N-semble-based method for identifying Parkinson’s disease genes

Parkinson’s disease (PD) was first described by Dr. James Parkinson as a “shaking palsy” in 1817 [1]. It is the second most common disease after Alzheimer’s, most prevalent among the elderly. PD is a chronic, progressive neurodegenerative disease associated with the central nervous system. PD is affected by continuous degeneration of dopamine-producing neurons in the pars compacta of the substantia nigra. Dopamine is a chemical messenger acting as a connector that sends messages from the body to the brain. PD mainly affects neurons, thereby reducing the level of dopamine; as a result, the abnormal brain movements that promote the onset of Parkinson's enable movement control [2]. Healthy people have higher dopamine levels than people with PD. Figure 1 illustrates the dopamine level of normal and Parkinson-affected neurons. The PD genes identification method helps to detect underlying molecular mechanisms and diagnose the disease efficiently. It is a time-consuming and expensive task to identify PD-related genes from a large number of unknown genes with experimental methods. Therefore, there is a need to identify genes with computational methods which have been used to discover similar features between disease genes and unknown genes.

We have introduced a novel n-semble method to identify Parkinson’s disease genes. Geary autocorrelation (GA), Moran autocorrelation (MA) and normalized Moreau–Broto autocorrelation (NA) representation methods on the basis of physicochemical properties of amino acids are applied to translate corresponding protein sequences into a feature vector. The t-Distributed Stochastic Neighbor Embedding (t-SNE) feature extraction technique is adopted to reduce high-dimensional features. In the absence of information about negative data, Jaccard similarity measure is employed to extract a reliable negative gene set from an unknown gene set. Finally, various ML methods such as Support Vector Machine (SVM), Random Forest (F), Adaboost, Decision Tree (DT), Xgboost, Neural network and Gradient Descent are used to identify genes responsible for PD.

In recent years, several Machine Learning (ML) methods have been proposed to identify the similarity between disease and candidate genes.

Xu and Li [14] applied K-nearest neighbor (KNN) with PPI topological features to identify disease- related genes. Smalter et al. [9] employed PPI-topological properties to generate features and the Support Vector Machine (SVM) classifier to identify Parkinson’s disease genes. Radivojac et al. [10] used three types of feature vectors such as PPI properties, protein sequences, and protein-functional information to propose a method by building three individual Support Vector Machine (SVM) classifiers to predict disease genes. All the above methods consider a two-class classification problem with disease data as a positive set and unknown genes data as a negative set. As the negative set may consist of several disease genes, the negative set may be noisy and leads to reduction in accuracy.

Mordelet et al. [12] developed an algorithm to prioritize disease genes using positive and unlabelled samples named ProDiGe. They had selected a random subset (RS) from Unknown genes (U). Also, they integrated various sources of information related to genes which can be divided into features including protein sequences, protein functional information, and PPI data. Then, they used SVM to train various classifiers to distinguish positive genes from the subset RS. The final result was obtained by combining all of the prediction results. Its performance is superior to traditional binary classification methods that used unknown genes as a negative set but still suffer from noise as the negative set is separated randomly from the unknown set, thereby deteriorating performance. Yang et al. [15] developed a method named PUDI that used the PPI network, gene ontology and protein domains biological networks. According to the similarity between both positive and negative genes, the unlabelled set (U) was separated into numerous subsets, which are called reliable negative, likely negative, likely positive and weak negative respectively. Finally, they applied multi-layer weighted SVM for disease gene prioritization. Yang et al. [16] expanded their prior work and proposed an overall positive unlabelled learning method (EPU) for disease gene identification. They integrated other data sources such as phenotype similarity networks and gene expression data along with previous data sources. However, they used PPI topological properties, the protein domain and gene ontology data to generate feature vectors which contain more than 4,000 features. For neurological diseases, EPU method achieved F Score of 78.6%. Experimental results have confirmed that the classifier built with high-dimensional features may not be an efficient in terms of detection of PD. Hwang [17] proposed a SRF (Stepwise Random Forests) approach for disease gene identification on biological data sources used by Yang. He had enhanced his method by considering only important features with filter-based feature selection method for classification. Further, it was analysed that he may not be able to achieve better classification results by using all the 4004 features. However, his method considered only 23 features and performed significantly better than the existing methods.

Yousef et al. [11] introduced a sequence-based one-class classification method to identify disease genes. The Support Vector Data Description (SVDD) method is used by them to train the model and selected the significant features with PCA. These methods select a few numbers of unknown genes as the negative set since the unknown genes set usually contains certain disease genes, which reduces the confusion in the classification process. However, these methods are not reliable or robust because the negative results obtained from unlabelled genes are plagued by noisy data.

Miao et al. [18] proposed an Alzheimer’s disease gene identification method based on multiple classifier integration with microarray data. They had adopted the ReliefF feature selection method to extract relevant features and then produced a two-stage cascading classifier to identify genes. Results from SVM, RF and Extreme Learning Machine (ELM) were merged through majority voting for classification. Peng et al. [19] developed a N2A-SVM method for PD gene prediction. They had used the node2vec method for feature extraction and reduced the features using the deep neural network (auto encoder) and finally SVM to predict the genes. Malhi et al. [20] put forth an ensemble method with five best machine learning methods out of 25 regression models on publically available datasets of voice measures of PD patients. Guruler et al. [21] introduced a hybrid method with k-means clustering-based feature weighting and a complex valued Artificial Neural Network (KMCFW) method on speech and sound signals to diagnose PD. The dataset consisted of only 31 people with 8 healthy and 23 PD patients. Senturk [22] used CART, SVM and ANN classifiers to classify Parkinson’s patients. They developed a feature selection-based system using voice signals features.

Some of the methods aim to prioritize disease genes using protein–protein interaction (PPI) [3] data, gene expression profiles [4] and gene ontology [5]. Unfortunately, all the above-mentioned methods depend on the knowledge of proteins attained from PPI data, protein domains and gene ontology. Therefore, these methods cannot be able to implement properly because the information is expensive, time consuming and suffers from a multitude of missing values. Protein sequences are the only data that can be used for proteins and contribute significantly to resolving issues such as protein–protein interactions [6, 7], predicting subcellular locations [8], and functional classes. The key difference between the computational methods is in the type of data used to generate feature vectors and the type of algorithm used to train the model. Some other methods considered the unknown proteins (genes) as a negative set and known disease proteins (genes) as a positive set [9, 10], while other considered this as one class classification method by training only positive data [11]. Since the unknown proteins usually contain certain disease proteins, some of the methods aim to reduce this problem in classification process by extracting the most reliable proteins as a negative protein set [12, 13].

From the above-mentioned methods, we concluded that the research conducted in the field of gene prediction is mainly restricted to the SVM classifier. Also, the existing methods are trained for multiple disease genes data, but not limited to Parkinson’s disease only. However, some of the methods used only six physicochemical properties of amino acids and recommended adding more to achieve better classification performance. Therefore, we have employed twelve physicochemical properties of amino acids to represent the features. Hence, using more physicochemical properties will allow us to provide more information about the interactions. Since there is no information about negative data, we have also selected reliably negative genes from unknown genes using the Jaccard distance metric. Then we have applied various classification methods to yield the final prediction results.

To identify Parkinson’s disease genes, sequence representation methods with physicochemical properties of amino acids are chosen to improve the efficiency of existing machine learning classifiers. In this paper, we have employed twelve physicochemical properties of amino acids to represent the amino acid features. Therefore, relying on more physicochemical properties will allow us to discover more information about the interactions. However, the increased characteristics lead to generate more features for each protein, which is why we have normalized the output feature vector, instead of concatenating the feature vector of two proteins. A novel n-semble method is proposed to develop an efficient disease gene identification method.

Mathematically, a problem statement is defined as follows: to classify G = {PD, nPD} for a protein sequence S where S = {a₁, a₂, a₃…, a_n}and a_i represents the amino acid in a sequence. Our task is to evaluate the best machine learning classifier with efficient features to calculate high efficiency in our proposed method.

The proposed n-semble method for identifying PD genes has been described in this section. The proposed approach consists of four steps: (1) adopting twelve physicochemical properties to transform corresponding protein sequences into feature vectors; (2) t-Distributed Stochastic Neighbor embedding (t-SNE) applied to reduce dimensionality (3) differentiating negative samples from unknown genes; (4) modelling features using n-semble method. The proposed method architecture is depicted in Fig. 2.

The performance of the proposed n-semble method on the imbalanced data set is evaluated in this section. First, we investigated the impact of three sequence representation methods such as GA, MA and NA on the performance of n-semble method. Additionally, an optimal number of features retrieved through the t-SNE method has been reviewed and optimized. Then, the effect of several machine learning methods has been evaluated, and on the predictions of top three models, a neural network is trained to develop an ensemble method. Finally, our method and another disease gene identification method were compared to confirm the method effectiveness. The ML methods applied in this work are Support Vector Machine (SVM), Random Forest (F), Adaboost, Decision Tree (DT), Xgboost, Neural network and Gradient Descent. Table 3 shows the values of various performance measures, i.e. Precision, Recall, and F Score for comparative analysis of the ML models experimented with. The top three models, random forest, adaboost and xgboost were selected on the basis of the highest F Score used to generate an ensemble method. The prediction values evaluated by means of each selected model are used as training data for the neural network, and the actual prediction values are used as target data. As shown in Table 3, the proposed ensemble method outperforms other methods with Precision (88.9%), Recall (90.9%) and F-Sore (89.8%). It was observed that the proposed method outperforms the Adaboost by 2.8%, Xgboost by 4.5% and Random Forest by 5.4%. To evaluate the predictive performance of all methods, the ROC (Receiver Operating Characteristic) curve is plotted. The performance of True Positive Rate (TPR) versus False Positive Rate (FPR) at various thresholds for the selected methods is shown in Fig. 4. Random Forest outperforms other methods with the area of 83.6% under ROC.

The main objective of this paper is to identify genes associated with Parkinson’s disease with the best known classification methods. To specify the conditions under which a classification method outperforms other classifiers is a key question in machine learning. This paper, introduced various methods, including Support Vector Machine (SVM), Random Forest (RF), Adaboost, Decision Tree (DT), Xgboost and Gradient Descent for genes identification. After evaluating and analysing the classification methods, more emphasis is placed on exploiting the strengths of a model to complement the weaknesses of another. Therefore, an n-semble method was proposed which trained a neural network in a special way and integrated three classification methods based on their F Score to ensemble the predictions and to achieve more accurate predictive analysis. On the basis of various performance measures, results from the proposed n-semble method show enhanced performance compared to state-of-the-art works. We have adopted protein sequences based on previous knowledge to extract features. GA, MA AND NA representation methods with twelve physicochemical properties of the amino acids are adopted to convert protein sequences into numerical feature vectors. Consequently, t-SNE is applied to extract relevant features. We found that physicochemical properties of amino acids would be highly beneficial in extracting features. Compared with the previous methods on unbalanced datasets, the proposed n-semble method improves the F Score.

In this paper, we have shown that the GA representation method is characterized by a higher success rate than other representation methods. Therefore, in the future, we will consider using a single GA feature vector to combine multiple different classifiers to improve classification. We will also use this method in the prediction of other related diseases.