A Case study on The Prediction of Feature Genes in Trauma Patients

Highlights:

1 A total of 133 feature genes were screened out in training set;

2 SVM classifier peaked at 100% and 86.2% correct rate in training and validation [A1]sets;

3 Feature genes were mainly enriched in cell proliferation and ribosome pathway;

Abstract

Purpose: Tumor necrosis factor (TNF)-α variation is closely linked to sepsis syndrome and mortality after severe trauma. We aimed to identify feature genes for prediction trauma patients with or without TNF-α variation to help direct them toward alternative successful treatment.

Methods:[A2] In this study, we used 58 sets of gene expression data of trauma patients from Gene Expression Omnibus to predict the feature genes involved in TNF-α variation. We applied support vector machine (SVM) classifier model for prediction combining leave-one-out cross validation method. Functional annotation of feature genes was carried out to study the biological function.

Results: By comparing the gene expression profiles of trauma patients with or without TNF-α variation, 133 feature genes were screened out and well differentiated the training set (14 patients with variant, 15 with wild type)[A3] with 100% correct rate and 86.2% in validation set[A4]. Interestingly, many cell proliferation-related genes and ribosome-associated genes were abnormally expressed in trauma patients with TNF-α variation.

Conclusion: A machine learning approach SVM has been applied for accurately identifying feature genes and classifying trauma patients with or without TNF-α variation. The method provides an efficient way to identify specific genes for finding novel biomedical markers and drug target in severe trauma treatment.

Keywords: severe trauma, tumor necrosis factor-α variation, support vector machine, feature gene, classification

Introduction

Severe trauma remains to be a global health problem and the second leading cause of death under the age of 45 years [1]. Patients subjected to severe trauma are at a risk of developing sepsis and organ system injury that result in organ dysfunction. It is well known that severe trauma and sepsis are related to a high rate of immune irregularities due to an acute inflammatory response in injured sites after encountering trauma or sepsis. The inflammatory response is the host’s response to extraneous threat such as invading pathogen or tissue damage and subsequent up-regulation of pro-inflammatory cytokines, chemokines and immune cells into tissues [2, 3].

The cytokine response plays an important role in the development of severe trauma in that pro-inflammatory cytokines including tumor necrosis factor α (TNF-α), interleukin (IL)-6, 8, 12 and granulocyte-macrophage colony-stimulating factor (GM-CSF) are released excessively. TNF-α is considered as an essential component in immune response to inflammation that primarily produced by immune cells including macrophages, monocytes and other non-immune cells [4, 5]. Currently, TNF-α acts as a key intermediary in inflammatory response that mediates a cascade of cytokines. Furthermore, TNF-α is responsible for organ failure as well as increased risk of sepsis after multiple injuries [6, 7]. A previous study has revealed that elevated circulating levels of TNF-α are observed in response to severe trauma [8]. Several single nucleotide polymorphisms (SNPs) within the promoter region or coding sequence have been thought to influence TNF-α production, therefore, identified as candidate variants that might be associated with outcome of trauma [9, 10]. Recently, a cohort study of patients with severe trauma has indicated that common TNF-α variants are related to sepsis syndrome and death following severe trauma [11]. The peripheral blood transcriptome in patients with or without TNF-α variant is different and worth to gaining more insight. Therefore, identification of genes related to TNF-α variant may help improve therapy of trauma. However, classification of trauma patients based on TNF-α variant has never been reported in previous study, as well as feature genes associated with TNF-α variation.

Prediction of feature genes associated with TNF-α variation in trauma patients help researchers to elucidate the molecular mechanism of severe trauma and is crucial for drug target selection in clinical application. Various feature gene identification methods in handling microarray data have been applied, especially support vector machine (SVM) which is one of the most effective method in classification [12].

In this study, we developed a systematic tool by applying the SVM method in conjunction with leave one out cross validation (LOOCV) for identifying feature genes using microarray data. Moreover, functional annotation of feature genes was conducted. The results indicate that SVM classifier achieved better predictive performance for feature gene prediction, as well as classification of patients with or without TNF-α.

2 Materials and methods

2.1 Data source

The dataset GSE5760 [A5]based on the platform of GE Healthcare/Amersham Biosciences CodeLink UniSet Human I Bioarray from GEO (Gene Expression Omnibus) of NCBI (http://www.ncbi.nlm.nih.gov/geo/) was used to classify trauma patients with or without variations in TNF-α. The complete dataset contained 28 blood samples with variation in TNF-α and 30 samples without variation. In this study, the first 29 out of 58 samples were used as training set (14 samples with variation and 15 samples without variation) and the remaining samples were used as validation set. A schematic diagram of analysis in this study was shown in Figure 1.

2.2 Data preprocessing

Raw expression data were normalized using Robust Multi-array Averaging (RMA) [13] method with quantile normalization. The probe data were converted into gene expression data and mean value of several probes was calculated as the gene expression value. Data were then log 2 transformed for further analysis.

2.3 Identification of feature gene in training set

To well distinguish samples with or without variation in TNF-α, Limma package [14] was applied to calculate the differential expression and generate feature genes in training set. P values were adjusted for multiple testing by Benjamini-Hochberg procedure as the false discovery rate (FDR) [15]. Genes with FDR less than 0.05 and fold change (FC, mutation/wild type) more than 1.5 were considered to be significant feature genes.

2.4 Class prediction analysis

An advanced classifier widely used statistical learning method, SVM algorithm [16] was applied to classify the samples based on the differentially expressed feature genes. SVM analysis was conducted using the e1071 package in R (http://cran.r-project.org/package=e1071) [17]. To evaluate the quality of the classifier, LOOCV [18] which involves iteratively leaving out 1 sample as testing sample and then generating feature gene using the remaining training samples was applied.

2.5 Evaluation parameters for SVM

LOOCV was applied to assess the performance of all classification models. In this report, the classification models for samples with or without variation in TNF-α were evaluated based on the calculation of sensitivity, specificity [19], positive predictive value (PPV), negative predictive value (NPV) [20] and the area under the receiver operating characteristic (AUROC) curve [21]. If the five parameters equal 1, a perfect classifier can be found. However, if they are equal to 0.5, the classifier has no discriminative power at all.

2.6 Functional annotation of the expression data

To further study the functions of feature genes, Gene Ontology (GO) functional analysis was conducted using the online tool DAVID (database for annotation, visualization, and integrated discovery, available at http://david.abcc.ncifcrf.gov/) [22]. KEGG (kyoto encyclopedia of genes and genomes) pathway enrichment analysis was carried out by KOBAS (KEGG Orthology Based Annotation System, available at http://kobas.cbi.pku.edu.cn/home.do) [23]. P value was calculated by hypergeometricdistribution. Go terms with FDR less than 0.05 and KEGG categories with P value less than 0.05 were considered to be significant.

3 Results

3.1 Identification of feature genes in training set

Using the statistical values (FDR < 0.05) and fold differences (FC > 1.5) as filtering criteria, 133 genes including 86 up- and 47 down-regulated genes were screened out that differentiated trauma patients with TNF-α variation from patients without TNF-α variation.

3.2 Evaluation of SVM predictive classifier

To generate a predictive classifier from this cohort, gene expression values were calculated. This classifier SVM, using LOOCV, peaked in predictive accuracy with 100% correct rate in training set indicating that all 29 samples were distinguished by SVM, but 86.2% in validation set (Figure 2 A and B). There were four samples were wrong differentiated, including two mutants and two wild types (Figure 2 C). To further assess the effects of SVM classifier, parameters including correct rate, sensitivity, specificity, PPV, NPV and AUROC were calculated and shown in Table 1. The ROC of training or validation set was shown in Figure 3. The results demonstrated the high correct rate, sensitivity, specificity, PPV and NPV for predicting trauma patients with TNF-α variation and without variation in the training and validation cohorts.

3.3 Functional annotation of feature genes in training set

To examine which GO function and signaling pathways were differentially expressed in peripheral blood from trauma patients, we performed GO functional and KEGG pathway enrichment analysis of the 133 feature genes. Interestingly, feature genes were mainly involved in GO terms of cell proliferation, protein import, negative regulation of cell proliferation and so on. Top 10 GO terms were shown in Table 2. The percentage of feature genes involved in GO terms was shown in Figure 4. Most of feature genes were enriched in cell proliferation and response to organic substance, including HMOX1 (heme oxygenase (decycling) 1), HDGFRP3 (hepatoma-derived growth factor, related protein 3) and ASPM (asp (abnormal spindle) homolog, microcephaly associated). In addition, ribosome pathway-related genes including MRPL1 (mitochondrial ribosomal protein L1), MRPL22 (mitochondrial ribosomal protein L22), RPS7 (ribosomal protein S7) and RPL6 (ribosomal protein L6) were significantly up-regulated in trauma patients with TNF-α variation. The significant signaling pathways were shown in Table 3.

4 Discussion

In this study, we developed a robust classification model for trauma patient prediction using machine learning method SVM. We identified 133 feature genes using filtering criteria of FDR < 0.05 and FC > 1.5 in the training set. The identified gene set could be valuable for screening of trauma patients with or without TNF-α variation. The analysis based on SVM using identified feature genes predicted trauma patients with TNF-α variation in the training set with a correct rate of 100% and 86.2% in the validation set, implying that SVM classifier successfully identified the trauma patients with TNF-α variation in the training and validation sets with a high accuracy. Functional annotation indicated that feature genes were mainly enriched in cell proliferation terms and ribosomes pathway, such as HMOX1 and PRS7.

The TNF-α rs1800629 polymorphism has been described to be associated with sepsis syndrome and death after severe trauma [11]. Moreover, TNF-α -308G/A variant is associated with survival in sepsis or septic shock of [24]. As we all know, TNF-α plays crucial roles in cellular response to inflammation, such as cell proliferation, apoptosis and production of inflammatory mediators as well as pathogenesis of septic shock [25]. Changes of associated gene expression would be found after TNF-α polymorphisms, such as HOMX1, RPS7. HOMX1 is a stress protein belonging to heat shock protein family that protects against oxidative stress and inflammation [26]. In other words, HOMX1 can attenuate inflammation in a wide range of conditions by decreasing leukocyte infiltration and inflammatory adhesion molecules, but down-regulation of HOMX1 contributes to inflammation [27]. It has been proved to be involved in joint damage and disease activity of rheumatoid arthritis that is an inflammation-induced disease [26]. More interestingly, HMOX1 has been found to be mediated by NFκB pathway that involved in cellular activity, such as injury, stress and complex immune response. Meanwhile, NFκB induces expression of TNF-α [28]. There is a great deal of evidence supporting correlation between NFκB signaling and sepsis suggesting that the pathway significantly regulates inflammation [29]. Therefore, TNF-α variant may induce differential expression of HMOX1 in trauma patients.

The local release of pro- and anti-inflammatory cytokines and chemokines after severe trauma implies their ability to alter immune response. The important pro-inflammatory cytokines associated to trauma consist of TNF-α, IL-6, IL-8 and so on. Ribosome is the essential part in transcription and translation of genes involved in inflammation and immune response. More importantly, overproduction of cytokines is also involved in ribosome pathway which was consistent with pathway enrichment analysis in our study. RPS7 is considered as a gene involved in ribosomal activation and maturation of ribosomal RNA that expressing in lymphocyte, neutrophil and monocyte [30]. Furthermore, RPS7 plays a vital role in cell development via interaction with p53 and regulation of matrix metalloproteinase family [31]. Genetic deficiency of RPS7 induces apoptosis and cell cycle arrest by activating p53. However, little is known about the role of RPS7 in inflammation and immunology after trauma. In our study, overexpressed RPS7 was observed in trauma patients with TNF-α variation. As a result, RPS7 may be involved in severe trauma with TNF-α variation through regulating ribosome pathway.

In classification using SVM algorithm, the specificity, sensitivity, PPV and NPV for the feature genes were well tolerated for the prediction of trauma patients with or without TNF-α in the training and validation sets. In addition, AUROC in training set was 1, whereas 0.89 in validation set, implying that the classifier significantly classified trauma patients. Nevertheless, these results need to be regarded as preliminary because of the still-limited number of investigated patients. Meanwhile, the results are urgent to be confirmed by experimental methods.

In summary, a machine learning approach SVM has been applied in this report for classification of trauma patients with or without TNF-α. A total of 133 feature genes were identified from training set by SVM. The predicted genes show clear expression patterns in corresponding trauma patients, and they have related functions consistent with TNF-α variation. The analysis of candidate genes suggests that our approach can well classify trauma patients for further experimental studies. The approach could also provide more insights for trauma patients-selective gene prediction and be used to find new drug targets.

[A1]在figure legends和Table中使用的是testing set，请统一说法

[A2]æ-¹æ³•æè¿°æœ‰äº›ç®€å•ï¼Œåœ¨å­-数要求范围内可以详细叙述添加使用软件工具

[A3]实验组和æ-¹æ³•ç»„最好在æ-¹æ³•ä¸­ä»‹ç»æ¸…楚

[A4]对于 SVM评估结果的sensitivity, specificity, PPV, NPV 和 AUROC也是重要的评价指标，可以添加一句话描述

[A5]添加原作者æˆ-者参考æ-‡çŒ®

Which was deposited by ***