Korean Journal of Biological Psychiatry

Address for correspondence: Ki-Yeol Kim, PhD, Oral Cancer Research Institute, Yonsei University College of Dentistry, Yonsei University, 50-1 Yonsei-ro, Seodaemun-gu, Seoul 03722, Korea
Tel: +82-2-2228-3043, Fax: +82-2-392-2959, E-mail: kky1004@yuhs.ac


Introduction

  Alzheimer’s disease (AD) is the most common cause of dementia in older adults with the loss of cognitive function and memory.1) The common symptoms of AD are difficulties in remembering recent events, thinking and reasoning, speaking and writing, making judgment and decisions, planning and performing familiar tasks, and changes in personality and behavior.2)

  AD is not a part of normal aging, but increasing age is the strongest risk factor for AD.3) Three in 10 people over the age of 85 years and one in every eight people over 65 years are estimated to develop AD.4) Family history and genetics, mild cognitive impairment (MCI), past head trauma, lifestyle, heart health, life-long learning, and social engagement are other risk factors of AD.5) The risk of developing AD is higher if a first-degree relative (parent or siblings) has the disease.6) The genetic mechanism of AD among families remains unexplained.7) People with MCI have a higher chance of developing AD but is not a certainty and can be prevented by developing a healthy lifestyle.8) Some studies showed that the risk factors of heart disease may also increase the risk of developing AD.9) The diagnosis of AD is usually based on the patient’s medical history, mental status testing, and physical testing, even though several histopathological markers, such as extracellular β-amyloid plaques and neurofibrillary tangles within neurons, can determine AD presence.10)

  Artificial intelligence (AI) is a method of simulating human intelligence processes using machines, especially computer systems. It can be used to analyze and improve the predictive performance of models in various research areas. Machine learning (ML), a major branch of AI, has been widely used.11-15)

  In this study, we investigated significant gene sets related to AD and identified combined biomarkers from selected significant gene sets. We also performed functional annotations for AD-related genes. We aimed to provide a systematic approach to discovering new therapeutic targets for the treatment of AD. A risk-predictive nomogram was constructed for practical usage. This nomogram could be used as an objective guideline to assess high risk for AD. With the identification of high-risk individuals, more efficient therapeutic trials can be designed to reduce the incidence of AD.

  Three publicly available gene expression datasets (GSE5281, GSE1297, and GSE48350) were used in this study. These datasets are accessible from a public microarray database (gene expression omnibus [GEO]). GSE5281 and GSE48350 consist of six brain regions and four brain regions (a total of 54675 probes), respectively. These datasets were conducted in the same GPL570 platform, Affy HG-U133 plus 2.0 (http://www.affymetrix.com/analysis/index.affx). GSE1297 includes hippocampal (HIP) gene expression data on nine controls and 22 AD patients of varying severity, incipient, and moderate and severe (a total of 22283 probes). This dataset was conducted in the GPL96 platform, Affy HG-U133A. In this study, we considered the HIP region in GSE5271 and GSE48350 and the severe stage in GSE1297. The datasets are summarized in Table 1.

  Gene ontology (GO)16) is a structured, defined, and controlled vocabulary for large-scale gene annotation. The Database for Annotation, Visualization, and Integrated Discovery (DAVID) was developed as a comprehensive functional annotation tool for relating functional terms to gene lists using a clustering algorithm.17) To analyze differentially expressed genes at the functional level, we performed pathway enrichment analysis using the online DAVID tool (https://david.ncifcrf.gov/).

  GO hierarchy contains three subontologies: biological pro- cesses, cellular components, and molecular functions. The p-value is a modified Fisher’s exact test p-value. Because DAVID examines thousands of gene sets, it is necessary to test multiple hypotheses. DAVID provides a Benjamini-Hochberg false discovery rate (FDR)-adjusted p-value, with a smaller p-value indicating enrichment. We used the p-value and Benjamini-Ho-chberg FDR to determine the significance of the enrichment of the terms for each annotation.

  In recent decades, rapid advancements in computational algorithms and the increased availability of big data have enabled AI, one of the most exciting technologies in our everyday lives, to analyze and improve the predictive performance of models in various research areas. Specifically, ML, a major branch of AI, has been used widely. The ML algorithms used in this study are as follows.18)

  Linear discriminant analysis (LDA) is a generalization of Fisher’s linear discriminant. It is a method used in statistics and other fields to find a linear combination of features that can characterize or separate two or more classes of objects or events. The resulting combination might be used as a linear classifier or for dimensionality reduction before classification. LDA is a kind of dimension-reduction technique commonly applied to supervised classification problems. It is used to model differences between groups, i.e., separating two or more groups from each other.19)

  The k-nearest neighbors (KNN) algorithm is one of the simplest techniques used in ML.20) It is used for both classification and regression and is preferred by many in the industry because of its ease of use and low calculation time. The KNN algorithm works by finding the distance between data points. The most common way to find this distance is to use the Euclidean distance. KNN computes the distance between each data point and the test data. It then finds the probability of these points being similar to the test data and classifies the data points based on which of them share the highest probability.21)

  A support vector machine (SVM) is a supervised learning model with associated learning algorithms that can analyze data for classification and regression analysis.22) The SVM algorithm is a popular ML tool that offers solutions to both classification and regression problems. It was developed at AT&T Bell Laboratories by Vapnik and colleagues.22) The objective of the SVM algorithm is to find a hyperplane in the N-dimensional space (where N is the number of features) that distinctly classifies the data points. Support vectors are data points that are closer to the hyperplane and can inf luence the position and orientation of the hyperplane. Using these support vectors, the margin of the classifier is maximized. Deleting the support vectors will change the position of the hyperplane.

  Random forest (RF) is an ensemble learning method for classification, regression, and other tasks. It operates by constructing a multitude of decision trees at the training time, and outputting the class that is the mode of the classes (classification) or the mean/average prediction (regression) of individual trees.23) RF algorithms can be used to solve both regression and classification problems, making it a diverse model that is widely used by engineers.

  All ML models were implemented using R programming language, version 4.1.3 (R Foundation for Statistical Computing, Vienna, Austria),1) including GEOquery and limma packages for downloading the GEO datasets and identifying significant genes. The nomogram for predicting AD risk was created based on the significant genes selected.

  In this study, we compared the accuracies of ML algorithms for predicting AD risk using the identified combined biomarkers. The study process is shown in Fig. 1.

  We detected 31 differently expressed common genes between the normal and AD groups using the limma package. Fig. 2 shows the expression patterns of the 31 common genes in the three different datasets (Fig. 2).

  The identified genes showed highly divergent expression patterns between the normal and AD groups in the GSE5281 and GSE1297 datasets but not in the GSE48350 dataset (Fig. 2C). The upregulated genes in the AD group included COX7C , DOCK3, CDK5, SLC25A12 , PLK2 , IQSEC1, NDUFB8 , CAPRIN2, WDFY3, and PKP4 . On the other hands, RXF2, TTN, RHOQ, SFTPB, and KIAA0485 were down-regulated in the AD groups (Fig. 2A and B).

  We next performed GO term annotation and pathway enrichment analysis of the differentially expressed genes at the functional level using DAVID tool. The results are summarized in Table 2.

  The AD-related genes were APC, NDUFB8 , CDK5, COX7C , and ITPR1. The Huntington’s disease-related pathway was also included, although the FDR-adjusted p-value was not significant. The pathway significance may have been underestimated by selecting a small number of common genes (31 common genes were used for annotation). The most significant enriched GO term was “serine/threonine-protein kinase” (Table 2).

  To select the optimal number of combined biomarkers for risk prediction, we tested random sets of 1–5 genes, with 200–500 replicates and evaluated the association between the number of genes and the predictive model accuracy (Fig. 3).

  Predictive accuracy indicates the probability of the concordance between the predicted and observed responses (logistic regression). The accuracy increased with increasing numbers of combined genes. We focused on identifying a predictive model based on the least number of genes. Therefore, we selected a two-gene set for further analysis. Table 3 summarizes the five combined sets of two genes resulting from the simulations shown in Fig. 3.

  We then compared the performance of the risk-predictive models using different ML algorithms. For this experiment, the dataset was randomly split into training (70% of data) and testing (30% of data) datasets. The random dataset spilt was processed repeatedly 100 times, and the model performance was summarized according to the mean values and standard deviations calculated for all processing cycles (Table 4).

  We compared the performance of two model types, one predicting the risk probability of AD based on a single gene and the other predicting the risk based on combined biomarkers. The performance of the models based on combined biomarkers was superior to that of the models based on single genes (Table 4). The predictive accuracies ranged from 0.814 to 1.000 for the training dataset and from 0.708 to 0.803 for the testing dataset for single gene models. For models based on combined biomarkers, RF was the best-performing model with the training dataset (accuracy = 1.0). Performance with the test datasets tended to depend on the combined biomarkers used.

  When we tested the models in GSE1297 and GSE48350, the performance in GSE1297 was much better than in GSE48350. This can be interpreted as a characteristic of the dataset as GSE1297 consists of normal and severe AD (Table 1). Considering the accuracy of the three datasets, we constructed a predictive nomogram by combining LZTS1 and SORBS2 genes.

  A nomogram was constructed to predict the AD risk probability using ATR and PRKCB (Fig. 4A) as the best combination of two genes for predicting the risk of AD (Table 4). The predictive accuracy was 80.9% in the GSE48350 dataset using the LDA model.

  ATR and PR KC B are genes upregulated and downregulated in AD, respectively. The risk probability for AD increased when the total nomogram points decreased. If the total points were 60 for a patient, the AD risk probability was about 50% (Fig. 4A). For practical usage of the nomogram, we constructed a nomogram in Hypertext Markup Language format and populated it with calculated total scores and probabilities (Fig. 4B). Fig. 4B shows that the AD risk probability of a patient with expression values of AT R and PR KC B of 0.6 and 0.4, respectively, was 32.5%. Additionally, the total points calculated by the nomogram could be used for stratifying patients according to AD risk probability.

  This study identified 31 significant genes for AD risk prediction. Among them, L z t s1 was reported to control both neuronal delamination and outer radial glial-like cell generation during mammalian cerebral development, suggesting a role in neuronal development.24)25)

  The neurodevelopmental spectrum was seen with CHD2 variants.26) DOCK3-related neurodevelopmental syndrome was reported in a boy with developmental delay and hypotonia.27) Genetic variants in SLC9A9 were associated with measures of attention-deficit/hyperactivity disorder symptoms in families.28)

  Kirtay et al.29) identified a physiological function of ATR beyond its DNA damage response role, in regulating neuronal activity. Gerschütz et al.30) reported that PRKCB and MAPK1 were increased in the late AD stages. MAPK1 and PRKCB levels were low in the brainstem and cerebellum. The authors proposed that alterations in the expression of these two genes occurred early in the pathogenesis of AD in a region-specific manner. Antonell et al.31) and Zhou et al.32) also reported that the low expression of PRKCB was a potential causative factor of AD. This study also confirmed that PRKCB was downregulated in the AD group (Fig. 2A and B).

  While ML can easily identify dataset trends and patterns, it requires massive datasets for training. Due to the small sample size, the limitation of this study was that the dataset could not represent the entire population of patients with AD. A model trained on a random sample of a dataset might have poor generalizability and perform poorly outside of that sample. Indeed, the use of larger training and test sets resulted in more accurate and reliable predictions.33)

  The implemented predictive model was presented in the form of a diagram, referred to as a nomogram. A nomogram is a graphical representation of a statistical model. It provides the probability of a particular clinical outcome. The nomogram introduced in this study can serve as an objective guideline to assess high risk for AD. With the identification of high-risk individuals, more efficient trials can be designed to reduce the incidence of AD. The constructed nomogram could be used as a test version. A predictive model using clinical variables and specific gene expression in large datasets of aging populations is needed in the future.

  This work was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (2021R1I1A1A01048175).

  The author has no financial conf licts of interest.

  Ki-Yeol Kim https://orcid.org/0000-0001-5357-1067