Identification of endoplasmic reticulum stress-associated genes and subtypes for prediction of Alzheimer's disease based on interpretable machine learning

Yongxing Lai; Xueyan Lin; Chunjin Lin; Xing Lin; Zhihan Chen; Li Zhang

doi:10.3389/fphar.2022.975774

Identification of endoplasmic reticulum stress-associated genes and subtypes for prediction of Alzheimer's disease based on interpretable machine learning

Front Pharmacol. 2022 Aug 19:13:975774. doi: 10.3389/fphar.2022.975774. eCollection 2022.

Authors

Yongxing Lai^{1

2}, Xueyan Lin³, Chunjin Lin^{1

2}, Xing Lin^{1

2}, Zhihan Chen⁴, Li Zhang⁵

Affiliations

¹ Department of Geriatric Medicine, Shengli Clinical Medical College of Fujian Medical University, Fujian Provincial Hospital, Fuzhou, China.
² Fujian Provincial Center for Geriatrics, Fujian Provincial Hospital, Fuzhou, China.
³ Department of Gastroenterology, Shengli Clinical Medical College of Fujian Medical University, Fujian Provincial Hospital, Fuzhou, China.
⁴ Department of Rheumatology and Immunology, Shengli Clinical Medical College of Fujian Medical University, Fujian Provincial Hospital, Fuzhou, China.
⁵ Department of Nephrology, Shengli Clinical Medical College of Fujian Medical University, Fujian Provincial Hospital, Fuzhou, China.

Abstract

Introduction: Alzheimer's disease (AD) is a severe dementia with clinical and pathological heterogeneity. Our study was aim to explore the roles of endoplasmic reticulum (ER) stress-related genes in AD patients based on interpretable machine learning. Methods: Microarray datasets were obtained from the Gene Expression Omnibus (GEO) database. We performed nine machine learning algorithms including AdaBoost, Logistic Regression, Light Gradient Boosting (LightGBM), Decision Tree (DT), eXtreme Gradient Boosting (XGBoost), Random Forest, K-nearest neighbors (KNN), Naïve Bayes, and support vector machines (SVM) to screen ER stress-related feature genes and estimate their efficiency of these genes for early diagnosis of AD. ROC curves were performed to evaluate model performance. Shapley additive explanation (SHAP) was applied for interpreting the results of these models. AD patients were classified using a consensus clustering algorithm. Immune infiltration and functional enrichment analysis were performed via CIBERSORT and GSVA, respectively. CMap analysis was utilized to identify subtype-specific small-molecule compounds. Results: Higher levels of immune infiltration were found in AD individuals and were markedly linked to deregulated ER stress-related genes. The SVM model exhibited the highest AUC (0.879), accuracy (0.808), recall (0.773), and precision (0.809). Six characteristic genes (RNF5, UBAC2, DNAJC10, RNF103, DDX3X, and NGLY1) were determined, which enable to precisely predict AD progression. The SHAP plots illustrated how a feature gene influence the output of the SVM prediction model. Patients with AD could obtain clinical benefits from the feature gene-based nomogram. Two ER stress-related subtypes were defined in AD, subtype2 exhibited elevated immune infiltration levels and immune score, as well as higher expression of immune checkpoint. We finally identified several subtype-specific small-molecule compounds. Conclusion: Our study provides new insights into the role of ER stress in AD heterogeneity and the development of novel targets for individualized treatment in patients with AD.

Keywords: Alzheimer’s disease; ER stress; machine learning; molecular subtypes; prediction model.