Schizophrenia is a complex psychiatric disease that is affected by multiple genes, some of which could be used as biomarkers for specific diagnosis of the disease. In this work, we explore the power of machine learning methodologies for predicting schizophrenia, through the derivation of a biomarker gene signature for robust diagnostic classification purposes. Postmortem brain gene expression from the anterior prefrontal cortex of schizophrenia patients were used as training data for the construction of the classifiers. Several machine learning algorithms, such as support vector machines, random forests, and extremely randomized trees classifiers were developed and their performance was tested. After applying the feature selection method of support vector machines recursive feature elimination a 21-gene model was derived. Using these genes for developing classification models, the random forests algorithm outperformed all examined algorithms achieving an area under the curve of 0.98 and sensitivity of 0.89, discriminating schizophrenia from healthy control samples with high efficiency. The 21-gene model that was derived. from the feature selection is suggested for classifying schizophrenic patients, as it was successfully applied on an independent dataset of postmortem brain samples from the superior temporal cortex, and resulted in a classification model that achieved an area under the curve score of 0.91. Additionally, the functional analysis of the statisticallysignificant genes indicated many mechanisms related to the immune system.

Keywords: classification, schizophrenia, machine learning, gene expression, microarray studies, support vector machines, adaboost

AUC, area under the curve; CV, cross-validation; GO, gene ontology; extra trees, extremely randomized trees; kNN, k-nearest neighbor; RF, random forests; RMA, robust multi-array analysis; ROC, receiver operating characteristic; SVM, support vector machines; SVM-RFE, support vector machines recursive feature elimination; SZ, schizophrenia

Schizophrenia (SZ) is a serious psychiatric disease, with a complex genetic basis that affects around 1% of the population worldwide. The symptoms of the disease are divided into positive, negative and cognitive symptoms. Positive symptoms include hallucinations, delusions as well as disorganised speech and behaviour. Negative symptoms include anhedonia, social withdrawal, and lack of motivation and energy. Finally, cognitive symptoms involve cognitive dysfunctions of patients suffering from SZ. Pharmacological treatment of the disease mostly deals with the positive, psychotic symptoms of the disease, but does not improve cognitive and social dysfunction. Moreover, the etiology of SZ predicates upon a combination of genetic and environmental factors, probably in early life, that affect neurogenesis and neuronal plasticity.¹ DNA microarray technologies enabling genome-wide gene expression profiling have been intensely exploited in the last decade, in order to promote the elucidation of the underlying biological mechanisms of SZ.^2–5 These studies, through the high dimensional data that they yield, can prove to be very useful for the generation of diagnostic biomarker signatures in the management of SZ. The usefulness of these data is based on the fact that they may reveal several genes that act synergistically. Probably, the genes that present these synergistic effects with other genes cannot be associated with SZ on their own. The importance of the development of classification models in SZ is great as, at the moment, the diagnosis of the disease is based exclusively on the evaluation of the clinical symptoms after they have manifested. Despite much research effort, some of the most crucial questions regarding SZ have not been answered. The heterogeneity and the multi-factorial background of SZ suggest the study of this disease through statistical methods for the identification of patterns in the data. Differentially expressed genes occurring from microarray experiments can be utilized as classifying biomarkers gain and can reveal underlying genetic factors in relation to important psychiatric diseases, such as SZ.⁶

Classification includes two main methodological models: the supervised and the unsupervised model. In unsupervised learning, the instances are unlabeled and the aim is to discover useful classes.⁷ Supervised learning includes instances with known labels. In this study, supervised methods are used.⁶ In supervised learning, the classes are first defined and then the aim is to build a classifier that can separate samples among the defined classes citation.⁸ The discrimination of the classes in this study is based on the gene expression profiles of the samples.

Algorithms

Support vector machines (SVM)

In SVM, good separation of classes is achieved by the hyper-plane that has the largest distance to the nearest training data points of any class. The instances that are on the boundaries of the margin and determine the position and the orientation of the hyperplane are called the support vectors.⁹ SVM have some mathematical attributes that make them advantageous for gene expression classification, such as their ability to deal with large feature spaces and their ability to recognise outliers.¹⁰

Extremely randomized trees (Extra Trees)

The Extra Trees classifier belongs to the tree classifier algorithms and is extremely randomized. Its difference from other tree algorithms lies in the way it is built. At the point where the algorithm seeks the the most discriminative thresholds for the separation of the samples of a node into two groups, random thresholds are drawn for each of the randomly-selected features. Then, the best randomly-generated threshold is chosen as the splitting rule.¹¹

Random forests (RF)

The RF classification algorithm is based on an ensemble of classification trees. Each classification tree is developed with bootstrap sampling of the data and for each split a random subset of the variables is used. RF uses two approaches: bagging or bootstrap aggregation that combines unstable learners and random variable selection for building the tree. No pruning is applied on the trees, in order to achieve low-bias trees. Additionally, bagging and random variable selection create trees with low correlation. As a classification method in microarray studies, it gives good performances even with noisy predictive variables and for this reason, it doesn’t need gene pre-selection. Finally, good performance is not so dependent on fine-tuning the parameters of the algorithm.¹²

 Nearest neighbors

The k-nearest neighbors (kNN) algorithm is one of the most widely used and simplest methods among machine learning classification algorithms. In the training process, the kNN algorithm classifies an unlabeled instance based on the most common label of its k-neighbors in the training set. The distance metric that is used for the identification of the nearest neighbors affects the performance of the classifier.^13–15

 Adaboost

This classification algorithm boosts the performance of a simple classifier by combining a set of weak classifiers to a stronger learning algorithm. In this way, the weak classifiers have to perform only a little better than a random guessing, but the final combined classifier usually results in a good performance. In order to boost a weak classifier, it is forced to solve a series of learning problems. After every learning round, the examples are weighted and the importance of the ones that were falsely classified by the previous weak classifier is increased.¹⁶

Evaluation

In this specific study, cross-validation (CV) has been used as an evaluation method of the classifier. In n-fold CV, the training set is divided into n subsets. One after the other, one subset is used as a test subset for the trained classifier and the remaining n-1 subsets are used as the training subset.¹⁷ The performance of the different classification algorithms is evaluated through receiver operating characteristic (ROC) curves.¹⁸ In binary classification the outcomes can be labeled either as positive or negative. The true positive (also known as sensitivity or recall) rate refers to the proportion of positive samples that are correctly predicted as positive, whereas the false positive (also known as 1-specificity) rate refers to the proportion of negative examples that are incorrectly predicted as positive. The Y axis of the ROC curve represents the true positive rate and the X axis represents the false positive rate. The upper-left corner of the plot is the “ideal” point, as the true positive rate equals 1 and the false positive rate equals 0. After constructing a ROC curve for each classifier, the area under the curve (AUC), defined as the area between the ROC curve and the X axis, is used for the prediction performance of each classifier. In this study, the ROC curve for each classifier is estimated using a 10-fold CV procedure and we compare the mean AUC occurring from each curve. A larger AUC usually means a better classifier.¹⁹ Other metrics for evaluating the performance of a classifier are precision, sensitivity and accuracy. Precision is the ability of the classifier to not label a sample that is negative as positive. As mentioned before, sensitivity equals to the proportion of positive samples that are correctly predicted as positive and, finally, accuracy is the number of correct predictions made divided by the total number of predictions made.²⁰

Feature selection

Feature selection can prove to be very important, as it can reveal subsets of informative genes that can discriminate schizophrenics from healthy control subjects. There are three main feature selection methods: filter, wrapper, and embedded methods. Filter methods filter out features that, based on statistical methods, are not informative. Filter feature selection is performed before applying classification (e.g. Fisher criterion score). Wrapper methods (e.g. stepwise forward selection and stepwise backward selection) search for optimal feature subsets, and utilize a classifier in order to evaluate the predictive power of the feature subsets. Compared to the filter methods, wrapper methods are usually more computationally demanding; but, they also provide more accurate results.²¹ The embedded methods select features while building a model. Embedded techniques are more computationally efficient than wrapper methods. An example of embedded methods is support vector recursive feature elimination (SVM-RFE), which is also used in this study. SVM-RFE is based on an iterative method of setting aside the feature with the lowest weight for each prediction method, until the optimal subset of genes is left.²²

The aim of this study is to test if the microarray gene expression data from a postmortem brain dataset contain enough information for the classification of SZ. For this reason several classification algorithms have been tested and their performance has been evaluated.

Data preprocessing and analysis

A dataset that includes brain postmortem gene expression data of 28 schizophrenic and 23 healthy control subjects, derived from Broadmann area 10 (anterior prefrontal cortex), accessible at NCBI GEO database²³ with the accession number GSE 17612, was analyzed using the Bioconductor package ‘affy’²⁴ through the R programming system.²⁵ Gene expression profiles were generated using the Affymetrix HG-U133 Plus 2.0 GeneChip. In this study, the robust multi-array analysis (RMA) method was used, which performs background correction on the Affymetrix perfect match data, applies quantile normalization and then performs summarization of the probe set information using median polish.²⁶ The limma (moderated t-test) Bioconductor package of R has been used towards the identification of differentially expressed genes among the two classes.²⁷ Transcripts were characterized as differentially expressed if their unadjusted p-value was less than 0.01. The differentially expressed genes were used as an input for the pathway analysis and gene prioritization as well as for the classification task.

Pathway analysis and gene prioritization

The differentially expressed genes were imported into the Bioinfominer web tool (available online: www.bioinfominer.com) for functional analysis based on established statistical tests and using different ontology databases, namely Gene Ontology (GO),²⁸ Reactome,²⁹ Human Phenotype Ontology³⁰ and MGI Mammalian Phenotype Ontology.³¹ In this way, significant biological mechanisms associated to the input data were revealed. The next part of the analysis include the identification and the prioritization of master regulatory genes, which represent hub nodes in the GO tree structure. These genes play a central role, as they are related to many distinct, cross-talking GO terms.³²

Classification algorithms and parameter optimization

In this study, the following classification techniques for the discrimination of the two classes (SZ and healthy controls) have been used: SVM, Extra Trees, RF, AdaBoost, and kNN classification algorithms. The classification algorithms come with a set of parameters. In this study, the parameters of the utilized classifiers were optimized with a CV grid search. Using this exhaustive search for each classifier, this method selects those parameters that maximize the mean AUC score of the CV.³³ For all the classification models developed in this paper, parameter optimization has been performed. All of the machine learning methods were implemented in scikit-learn.³⁴

Feature selection

The differentially expressed genes were used as the dataset for SVM-RFE method, in order to filter out the optimum informative feature set.³⁵ Generally, SVM-RFE selects the minimum informative subset of features that separates classes, by progressively removing features that are not informative. This procedure has many rounds. At each round one gene is eliminated and an SVM classifier is trained based on the rest of the genes. That procedure is recursively repeated on the pruned sets until the number of features that present the best performance according to CV is reached.³⁶ The reason for using SVM-RFE is that we aim at developing a sensitive and specific classification algorithm, based on realistic clinical biomarkers, assaying a small number of genes.³⁷

Data collection and classification of the independent test cohort

The second dataset (NCBI GEO accession number: GSE 21935) was used as an independent group of samples in order to examine if the final genes occurring from the feature selection can be used as biomarkers in SZ. For this reason, the 21-gene model was used as an input for testing if those genes can discriminate SZ samples from healthy control samples on this independent dataset. The classification task was applied on the normalized gene expression values of the dataset, also resulting from RMA. The dataset included samples from the Brodmann Area 22 (superior temporal cortex) of 23 schizophrenic patients and 19 healthy controls.

Model evaluation

ROC curve was mainly used in this study as a metric to evaluate the output quality of each classification model, created from the 10-fold CV. Each of the 10 different splits of the dataset generated by the 10-fold CV results in a curve. Taking all of these curves, the mean AUC of each classifier is calculated. A classifier with larger mean AUC is considered to have better performance. Other metrics were also used as evaluation criteria for the performance of the classifiers, including accuracy, precision, and sensitivity.

This study revealed 164 genes that were statistically significant. Furthermore, among the differentially expressed genes, CCL3, S100A8, SYK and VEGFA have been previously implicated in SZ and other psychiatric diseases. The main identified statistically significant ontological terms of interest that have been previously related to SZ are immune-related mechanisms. Other interesting mechanisms have also been found to be overrepresented, such as central nervous system synaptic transmission, integrin mediated signaling and retinoic acid signaling. In this study, the importance of the integrated feature selection and classification algorithm for the prediction of classes and for the identification of significant genes have been revealed once more.⁵² In summary, RF after the feature selection method of SVM-RFE outperformed the other tested classification methods with an AUC score of 0.98. The feature selection resulted to 21 genes that could discriminate schizophrenic and healthy control samples in two different independent datasets of postmortem brain samples obtained from two different brain regions.

None.

Author declares that there are no conflicts of interest.

1. Kahn RS, Sommer IE, Murray RM, et al. Schizophrenia. Nature Reviews Disease Primers. 2015;1:15067.
2. Middleton FA, Mirnics K, Pierri JN, et al. Gene expression profiling reveals alterations of specific metabolic pathways in schizophrenia. J Neurosci. 2002;22(7):2718‒2729.
3. Iwamoto K, Kato T. Gene expression profiling in schizophrenia and related mental disorders. Neuroscientist. 2006;12(4):349‒361.
4. Dean B, Keriakous D, Scarr E, et al. Gene expression profiling in Brodmann's area 46 from subjects with schizophrenia. Aust N Z J Psychiatry. 2007;41(4): 308‒320.
5. Cattane N, Minelli A, Milanesi E, et al. Altered gene expression in schizophrenia: findings from transcriptional signatures in fibroblasts and blood. PLoS One. 2015;10(2):e0116686.
6. Kotsiantis SB. Supervised Machine Learning: A Review of Classification Techniques. Real Word AI Systems with Applications in eHealth, HCI, Information Retrieval and Pervasive Technologies. Emerging Artificial Intelligence Applications in Computer Engineering, Greece. 2007.
7. Takahashi M, Hayashi H, Watanabe Y, et al. Diagnostic classification of schizophrenia by neural network analysis of blood‒based gene expression signatures. Schizophr Res. 2010;119(1‒3):210‒218.
8. Tarca AL, Romero R, Draghici S. Analysis of microarray experiments of gene expression profiling. Am J Obstet Gynecol. 2006;195(2):373‒388.
9. Struyf J, Dobrin S, Page D. Combining gene expression, demographic and clinical data in modeling disease: a case study of bipolar disorder and schizophrenia. BMC Genomics. 2008;9:531.
10. Brown MP, Grundy WN, Lin D, et al. Knowledge‒based analysis of microarray gene expression data by using support vector machines. Proc Natl Acad Sci USA. 2000;97(1):262‒267.
11. Geurts P, Ernst D, Wehenkel L. Extremely randomized trees. Machine Learning. 2006;63(1):3‒42.
12. Díaz‒Uriarte R, Alvarez de Andrés S. Gene selection and classification of microarray data using random forest. BMC Bioinformatics. 2006;7(3):1‒13.
13. Sarkar M, Leong TY. Application of K‒nearest neighbours algorithm on breast cancer diagnosis problem. Proc AMIA Symp. 2000;759‒63.
14. Weinberger KQ, Saul LK. Distance Metric Learning for Large Margin nearest Neighbour Classification. Journal of Machine Learning Research. 2009;10:207‒244.
15. Asyali MH, Colak D, Demirkaya O, et al. Gene Expression Profile Classification: A Review. Current Bioinformatics. 2006;1(1):55‒73.
16. Mozos OM, Stachniss C, Burgard W. Supervised Learning of Places from Range Data using Adaboost. IEEE International Conference on Robotics and Automation. 2005;1730‒1735.
17. Hsu CW, Chang CC, Lin CJ. A Practical Guide to Support Vector Classification. Department of Computer Science, National Taiwan University. 2003;1‒16.
18. Bradley AP. The use of the area under the Roc curve in the evaluation of machine learning algorithms. Pattern Recognition. 1997;30(7):1145‒1159.
19. Hajian Tilaki K. Receiver Operating Characteristic (ROC) Curve Analysis for Medical Diagnostic Test Evaluation. Caspian J Intern Med. 2013;4(2):627‒635.
20. Sokolova M, Lapalme G. A systematic analysis of performance measures for classification tasks. Information Processing & Management. 2009;45(4):427‒437.
21. Saeys Y, Inza I, Larranaga P. A review of feature selection techniques in bioinformatics. Bioinformatics. 2007;23(19):2507‒2517.
22. Lin X, Yang F, Zhou L, et al. A support vector machine‒recursive feature elimination feature selection method based on artificial contrast variables and mutual information. J Chromatogr B Analyt Technol Biomed Life Sci. 2012;910:149‒155.
23. Clough E, Barrett T. The Gene Expression Omnibus Database. Methods Mol Biol. 2016;1418:93‒110.
24. Gentleman RC, Carey VJ, Bates DM, et al. Bioconductor: open software development for computational biology and bioinformatics. Genome Biol. 2004;5(10):R80.
25. Team RDC. A language and environment for statistical computing. Foundation for Statistical Computing: Vienna, Austria. 2010
26. Irizarry RA, Hobbs B, Collin F, et al. Exploration, normalization, and summaries of high density oligonucleotide array probe level data. Biostatistics. 2003;4(2):249‒264.
27. Ritchie ME, Phipson B, Wu D, et al. Limma powers differential expression analyses for RNA‒sequencing and microarray studies. Nucleic Acids Res. 2015;43(7):e47.
28. Ashburner M, Ball CA, Blake JA, et al. Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet. 2000;25(1):25‒29.
29. D'Eustachio P. Reactome knowledgebase of human biological pathways and processes. Methods Mol Biol. 2011;694:49‒61.
30. Groza T, Kohler S, Moldenhauer D, et al. The Human Phenotype Ontology: Semantic Unification of Common and Rare Disease. Am J Hum Genet. 2015;97(1):111‒124.
31. Smith CL, Eppig JT. The Mammalian Phenotype Ontology as a unifying standard for experimental and high‒throughput phenotyping data. Mamm Genome. 2012;23(9‒10):653‒668.
32. Koutsandreas T, Pilalis E, Vlachavas EI, et al. Making sense of the biological complexity through the platform‒driven unification of the analytical and visualization tasks. IEEE 15th International Conference on Bioinformatics and Bioengineering. 2015;1‒6.
33. Bergstra J, Bengio Y. Random search for hyper‒parameter optimization. Journal of Machine Learning Research. 2012;13:281‒305.
34. Pedregosa F, Varoquaux G, Gramfort A, et al. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research. 2011;12:2825‒2830.
35. Maldonado S, Weber R. A wrapper method for feature selection using Support Vector Machines. Information Sciences. 2009;179(13):2208‒2217.
36. Guyon I, Weston J, Barnhill S, et al. Gene Selection for Cancer Classification using Support Vector Machines. Mach Learn. 2002;46:389‒422.
37. Clelland CL, Read LL, Panek LJ, Nadrich RH, et al. Utilization of never‒medicated bipolar disorder patients towards development and validation of a peripheral biomarker profile. PLoS One. 2013;8(6):e69082.
38. Berridge MJ. Calcium signalling and psychiatric disease: bipolar disorder and schizophrenia. Cell Tissue Res. 2014;357(2):477‒492.
39. Lidow MS. Calcium signaling dysfunction in schizophrenia: a unifying approach. Brain Res Brain Res Rev. 2003;43(1):70‒84.
40. Huang KC, Yang KC, Lin H, et al. Analysis of schizophrenia and hepatocellular carcinoma genetic network with corresponding modularity and pathways: novel insights to the immune system. BMC Genomics. 2013;14 Suppl:5:S10.
41. Gardiner EJ, Cairns MJ, Liu B, et al. Gene expression analysis reveals schizophrenia‒associated dysregulation of immune pathways in peripheral blood mononuclear cells. J Psychiatr Res. 2013;47(4):425‒437.
42. Rao JS, Kim HW, Harry GJ, et al. Increased neuroinflammatory and arachidonic acid cascade markers, and reduced synaptic proteins, in the postmortem frontal cortex from schizophrenia patients. Schizophr Res. 2013;147(1):24‒31.
43. Faludi G, Mirnics K. Synaptic changes in the brain of subjects with schizophrenia. Int J Dev Neurosci. 2011;29(3):305‒309.
44. Ranjan A, Gupta P, Srivastava SK. Penfluridol: An Antipsychotic Agent Suppresses Metastatic Tumor Growth in Triple‒Negative Breast Cancer by Inhibiting Integrin Signaling Axis. Cancer Res. 2016;76(4):877‒890.
45. Liu B, Cui Q, Jiang T, et al. A combinational feature selection and ensemble neural network method for classification of gene expression data. BMC Bioinformatics. 2004;5:136.
46. Xu J, Sun J, Chen J, et al. RNA‒Seq analysis implicates dysregulation of the immune system in schizophrenia. BMC Genomics. 2012;13 Suppl 8: S2.
47. Perez-Santiago J, Diez Alarcia R, Callado LF, et al. A combined analysis of microarray gene expression studies of the human prefrontal cortex identifies genes implicated in schizophrenia. J Psychiatr Res. 2012;46(11):1464‒1474.
48. Stuart MJ, Baune BT. Chemokines and chemokine receptors in mood disorders, schizophrenia, and cognitive impairment: a systematic review of biomarker studies. Neurosci Biobehav Rev. 2014;42:93‒115.
49. Rodier F, Coppe JP, Patil CK, et al. Persistent DNA damage signalling triggers senescence‒associated inflammatory cytokine secretion. Nat Cell Biol. 2009;11(8):973‒979.
50. Papanastasiou E, Gaughran F, Smith S. Schizophrenia as segmental progeria. Journal of the Royal Society of Medicine. 2011;104(11):475‒484.
51. Gao K, Wang Q, Zhang Y, et al. Association study of VEGFA polymorphisms with schizophrenia in Han Chinese population. Neurosci Lett. 2015;590:121‒125.
52. Pirooznia M, Yang JY, Yang MQ, et al. A comparative study of different machine learning methods on microarray gene expression data. BMC Genomics. 2008;9 Suppl(1):S13.