1–9. [ Google Scholar ] Monod M, Krusche P, Cao Q  et al. Torchsurv: a lightweight package for deep survival analysis arXiv preprint arXiv : 2404.1076 12024. preprint: not peer reviewed. Nohara Y, Matsumoto K, Soejima H  et al.  Explanation of machine learning models using shapley additive explanation and application for real data in hospital. Comput Methods Programs Biomed  2022;214:106584. [ DOI ] [ PubMed ] [ Google Scholar ] Nuinoon M, Saiphak W, Nawaka N  et al.  Association of CELSR2, APOB100, ABCG5/8, LDLR, and APOE polymorphisms and their genetic risks with lipids among the Thai subjects. Saudi J Biol Sci  2023;30:103554. 10.1016/j.sjbs.2022.103554 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Paszke A, Gross S, Massa F  et al.  PyTorch: an imperative style, high-performance deep learning library. Adv Neural Inf Process Syst  2019;32. [ Google Scholar ] Pedregosa F, Varoquaux G, Gramfort A  et al.  Scikit-learn: machine learning in python. J Mach Learn Res  2011;12:2825–30. [ Google Scholar ] Petit J-M, Masson D, Guiu B  et al.  Gckr polymorphism influences liver fat content in patients with type 2 diabetes. Acta Diabetol  2016;53:237–42. 10.1007/s00592-015-0766-4 [ DOI ] [ PubMed ] [ Google Scholar ] Podboi ICR, Stephenson S, Pilic L  et al.  Dietary intake and TCF7L2 rs7903146 t allele are associated with elevated blood glucose levels in healthy individuals. Lifestyle Genom  2021;14:117–23. 10.1159/000518523 [ DOI ] [ PubMed ] [ Google Scholar ] Quevedo Alves F, Reuter CP, Neumann I  et al.  Relationship between rs9939609 FTO polymorphism with waist circumference and body fat is moderated by ponderal index at birth in youth. Am J Hum Biol  2022;34:e23575. 10.1002/ajhb.23575 [ DOI ] [ PubMed ] [ Google Scholar ] Rees SD, Hydrie MZI, O’Hare JP  et al.  Effects of 16 genetic variants on fasting glucose and type 2 diabetes in South asians: adcy5 and glis3 variants may predispose to type 2 diabetes. PLoS One  2011;6:e24710. 10.1371/journal.pone.0024710 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Rumelhart DE, Hinton GE, Williams RJ.  Learning representations by back-propagating errors. Nature  1986;323:533–6. [ Google Scholar ] Sherry ST, Ward M-H, Kholodov M  et al.  Dbsnp: the ncbi database of genetic variation. Nucleic Acids Res  2001;29:308–11. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Vaxillaire M, Veslot J, Dina C  et al.  Impact of common type 2 diabetes risk polymorphisms in the desir prospective study. Diabetes  2008;57:244–54. 10.2337/db07-0615 [ DOI ] [ PubMed ] [ Google Scholar ] Verghese PB, Castellano JM, Holtzman DM.  Apolipoprotein E in Alzheimer’s disease and other neurological disorders. Lancet Neurol  2011;10:241–52. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Wang J, Ban MR, Zou GY  et al.  Polygenic determinants of severe hypertriglyceridemia. Hum Mol Genet  2008;17:2894–9. 10.1093/hmg/ddn188 [ DOI ] [ PubMed ] [ Google Scholar ] Wang J, Zhang J, Li L  et al.  Association of rs12255372 in thetcf7l2 gene with type 2 diabetes mellitus: a meta-analysis. Braz J Med Biol Res  2013;46:382–93. 10.1590/1414-431x20132677. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Weibull W.  A statistical distribution function of wide applicability. Journal of Applied Mechanics  1951;18:293–7. [ Google Scholar ] Xu C, Bai R, Zhang D  et al.  Effects of APOA5 -1131t>c (rs662799) on fasting plasma lipids and risk of metabolic syndrome: evidence from a case-control study in China and a meta-analysis. PLoS One  2013;8:e56216. 10.1371/journal.pone.0056216 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Xu J, Zhang W, Song W  et al.  Relationship between KCNQ1 polymorphism and type 2 diabetes risk in northwestern China. Pharmgenomics Pers Med  2021;14:1731–51. 10.2147/pgpm.s340813 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Yanasegaran K, Ng JYE, Chua EW  et al.  Single nucleotide polymorphisms (snps) that are associated with obesity and type 2 diabetes among Asians: a systematic review and meta-analysis. Sci Rep  2024;14:20062. 10.1038/s41598-024-70674-2. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Yuan S, Liu Q, Huang X  et al.  Development of an individualized dementia risk prediction model using deep learning survival analysis incorporating genetic and environmental factors. Alzheimers Res Ther  2024;16:278. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Zeng X-N, Yin R-X, Huang P  et al.  Association of the MLXIPL/TBL2 rs17145738 SNP and serum lipid levels in the Guangxi mulao and Han populations. Lipids Health Dis  2013;12:156. 10.1186/1476-511x-12-156. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Zeng Y, He H, Zhang L  et al.  GWA-based pleiotropic analysis identified potential SNPS and genes related to type 2 diabetes and obesity. J Hum Genet  2021;66:297–306. 10.1038/s10038-020-00843-4. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Zheng H, Yang Z, Liu W  et al. Improving deep neural networks using SoftPlus units. In: 2015 International joint conference on neural networks (IJCNN) . IEEE: Piscataway, NJ, USA 2015, p. 1–4. Zhu X-J, Lin Y-J, Chen W  et al.  Physiological study on association between nicotinamide n-methyltransferase gene polymorphisms and hyperlipidemia. Biomed Res Int  2016;2016:7521942–8. 10.1155/2016/7521942 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials btag213_Supplementary_Data btag213_supplementary_data.pdf (1.2MB, pdf) Articles from Bioinformatics are provided here courtesy of Oxford University Press ACTIONS View on publisher site PDF (2.5 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Robust transcriptomic signatures of Alzheimer's disease progression: validated explainable AI approach.#213872
• Imaging-genetics-based dementia risk prediction using deep survival neural networks in the Rotterdam Study#1052961
• A Deep-Learning Based Biomarker of Systemic Cellular Senescence Burden to Predict Mortality and Health Outcomes.#1107664
• Genotoxic and metabolic stress drive divergent senescence programs in human microglia#1169723
• Generative approaches to kinetic parameter inference in metabolic networks via latent space exploration.#297357