Interpretable neural network for risk stratification and drug target discovery based on PBMC transcriptomes - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy 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 iScience . 2026 Mar 30;29(4):115538. doi: 10.1016/j.isci.2026.115538 Search in PMC Search in PubMed View in NLM Catalog Add to search Interpretable neural network for risk stratification and drug target discovery based on PBMC transcriptomes Ziqing Yu Ziqing Yu 1 Department of Cardiology, Zhongshan Hospital, Fudan University, Shanghai 200032, China Find articles by Ziqing Yu 1, 7 , Liwei Liu Liwei Liu 2 Shanghai Institute of Cardiovascular Diseases, Zhongshan Hospital, Fudan University, Shanghai 200032, China Find articles by Liwei Liu 2, 7 , Yugang Zou Yugang Zou 3 The Second People’s Hospital of Liao Cheng, Linqing, Shandong Province 252699, China Find articles by Yugang Zou 3 , Yue Liu Yue Liu 4 Department of Obstetrics and Gynecology, Obstetrics and Gynecology Hospital, Fudan University, Shanghai 200011, China Find articles by Yue Liu 4 , Liming Chen Liming Chen 2 Shanghai Institute of Cardiovascular Diseases, Zhongshan Hospital, Fudan University, Shanghai 200032, China Find articles by Liming Chen 2 , Lulu Peng Lulu Peng 5 Reproductive Medicine Centre, Zhongshan Hospital, Fudan University, Shanghai 200032, China Find articles by Lulu Peng 5 , Wei Chen Wei Chen 5 Reproductive Medicine Centre, Zhongshan Hospital, Fudan University, Shanghai 200032, China Find articles by Wei Chen 5 , Jing Wang Jing Wang 5 Reproductive Medicine Centre, Zhongshan Hospital, Fudan University, Shanghai 200032, China Find articles by Jing Wang 5 , Kun Yin Kun Yin 6 School of Global Health, Chinese Center for Tropical Diseases Research, Shanghai Jiao Tong University School of Medicine, Shanghai 200025, China Find articles by Kun Yin 6 , Zhiwei Zhang Zhiwei Zhang 2 Shanghai Institute of Cardiovascular Diseases, Zhongshan Hospital, Fudan University, Shanghai 200032, China 5 Reproductive Medicine Centre, Zhongshan Hospital, Fudan University, Shanghai 200032, China Find articles by Zhiwei Zhang 2, 5, ∗ , Suling Ding Suling Ding 2 Shanghai Institute of Cardiovascular Diseases, Zhongshan Hospital, Fudan University, Shanghai 200032, China Find articles by Suling Ding 2, ∗∗ , Qinqin Hu Qinqin Hu 2 Shanghai Institute of Cardiovascular Diseases, Zhongshan Hospital, Fudan University, Shanghai 200032, China 6 School of Global Health, Chinese Center for Tropical Diseases Research, Shanghai Jiao Tong University School of Medicine, Shanghai 200025, China Find articles by Qinqin Hu 2, 6, 8, ∗∗∗ Author information Article notes Copyright and License information 1 Department of Cardiology, Zhongshan Hospital, Fudan University, Shanghai 200032, China 2 Shanghai Institute of Cardiovascular Diseases, Zhongshan Hospital, Fudan University, Shanghai 200032, China 3 The Second People’s Hospital of Liao Cheng, Linqing, Shandong Province 252699, China 4 Department of Obstetrics and Gynecology, Obstetrics and Gynecology Hospital, Fudan University, Shanghai 200011, China 5 Reproductive Medicine Centre, Zhongshan Hospital, Fudan University, Shanghai 200032, China 6 School of Global Health, Chinese Center for Tropical Diseases Research, Shanghai Jiao Tong University School of Medicine, Shanghai 200025, China ∗ Corresponding author [email protected] ∗∗ Corresponding author [email protected] ∗∗∗ Corresponding author [email protected] 7 These authors contributed equally 8 Lead contact Received 2025 Aug 5; Revised 2026 Jan 19; Accepted 2026 Mar 26; Collection date 2026 Apr 17. © 2026 Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). PMC Copyright notice PMCID: PMC13091452 PMID: 42006310 Summary Peripheral blood mononuclear cells (PBMCs) play a central role in immune surveillance and disease pathogenesis and provide a minimally invasive source for molecular screening. Here, we developed ME-NET, an intrinsically interpretable deep learning framework for multi-disease screening and risk stratification from PBMC transcriptomic profiles. Using harmonized datasets from heterogeneous clinical backgrounds, ME-NET showed consistent performance and outperformed standard machine-learning baselines. By incorporating curated pathway priors in a pathway-informed layer, ME-NET connects predictions to pathway-level signals and contributing genes, enabling biologically grounded interpretation of disease-associated immune programs. Finally, integrating ME-NET-prioritized targets with drug-gene interaction resources identified candidate compounds for potential therapeutic repurposing. Collectively, these results support the use of interpretable deep learning on blood-derived transcriptomes for scalable multi-disease screening and for prioritizing biomarkers and drug targets. Subject areas: Health sciences, Medicine, Immunology Graphical abstract Open in a new tab Highlights • We develop ME-NET, an interpretable deep learning model for health risk prediction • It leverages PBMC transcriptomes from 2,600+ individuals across nine major diseases • The model identifies biomarkers and pathways with superior accuracy over baselines • ME-NET reveals 400+ candidate drugs for repurposing, enabling translational medicine Health sciences; Medicine; Immunology Introduction Peripheral blood mononuclear cells (PBMCs), including lymphocytes and monocytes, are critical components of immune responses to infectious and non-infectious diseases. 1 Conventional assays such as complete blood counts and flow cytometry are widely employed to quantify PBMC populations, and provide diagnostic information across diseases. 2 However, these assays provide limited molecular resolution for disease-specific risk assessment. Although panels of blood biomarkers measured by enzyme-linked immunosorbent assay (ELISA) and related assays can improve diagnostic specificity, a scalable and generalizable framework for multi-disease screening in heterogeneous populations remains needed. 3 , 4 Machine-learning approaches have been widely applied in biomedical research, including pathology imaging and proteomics. 5 However, many conventional models rely on manual feature engineering and may generalize poorly in high-dimensional omics settings. A growing body of work has explored transcriptome-based disease stratification using machine-learning models. Prior work has shown that conventional machine-learning models can stratify patients within a single disease using blood/PBMC transcriptomic profiles. 6 In parallel, deep learning can learn hierarchical, non-linear representations directly from omics measurements, and has enabled multi-endpoint prediction of future health risks from blood-omics data. 5 Despite these advances, limited transparency remains a major barrier for clinical translation, where understanding the decision-making process is crucial. 7 To address these challenges, we developed ME-NET, an interpretable deep learning model for multi-disease screening and risk stratification using PBMC-derived transcriptomic data. ME-NET adopts a pathway-informed architecture in which gene-level inputs are first aggregated into curated pathway nodes and then processed by downstream hidden layers. This design enables direct gene- and pathway-level attribution of predictions, providing biologically grounded insight beyond post hoc feature attribution alone. Using this framework, we estimated disease-related risk scores in clinically heterogeneous cohorts and prioritized disease-associated pathways and candidate therapeutic targets for downstream analysis. Results Overview of datasets To evaluate whether a single PBMC-transcriptome model can generalize across distinct clinical domains, we assembled datasets spanning multiple disorder categories (cardiovascular/respiratory, autoimmune, and neuropsychiatric/neurodegenerative), constrained by the availability of publicly accessible PBMC (or PBMC-enriched blood) transcriptomic data with compatible case/control annotations. A total of 2,699 participants with PBMC-derived transcriptomic data were included in the study, spanning nine distinct diseases: Alzheimer’s disease (AD), 8 amyotrophic lateral sclerosis (ALS), bipolar disorder (BP), 9 cardiovascular disease (CD), 10 , 11 chronic obstructive pulmonary disease (COPD), 12 , 13 multiple sclerosis (MS), 14 Parkinson’s disease (PD), 15 rheumatoid arthritis (RA), 16 and schizophrenia (SCZ) 17 ( Figure 1 A). To mitigate cross-platform batch effects, all datasets underwent normalization and YuGene transformation, after which 1,799 genes shared across all datasets were retained as the cross-platform intersection for downstream modeling. Figure 1. Open in a new tab Integration of blood-derived datasets across multiple diseases (A) t-SNE visualization of pooled case and control samples. t-SNE was applied to the first 10 principal components using Euclidean distance in PCA space (perplexity = 20; learning rate = 224.6; max iterations = 3,000; theta = 0.3; output dimensions = 2; random seed = 1). (B) Sample size across disease categories: AD ( n = 293), ALS ( n = 52), BP ( n = 48), CD ( n = 343), COPD ( n = 215), MS ( n = 104), PD ( n = 90), RA ( n = 441), and SCZ ( n = 153). To strengthen disease-specific discrimination beyond generic “unhealthy” immune activation, we labeled samples from the target disease as positive, while controls together with samples from other diseases were used as a heterogeneous negative background (“hard negatives”) to enhance specificity in clinically mixed settings ( Figure 1 B). Performance of the ME-NET framework ME-NET demonstrated consistently high classification performance across all nine diseases. Receiver operating characteristic area under the curve (ROC-AUC) values exceeded 0.80 for all conditions, outperforming support vector machine (SVM), K-nearest neighbor (KNN), random forest (RF), and extreme gradient boosting (XGBoost) models ( Figure 2 A). Figure 2. Open in a new tab Model performance evaluation (A) Comparison of receiver operating characteristic (ROC) curves and area under the curve (AUC) values for ME-NET and baseline machine learning models (SVM, KNN, RF, and XGBoost) across nine diseases. (B) External evaluation on an independent acute coronary syndrome cohort ( n = 71; STEMI vs. non-STEMI). Models were compared with ME-NET (DeLong’s test, ∗ p < 0.05; ∗∗ p < 0.01). External validation was performed by applying the CD classifier of ME-NET to an independent acute coronary syndrome cohort ( n = 71), including ST-elevation myocardial infarction (STEMI) and non-STEMI patients, to assess cross-cohort generalization of the learned CD risk signal. Despite clinical heterogeneity in this real-world cohort, ME-NET achieved an AUC of 0.72), whereas baseline models achieved AUC values below 0.60 ( Figure 2 B). Biological interpretability The interpretable layer of ME-NET facilitated the identification of disease-relevant genes and pathways. Comparison with DEGs via Venn diagrams revealed a distinct subset of ME-NET-prioritized genes not captured by standard differential analysis ( Figure 3 A). Compared with DEGs identified solely by statistical significance, ME-NET prioritized a more disease-specific and predictive gene set, characterized by stronger disease-background separability and higher model-derived importance, consistent with greater contributions to classification ( Figure 3 B). Interestingly, ME-NET identified more concentrated set of pathways than DEG ( Figure 4 A). Figure 3. Open in a new tab Comparison between differentially expressed genes and ME-NET-prioritized genes (A) Venn diagram illustrating the up-/down-regulated overlap between DEGs and ME-NET-prioritized genes. (B) Violin plots showing expression distributions of representative top-ranked genes from the DEG list and the ME-NET list across the corresponding sample groups. Figure 4. Open in a new tab Pathway and GO enrichment (A) Pathway enrichment results for DEG- and ME-NET-derived gene sets. Normalized enrichment score (NES) was computed using GSEA with MSigDB Hallmark gene sets. (B–D) Sankey plots summarizing enriched GO terms across biological process (BP), cellular component (CC), and molecular function (MF) categories for ME-NET-prioritized genes and DEGs. To further clarify the differences between genes identified by ME-NET and DEGs, gene ontology (GO) analysis was applied, including biological function, cellular compartment, and molecular function ( Figures 4 B–4D). ME-NET-prioritized genes were enriched for nuclear organization, mitochondrial respiratory chain components (cytochrome complexes), and RNA-binding/translation-related functions. In contrast, DEGs were enriched for leukocyte chemotaxis, dynein complex, and ubiquitin-like protein ligase binding. Drug discovery by ME-NET Drug discovery was performed by cross-referencing ME-NET-derived targets with DrugBank. Over 400 drug candidates were identified, including FDA-approved and investigational agents ( Figure 5 ). Notable examples include Fingolimod (Gilenya)—clinically used for MS—and Forskolin, a known activator of adenylate cyclase with cardioprotective properties. These findings demonstrate the utility of ME-NET not only for diagnosis but also for therapeutic discovery. Figure 5. Open in a new tab Drug prediction and repurposing analysis Drug-target mapping workflow based on ME-NET-prioritized genes. Nodes represent diseases, prioritized target genes, and nominated compounds. Compounds are categorized by clinical status: approved (red), investigational (orange), and not previously linked to the analyzed diseases (blue). Discussion Despite the rapid growth of artificial intelligence (AI) applications in medicine, the complexity and heterogeneity of disease remain significant obstacles to deploying data-driven decision-support models. 18 In this study, we present ME-NET, an interpretable deep learning framework for disease risk assessment across nine distinct conditions, using blood-derived transcriptomic data ( Figure 6 ). By integrating prediction with interpretation, ME-NET demonstrates strong potential for large-scale disease screening, early biomarker discovery, and drug repurposing. Figure 6. Open in a new tab Schematic overview of the ME-NET architecture The model takes normalized PBMC transcriptomic profiles (1,799 genes) as input, where each node corresponds to a gene. Inputs are first mapped to a biologically informed, partially connected pathway layer comprising functional modules curated from KEGG and Reactome. These pathway nodes represent prior knowledge and can be updated during training as model parameters are optimized. The pathway layer is followed by fully connected hidden layers that integrate pathway-level signals and learn higher-order nonlinear representations. Finally, the output layer produces disease risk stratification scores (predicted probabilities) for the target conditions. In recent years, deep learning has revolutionized biomedical research, particularly in areas such as medical imaging, proteomics, and genomics. 19 , 20 However, its application to transcriptomic data—especially in a clinically relevant, interpretable manner—remains underdeveloped. 21 , 22 Traditional machine learning models typically rely on manually selected features and often fall short in handling the high dimensionality of omics data or delivering biological insights. 23 , 24 ME-NET addresses these limitations by integrating gene expression with pathway-level biological knowledge in a fully automated deep learning architecture. Unlike black-box models, ME-NET incorporates an interpretability module, allowing clinicians and researchers to trace model predictions back to specific genes and pathways. A growing body of research indicates that the immune system plays a sentinel role in detecting and responding to disease. Immune cells circulating in peripheral blood continuously adapt to internal physiological states, making them valuable, non-invasive reporters of disease status. 25 , 26 However, most blood immune transcriptome-based predictive models have been developed for single diseases (or narrowly defined differential-diagnosis tasks), rather than as scalable multi-disease screening frameworks. 7 In contrast, multi-endpoint future risk stratification has been more extensively demonstrated in large-scale blood proteomics. 27 Building on this paradigm, ME-NET aims to operationalize PBMC transcriptomics for multi-disease risk stratification within a unified architecture, while providing model interpretability through an explicitly biologically structured pathway layer that enables direct gene- and pathway-level attribution (rather than relying only on post hoc feature-importance analyses, such as SHAP, commonly used in prior work). To further improve the robustness of our framework, we used data from healthy controls and patients with other diseases for comparison and training. This mixed-negative design reflects the reality that PBMC transcriptomes often capture shared host-response programs across disorders, and it prevents the model from relying on a trivial “unhealthy vs. healthy” separator. 28 Instead, non-target diseases serve as “hard negatives,” encouraging ME-NET to learn comparatively disease-specific signals in a clinically heterogeneous background. The results indicate that our model achieves more robust performance than traditional prediction models, as reflected by consistently higher AUCs across diseases. Notably, among the baseline models, XGBoost was generally the strongest comparator and, for some diseases, approached ME-NET performance, which is consistent with the strong empirical performance of gradient-boosted trees on high-dimensional transcriptomic data. 28 , 29 In contrast, RF showed weaker performance across multiple diseases and exhibited more pronounced deviations for AD and SCZ, which may reflect the interaction between RF’s sensitivity to correlated high-dimensional features and cohort heterogeneity and/or subtler peripheral-blood signals in these conditions. Moreover, to further verify the accuracy of our model, we also collected transcriptomic data from independent acute coronary syndrome cohort for testing. Compared to traditional methods, the CD risk scores generated by ME-NET showed meaningful separation between STEMI and non-STEMI patients (AUC = 0.72), suggesting that the CD-related transcriptomic signal learned from public datasets can generalize to an independent clinical cohort. In addition, we compared ME-NET-prioritized genes with DEGs derived from an independent STEMI cohort (not used for training) and observed limited but biologically plausible overlap (e.g., RGS19), indicating that ME-NET-derived importance can capture reproducible signals in an external cohort. Overall, these results support the generalization potential of ME-NET in clinically heterogeneous backgrounds and highlight the value of learning high-dimensional disease-associated patterns beyond non-specific immune activation. Interpretable and explainable neural networks for clinical and biological research have advanced rapidly, and existing approaches can be broadly grouped into intrinsic interpretability (architecture-level transparency) and post hoc explainability (explanations applied after model training). 30 On the intrinsic side, one prominent line of work embeds biological structure into network topology so that intermediate units correspond to interpretable entities. Examples include knowledge-primed neural networks (KPNNs) that map model nodes/edges onto curated biological networks, 31 and “visible” neural networks such as DCell that explicitly align neurons with hierarchical biological knowledge. 32 More recent variants further emphasize parsimony and hierarchical interpretability. 33 Beyond biology-structured models, intrinsic interpretability is also pursued through additive and modular architectures, such as neural additive models (NAMs), which preserve a GAM-like decomposition while retaining neural-network flexibility. 34 Complementarily, post hoc explainability methods quantify feature attributions for trained models, including Integrated Gradients, DeepLIFT, and SHAP, which are widely used to prioritize influential inputs and to audit model behavior. 35 In this study, ME-NET follows the intrinsic interpretability paradigm by introducing a pathway-informed layer (curated from KEGG/Reactome) that enables direct gene- and pathway-level attribution in a heterogeneous “hard-negative” PBMC screening setting, while post hoc attributions are used only for contextualization rather than as the primary basis for mechanistic interpretation. Our interpretable deep learning model provides not only potential of clinical application, but also crucial insights for further research. 22 , 23 Besides enhancing its predictive capability, the interpretive module of ME-NET deciphers the contribution of each node. For example, DEG analysis highlighted both KRAS signaling (e.g., LAT2) and DNA repair-related genes (e.g., POLR2A, POLB, and POLR2G) in RA. In contrast, ME-NET primarily emphasized KRAS-related signals, suggesting that although DNA repair-related genes are differentially expressed, they contribute less to discriminating RA from the heterogeneous background under our classification setting. 36 , 37 , 38 Moreover, to further clarify the differences between genes-derived DEG and ME-NET, we analyzed the cellular distribution, biological progression and molecular function of the both. More DEGs were clustered in the RNA splicing, epigenetic modification, and kinase system, while ME-NET genes were grouped in ribosomal subunits, endoplasmic reticulum, and plasma membrane. 25 , 26 , 39 , 40 It was considered that immune stem cells differentiate into specific lineages to adapt disease status, by epigenetic regulation and RNA splicing. 26 These enrichment patterns indicate that epigenetic and RNA-processing programs are broadly represented during immune-cell differentiation, whereas the ME-NET-prioritized gene set is enriched for protein synthesis/secretory processes and plasma membrane signaling, including pathways related to inflammation and cell-cell communication. 41 Abnormal inflammatory responses occur in patients with various diseases, including cardiovascular, autoimmune, and neurodegenerative diseases, targeting this process could be a practical strategy to alleviate illness. 42 , 43 , 44 Based on our model and database of DrugBank, more than 400 compounds have been identified, including approved or investigational agonists/antagonists. Notably, Fingolimod (Gilenya), used for MS by sequestering lymphocytes in lymph nodes, emerged as a notable candidate—mirroring immune modulation identified by ME-NET. 45 , 46 Similarly, Forskolin, an adenylate cyclase activator with cardioprotective effects, was highlighted in cardiovascular contexts for its role in reducing inflammation and tissue injury. 47 , 48 Moreover, connectivity map (CLUE platform) and drug-gene interaction database (DGIdb) are also included for target validation ( Tables S1 and S2 ). By leveraging immune cell transcriptomics and interpretable deep learning, ME-NET offers a robust, generalizable framework for disease risk stratification and mechanistic insight. It bridges a crucial gap between predictive modeling and biological understanding, uncovering features relevant for diagnosis and therapeutic exploration. Despite certain limitations, ME-NET represents a significant advancement toward the clinical application of AI for scalable, cost-effective disease screening and health risk assessment. Limitations of the study While ME-NET demonstrates strong predictive performance and interpretability, several limitations warrant consideration and offer avenues for future improvement. First, the model’s performance is highly dependent on the protocol quality and integrity of immune cell isolation and transcriptomic sequencing. Although ME-NET showed robustness in an independent real-world evaluation with cardiovascular cohort, its cross-cohort generalization should be assessed in multicenter studies, particularly in comorbid patient. Second, like other biologically informed models, ME-NET relies on existing biological pathway databases for network construction and interpretation. As our understanding of gene functions and interactions evolves, the integration of new or updated pathway databases and novel gene-gene connections could further enhance model performance and expand its biological relevance. Finally, while ME-NET has identified several promising biomarkers, pathways and targets/drugs, additional data would be essential, including clinical database, molecular docking, and functional validation in organoids and animals. Resource availability Lead contact Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Qinqin Hu ( [email protected] ). Materials availability This study did not generate new unique reagents. Data and code availability • Raw data of the study are deposited at https://ngdc.cncb.ac.cn/gsa-human GSA-Human: HRA007646. • ME-NET is available at GitHub: ZSBiolab/ME-NET https://github.com/ZSBiolab/ME-NET . • All data presented in this study are available upon request from the lead contact upon reasonable request. Acknowledgments This study was supported by grants from the National Key Research and Development Program of China (2024YFC3406504), the Basic Science Center Project (T2288101), the National Natural Science Foundation of China, China (82000270, 82300563, and 82370266), and Shanghai Municipal Health Commission, China (202240292). We would like to thank Liu Yang for inspiring this study. Author contributions Q.H., S.D., and Z.Z. were involved in the funding acquirement, study design, and manuscript preparation; Y.Z., L.Y., L.C., L.P., W.C., and J.W. participated in sample collection; Z.Z., L.L., and Z.Y. participated in experiments and model construction; L.L., Z.Y., K.Y., and S.D. contributed to data analysis. All the authors have read and approved the final manuscript. Declaration of interests The authors declare no conflict of interests. Declaration of generative AI and AI-assisted technologies in the writing process During the preparation of this work, the authors used ChatGPT5.2 in order to polish the paper. After using this service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication. STAR★Methods Key resources table REAGENT or RESOURCE SOURCE IDENTIFIER Biological samples Human blood samples Zhongshan Hospital N/A Critical commercial assays Lymphocyte separation medium MP Biomedicals Cat# 1692249 Deposited data Gene expression data of PBMC This paper HRA007646 Code for model This paper https://github.com/ZSBiolab/ME-NET Software and algorithms Bowtie (v2.0.6) Langmead B et al. https://bowtie-bio.sourceforge.net/ TopHat (v2.0.9) Trapnell C et al. https://ccb.jhu.edu/software/tophat/ PyTorch (v1.11.0) PyTorch Foudation https://pytorch.org/ Open in a new tab Experimental model and study participant details Patient enrollment and sample collection The protocol used in this investigation was reviewed and approved by the ethics committee (B2023-260R), in accordance with the principles of the Declaration of Helsinki. Patients and control subjects were enrolled with written informed consent from Zhongshan Hospital, Fudan University. Patients with ST-elevation myocardial infarction (STEMI) underwent coronary angiography and percutaneous coronary intervention of the infarct-related artery, with blood collection on the day 1 post-infarction. Study participation did not influence clinical management, including pharmacological treatment and procedures. Control subjects were excluded if they exhibited coronary artery stenosis exceeding 50% (as assessed by angiography), had undergone previous percutaneous coronary intervention or coronary artery bypass grafting, or tested positive in non-invasive cardiac assessments. Exclusion criteria included the presence of conditions expected to substantially confound transcriptomic profiling or clinical classification (e.g., active systemic infection, known autoimmune disease, malignancy under active treatment, or ongoing immunosuppressive therapy). The final external validation cohort comprised a total of n = 71 participants. Sex-related bias was minimized where possible by performing sex-consistency quality control with massiR in datasets with available sex annotation. However, the potential influence of sex, age and race on the study outcomes cannot be excluded. Peripheral blood was collected into EDTA (Ethylene Diamine Tetraacetic Acid) anticoagulant tubes, and PBMCs were isolated using lymphocyte separation medium (MP Biomedicals, USA) according to the manufacturer’s instructions. Method details RNA extraction and sequencing Total RNA was isolated from PBMCs and quantified by UV spectrophotometry (NanoDrop, UK). RNA integrity was assessed, and samples with an RNA integrity number (RIN) of at least 8 were retained and stored at −80°C until processing. Sequencing libraries were prepared from qualified RNA, followed by adaptor ligation and PCR amplification. Libraries were sequenced on an Illumina HiSeq 2500 platform. Raw reads were aligned using the Bowtie (v2.0.6)-TopHat (v2.0.9) pipeline for downstream RNA-seq analysis. Raw data of the study (HRA007646) are deposited at https://ngdc.cncb.ac.cn/gsa-human/ . Data acquisition and gene expression analysis Gene expression data were sourced from publicly available repositories Gene Expression Omnibus (GEO) ( https://www.ncbi.nlm.nih.gov/geo/ ) and Array Express ( https://www.ebi.ac.uk/arrayexpress/ ), processed using a unified pipeline as described previously. 28 Briefly, YuGene transformation was applied independently within each dataset after background correction and normalization to improve cross-study comparability. For datasets with recorded sex, we performed a sex-consistency quality-control step using massiR to identify and exclude samples with discordant predicted versus recorded sex. 49 To enable cross-platform modeling, we defined the “shared gene set” as genes with consistent identifiers and available expression measurements across all included datasets after gene-identifier harmonization and preprocessing. The final shared set comprised 1,799 genes and was used as the common feature space for all models. The differentially expressed genes (DEGs) were analyzed, using R-package limma, for training datasets and external PBMC dataset (GEO: GSE103182 ). DEGs were defined as FDR < 0.01 and |log2FC| ≥ 0.1. Functional analysis was performed using clusterProfiler and Gene set enrichment analysis (GSEA), and databases, including Connectivity Map (CLUE platform), Drug-gene interaction database (DGIdb), DrugBank, Gene Ontology (GO), and hallmark gene set of Molecular signatures database (MSigDB). 50 , 51 Model architecture The architecture of ME-NET consists of an input layer, an interpretable pathway-informed layer and multiple hidden layers. 23 The patient samples with normalized genes were transferred into the input layer, with the same dimension of features. The pathway information is incorporated in the interpretable layer as a prior knowledge, while the importance scores of each node were extracted from this layer for the further analysis. The connected hidden layers express the higher-level information of genes and pathways, which facilitates the further analysis, by indicating active/inactive status of nodes ( Figure 6 ). ME-NET is available on GitHub ( https://github.com/ZSBiolab/ME-NET ). Architecture for interpretability In a conventional fully connected neural network, the first hidden layer is densely connected to all input genes and the resulting latent units are not directly interpretable. In ME-NET, the first embedding layer is replaced by an interpretable pathway layer whose nodes correspond to curated biological pathways and whose gene-to-node edges are constrained by the incidence matrix. Consequently, the learned gene-pathway weights and derived node importance scores provide quantitative attribution at the gene and pathway levels, enabling mechanistic interpretation of model predictions. Subsequent hidden layers are used to learn higher-order non-linear representations for risk stratification. Model training and evaluation ME-NET was trained on 80% of the dataset, with the remaining 20% reserved for internal validation. Model performance was benchmarked against four widely used machine learning algorithms: Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Random Forest (RF), and Extreme gradient boosting (XGBoost). 52 , 53 All models were tuned using stratified 3-fold cross-validation to ensure fair comparison. ME-NET was implemented in PyTorch (v1.11.0) and trained on an NVIDIA RTX 4090 GPU (24 GB). Model optimization employed an adaptive learning rate starting at 0.03, which decreased dynamically upon validation loss plateauing. Regularization was achieved through L2-penalties ranging from 0.1 to 0.001 and dropout techniques to prevent overfitting. Finally, models were trained and validated using an external case-control cohort. Model interpretation To quantify disease-relevant genes and pathway nodes from ME-NET, we extracted a model-derived weight matrix W with the same dimension as the input expression matrix, where each entry represents the gene-level contribution score assigned by the trained model for a given sample. For each disease task, we compared the distribution of gene weights between the target-disease group and the heterogeneous non-target background (healthy controls plus samples from other diseases). Statistical significance was assessed using the Wilcoxon rank-sum test for each gene, followed by Benjamini-Hochberg false discovery rate (FDR) correction (q value) across all genes. Genes with FDR q < 0.05 were defined as ME-NET-prioritized genes for that disease. DEG and ME-NET-prioritized genes were used as targets for new compounds. Quantification and statistical analysis For internal evaluation, stratified 3-fold cross-validation was used to preserve class proportions across splits. Model performance was primarily assessed using the receiver operating characteristic area under the curve (ROC-AUC), which is reported as the mean with 95% confidence intervals (CIs) across cross-validation folds. For external validation, classifier performance was further evaluated in an independent cohort, and ROC-AUC with 95% CI was calculated. Statistical significance of AUC differences between ME-NET and baseline models was assessed using DeLong’s test for correlated ROC curves. Statistical analyses were performed using R (4.2.1) and Python (3.13.4). Published: March 30, 2026 Footnotes Supplemental information can be found online at https://doi.org/10.1016/j.isci.2026.115538 . Contributor Information Zhiwei Zhang, Email: [email protected]. Suling Ding, Email: [email protected]. Qinqin Hu, Email: [email protected]. Supplemental information Table S1. Connectivity map analysis of candidate compounds associated with ME-NET-prioritized targets mmc1.csv (56.3KB, csv) Table S2. DGIdb-based drug-gene interaction annotation of ME-NET-prioritized targets mmc2.csv (507.7KB, csv) References 1. Shi C., Pamer E.G. Monocyte recruitment during infection and inflammation. Nat. Rev. Immunol. 2011;11:762–774. doi: 10.1038/nri3070. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Amani F., Amanzadeh M., Hamedan M., Amani P. Diagnostic accuracy of deep learning in prediction of osteoporosis: a systematic review and meta-analysis. BMC Musculoskelet. Disord. 2024;25:991. doi: 10.1186/s12891-024-08120-7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Lackner K.J. Cardiac troponins - a paradigm for diagnostic biomarker identification and development. Clin. Chem. Lab. Med. 2023;61:795–800. doi: 10.1515/cclm-2022-1112. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Yazdani A.N., Pletsch M., Chorbajian A., Zitser D., Rai V., Agrawal D.K. Biomarkers to monitor the prognosis, disease severity, and treatment efficacy in coronary artery disease. Expert Rev. Cardiovasc Ther. 2023;21:675–692. doi: 10.1080/14779072.2023.2264779. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Carrasco-Zanini J., Pietzner M., Davitte J., Surendran P., Croteau-Chonka D.C., Robins C., Torralbo A., Tomlinson C., Grünschläger F., Fitzpatrick N., et al. Proteomic signatures improve risk prediction for common and rare diseases. Nat. Med. 2024;30:2489–2498. doi: 10.1038/s41591-024-03142-z. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Acquaviva M., Menon R., Di Dario M., Dalla Costa G., Romeo M., Sangalli F., Colombo B., Moiola L., Martinelli V., Comi G., Farina C. Inferring Multiple Sclerosis Stages from the Blood Transcriptome via Machine Learning. Cell Rep. Med. 2020;1 doi: 10.1016/j.xcrm.2020.100053. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Meier S., Henkens M., Heymans S., Robinson E.L. Unlocking the Value of White Blood Cells for Heart Failure Diagnosis. J. Cardiovasc. Transl. Res. 2021;14:53–62. doi: 10.1007/s12265-020-10007-6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Sood S., Gallagher I.J., Lunnon K., Rullman E., Keohane A., Crossland H., Phillips B.E., Cederholm T., Jensen T., van Loon L.J.C., et al. A novel multi-tissue RNA diagnostic of healthy ageing relates to cognitive health status. Genome Biol. 2015;16:185. doi: 10.1186/s13059-015-0750-x. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Clelland C.L., Read L.L., Panek L.J., Nadrich R.H., Bancroft C., Clelland J.D. Utilization of Never-Medicated Bipolar Disorder Patients towards Development and Validation of a Peripheral Biomarker Profile. PLoS One. 2013;8 doi: 10.1371/journal.pone.0069082. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Sinnaeve P.R., Donahue M.P., Grass P., Seo D., Vonderscher J., Chibout S.-D., Kraus W.E., Sketch M., Jr., Nelson C., Ginsburg G.S., et al. Gene Expression Patterns in Peripheral Blood Correlate with the Extent of Coronary Artery Disease. PLoS One. 2009;4 doi: 10.1371/journal.pone.0007037. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Maciejak A., Kiliszek M., Michalak M., Tulacz D., Opolski G., Matlak K., Dobrzycki S., Segiet A., Gora M., Burzynska B. Gene expression profiling reveals potential prognostic biomarkers associated with the progression of heart failure. Genome Med. 2015;7:26. doi: 10.1186/s13073-015-0149-z. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Singh D., Fox S.M., Tal-Singer R., Bates S., Riley J.H., Celli B., Investigators E. Altered Gene Expression in Blood and Sputum in COPD Frequent Exacerbators in the ECLIPSE Cohort. PLoS One. 2014;9 doi: 10.1371/journal.pone.0107381. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Bahr T.M., Hughes G.J., Armstrong M., Reisdorph R., Coldren C.D., Edwards M.G., Schnell C., Kedl R., LaFlamme D.J., Reisdorph N., et al. Peripheral Blood Mononuclear Cell Gene Expression in Chronic Obstructive Pulmonary Disease. Am. J. Respir. Cell Mol. Biol. 2013;49:316–323. doi: 10.1165/rcmb.2012-0230OC. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Irizar H., Muñoz-Culla M., Sepúlveda L., Sáenz-Cuesta M., Prada Á., Castillo-Triviño T., Zamora-López G., López de Munain A., Olascoaga J., Otaegui D. Transcriptomic Profile Reveals Gender-Specific Molecular Mechanisms Driving Multiple Sclerosis Progression. PLoS One. 2014;9 doi: 10.1371/journal.pone.0090482. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Calligaris R., Banica M., Roncaglia P., Robotti E., Finaurini S., Vlachouli C., Antonutti L., Iorio F., Carissimo A., Cattaruzza T., et al. Blood transcriptomics of drug-naive sporadic Parkinson's disease patients. BMC Genomics. 2015;16:876. doi: 10.1186/s12864-015-2058-3. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Sellam J., Marion-Thore S., Dumont F., Jacques S., Garchon H.-J., Rouanet S., Taoufik Y., Hendel-Chavez H., Sibilia J., Tebib J., et al. Use of Whole-Blood Transcriptomic Profiling to Highlight Several Pathophysiologic Pathways Associated With Response to Rituximab in Patients With Rheumatoid Arthritis Data From a Randomized, Controlled, Open-Label Trial. Arthritis Rheumatol. 2014;66:2015–2025. doi: 10.1002/art.38671. [ DOI ] [ PubMed ] [ Google Scholar ] 17. Astafeva D., Gonda X., Ossina N., Volova L., Sheyfer M., Strelnik A., Lyamin A., Ashurov Z., Syunyakov T., Smirnova D. Peripheral Blood Mononuclear Cell-Related Biomarkers in Schizophrenia. Psychiatr. Danub. 2023;35:114–122. [ PubMed ] [ Google Scholar ] 18. Dann E., Cujba A.-M., Oliver A.J., Meyer K.B., Teichmann S.A., Marioni J.C. Precise identification of cell states altered in disease using healthy single-cell references. Nat. Genet. 2023;55:1998–2008. doi: 10.1038/s41588-023-01523-7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Onnis C., van Assen M., Muscogiuri E., Muscogiuri G., Gershon G., Saba L., De Cecco C.N. The Role of Artificial Intelligence in Cardiac Imaging. Radiol. Clin. North Am. 2024;62:473–488. doi: 10.1016/j.rcl.2024.01.002. [ DOI ] [ PubMed ] [ Google Scholar ] 20. Michelhaugh S.A., Januzzi J.L., Jr. Using Artificial Intelligence to Better Predict and Develop Biomarkers. Clin. Lab. Med. 2023;43:99–114. doi: 10.1016/j.cll.2022.09.021. [ DOI ] [ PubMed ] [ Google Scholar ] 21. Kiddle S.J., Sattlecker M., Proitsi P., Simmons A., Westman E., Bazenet C., Nelson S.K., Williams S., Hodges A., Johnston C., et al. Candidate Blood Proteome Markers of Alzheimer’s Disease Onset and Progression: A Systematic Review and Replication Study. J. Alzheimers Dis. 2014;38:515–531. doi: 10.3233/jad-130380. [ DOI ] [ PubMed ] [ Google Scholar ] 22. Hartman E., Scott A.M., Karlsson C., Mohanty T., Vaara S.T., Linder A., Malmström L., Malmström J. Interpreting biologically informed neural networks for enhanced proteomic biomarker discovery and pathway analysis. Nat. Commun. 2023;14:5359. doi: 10.1038/s41467-023-41146-4. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Hao J., Kim Y., Mallavarapu T., Oh J.H., Kang M. Interpretable deep neural network for cancer survival analysis by integrating genomic and clinical data. BMC Med. Genomics. 2019;12:189. doi: 10.1186/s12920-019-0624-2. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Maouche S., Poirier O., Godefroy T., Olaso R., Gut I., Collet J.-P., Montalescot G., Cambien F. Performance comparison of two microarray platforms to assess differential gene expression in human monocyte and macrophage cells. BMC Genomics. 2008;9:302. doi: 10.1186/1471-2164-9-302. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Guilliams M., Mildner A., Yona S. Developmental and Functional Heterogeneity of Monocytes. Immunity. 2018;49:595–613. doi: 10.1016/j.immuni.2018.10.005. [ DOI ] [ PubMed ] [ Google Scholar ] 26. Hoeksema M.A., de Winther M.P.J. Epigenetic Regulation of Monocyte and Macrophage Function. Antioxid. Redox Signal. 2016;25:758–774. doi: 10.1089/ars.2016.6695. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. You J., Guo Y., Zhang Y., Kang J.-J., Wang L.-B., Feng J.-F., Cheng W., Yu J.-T. Plasma proteomic profiles predict individual future health risk. Nat. Commun. 2023;14:7817. doi: 10.1038/s41467-023-43575-7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Patel H., Iniesta R., Stahl D., Dobson R.J.B., Newhouse S.J. Working Towards a Blood-Derived Gene Expression Biomarker Specific for Alzheimer's Disease. J. Alzheimers Dis. 2020;74:545–561. doi: 10.3233/JAD-191163. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Chen T., Guestrin C. XGBoost: A Scalable Tree Boosting System. arXiv. 2016 doi: 10.48550/arXiv.1603.02754. Preprint at. [ DOI ] [ Google Scholar ] 30. Wagle M.M., Long S., Chen C., Liu C., Yang P. Interpretable deep learning in single-cell omics. Bioinformatics. 2024;40 doi: 10.1093/bioinformatics/btae374. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Fortelny N., Bock C. Knowledge-primed neural networks enable biologically interpretable deep learning on single-cell sequencing data. Genome Biol. 2020;21:190. doi: 10.1186/s13059-020-02100-5. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Ma J., Yu M.K., Fong S., Ono K., Sage E., Demchak B., Sharan R., Ideker T. Using deep learning to model the hierarchical structure and function of a cell. Nat. Methods. 2018;15:290–298. doi: 10.1038/nmeth.4627. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Huang X., Huang K., Johnson T., Radovich M., Zhang J., Ma J., Wang Y. ParsVNN: parsimony visible neural networks for uncovering cancer-specific and drug-sensitive genes and pathways. NAR Genom. Bioinform. 2021;3 doi: 10.1093/nargab/lqab097. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Agarwal R., Melnick L., Frosst N., Zhang X., Lengerich B., Caruana R., Hinton G.E. Neural additive models: Interpretable machine learning with neural nets. Adv. Neural Inf. Process. Syst. 2021;34:4699–4711. [ Google Scholar ] 35. Chen H., Lundberg S., Lee S.-I. Explainable AI in Healthcare and Medicine: Building a Culture of Transparency and Accountability. Springer; 2020. Explaining models by propagating Shapley values of local components; pp. 261–270. [ Google Scholar ] 36. Shao L. DNA Damage Response Signals Transduce Stress From Rheumatoid Arthritis Risk Factors Into T Cell Dysfunction. Front. Immunol. 2018;9:3055. doi: 10.3389/fimmu.2018.03055. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Singh K., Deshpande P., Li G., Yu M., Pryshchep S., Cavanagh M., Weyand C.M., Goronzy J.J. K-RAS GTPase- and B-RAF kinase-mediated T-cell tolerance defects in rheumatoid arthritis. Proc. Natl. Acad. Sci. USA. 2012;109:E1629–E1637. doi: 10.1073/pnas.1117640109. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Chen J., Wang X., Liu Y., Zhang X. Recent advances on neutrophil dysregulation in the pathogenesis of rheumatic diseases. Curr. Opin. Rheumatol. 2024;36:142–147. doi: 10.1097/bor.0000000000000986. [ DOI ] [ PubMed ] [ Google Scholar ] 39. Srinivasan R., Bogdanov V.Y. Alternatively spliced tissue factor: discovery, insights, clinical implications. Front. Biosci. 2011;16:3061–3071. doi: 10.2741/3899. [ DOI ] [ PubMed ] [ Google Scholar ] 40. Rigamonti A., Villar J., Segura E. Monocyte differentiation within tissues: a renewed outlook. Trends Immunol. 2023;44:999–1013. doi: 10.1016/j.it.2023.10.005. [ DOI ] [ PubMed ] [ Google Scholar ] 41. Raj T., Kuchroo M., Replogle J.M., Raychaudhuri S., Stranger B.E., De Jager P.L. Common Risk Alleles for Inflammatory Diseases Are Targets of Recent Positive Selection. Am. J. Hum. Genet. 2013;92:517–529. doi: 10.1016/j.ajhg.2013.03.001. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 42. Chen B., Guo J., Ye H., Wang X., Feng Y. Role and molecular mechanisms of SGLT2 inhibitors in pathological cardiac remodeling (Review) Mol. Med. Rep. 2024;29:73. doi: 10.3892/mmr.2024.13197. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Kanika. Khan R. Functionalized nanomaterials targeting NLRP3 inflammasome driven immunomodulation: Friend or Foe. Nanoscale. 2023;15:15906–15928. doi: 10.1039/d3nr03857b. [ DOI ] [ PubMed ] [ Google Scholar ] 44. Emami Kazemabad M.J., Asgari Toni S., Tizro N., Dadkhah P.A., Amani H., Akhavan Rezayat S., Sheikh Z., Mohammadi M., Alijanzadeh D., Alimohammadi F., et al. Pharmacotherapeutic potential of pomegranate in age-related neurological disorders. Front. Aging Neurosci. 2022;14 doi: 10.3389/fnagi.2022.955735. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 45. Framke E., Thygesen L.C., Malmborg M., Schou M., Sellebjerg F., Magyari M. Risk of cardiovascular disease in patients with multiple sclerosis treated with fingolimod compared to natalizumab: A nationwide cohort study of 2095 patients in Denmark. Mult. Scler. 2024;30:184–191. doi: 10.1177/13524585231221415. [ DOI ] [ PubMed ] [ Google Scholar ] 46. Framke E., Thygesen L.C., Malmborg M., Schou M., Sellebjerg F., Magyari M. Risk of cardiovascular disease in patients treated with fingolimod compared to natalizumab: A nationwide cohort study in Denmark. Mult. Scler. 2023;29:368. doi: 10.1177/13524585231221415. [ DOI ] [ PubMed ] [ Google Scholar ] 47. Seifert R., Beste K.Y. Allosteric Regulation of Nucleotidyl Cyclases: An Emerging Pharmacological Target. Sci. Signal. 2012;5:pe37. doi: 10.1126/scisignal.2003466. [ DOI ] [ PubMed ] [ Google Scholar ] 48. Luchowska-Stańska U., Morgan D., Yarwood S.J., Barker G. Selective small-molecule EPAC activators. Biochem. Soc. Trans. 2019;47:1415–1427. doi: 10.1042/bst20190254. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Lê Cao K.A., Rohart F., McHugh L., Korn O., Wells C.A. YuGene: A simple approach to scale gene expression data derived from different platforms for integrated analyses. Genomics. 2014;103:239–251. doi: 10.1016/j.ygeno.2014.03.001. [ DOI ] [ PubMed ] [ Google Scholar ] 50. Wu T., Hu E., Xu S., Chen M., Guo P., Dai Z., Feng T., Zhou L., Tang W., Zhan L., et al. clusterProfiler 4.0: A universal enrichment tool for interpreting omics data. Innovation. 2021;2 doi: 10.1016/j.xinn.2021.100141. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 51. Knox C., Wilson M., Klinger C.M., Franklin M., Oler E., Wilson A., Pon A., Cox J., Chin N.E.L., Strawbridge S.A., et al. DrugBank 6.0: the DrugBank Knowledgebase for 2024. Nucleic Acids Res. 2024;52:D1265–D1275. doi: 10.1093/nar/gkad976. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 52. Daidone M., Ferrantelli S., Tuttolomondo A. Machine learning applications in stroke medicine: advancements, challenges, and future prospectives. Neural Regen. Res. 2024;19:769–773. doi: 10.4103/1673-5374.382228. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 53. Meng X., Wu Y., Liu W., Wang Y., Xu Z., Jiao Z. Research on Voxel-Based Features Detection and Analysis of Alzheimer's Disease Using Random Survey Support Vector Machine. Front. Neuroinform. 2022;16 doi: 10.3389/fninf.2022.856295. [ 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 Table S1. Connectivity map analysis of candidate compounds associated with ME-NET-prioritized targets mmc1.csv (56.3KB, csv) Table S2. DGIdb-based drug-gene interaction annotation of ME-NET-prioritized targets mmc2.csv (507.7KB, csv) Data Availability Statement • Raw data of the study are deposited at https://ngdc.cncb.ac.cn/gsa-human GSA-Human: HRA007646. • ME-NET is available at GitHub: ZSBiolab/ME-NET https://github.com/ZSBiolab/ME-NET . • All data presented in this study are available upon request from the lead contact upon reasonable request. Articles from iScience are provided here courtesy of Elsevier ACTIONS View on publisher site PDF (8.8 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel 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