Predicting SSRIs Resistance In Patients With Recurrent Major Depressive Disorder:A Study Of Prediction Models Based On Support Vector Machine

ObjectiveRecurrent depressive patients who had been treated with selective serotonin reuptake inhibitors were used as the research subjects,tagSNPs of AC-cAMP pathway related genes mediated by 5-HT and their clinical features were used to construct databases respectively.Through data mining,we searched for the clinical and genetic features of SSRIs resistance to explore the method of early prediction of SSRIs resistant patients,in order to provide evidence for individualized precision medicine within major depressive disorder.Methods1.Construction of Clinical Features Database:857 patients with recurrent major depressive disorder were followed up to establish a database of clinical characteristics includingocio-demographic,clinical features and SSRIs treatment features during the first course treatment.2.TagSNPs Gene Database:We applied mass spectrometry analysis technique to complete the detection of 34 tagSNPs around 9 target genes related to 5-HT signaling pathway,such as HTR1A,HTR2A,CREB1,BDNF,ADCY7,ADCY9,ADCY3,NOS1 and PDE4A,and tagSNPs databases were constructed.3.857 subjects with the diagnosis of recurrent Major Depressive Episode according to DSM-IV,302 SSRIs resistance and 304 SSRIs not-resistance patients were selected according to SSRIs treatment outcome(the reduction rate of HDRS-24score)in order to use to classify training samples and test samples for case-control study and machine learning.4.SPSS 21.0 was used for general data processing and analysis.correlation analysis was used for primary screening of tagSNPs.5.SVM prediction model was established using support vector machine(SVM).Results1.Comparison of clinical variables between SSRIs-R group and SSRIs-NR groupComparing clinical features between 302 SSRIs-R and 304 SSRIs-NR patients groups,we found that there were significant differences in 12 clinical features between SSRIs-R and SSRIs-NR groups,including psychomotor retardation(c~2=11.068,p=0.001),psychotic features(c~2=13.795,p=0.000),suicidality(c~2=9.559,p=0.002),weight loss(c~2=9.145,p=0.002),SSRIs average tolerance dose(c~2=10.049,p=0.002),first-course treatment response(c~2=25.343,p=0.000),sleep disturbance(c~2=8.386,p=0.004),residual symptom(c~2=9.650,p=0.002),personality(c~2=18.091,p=0.000),age of onset(p=0.048),frequency of episode(p=0.031),duration(p=0.014).However,there was no significant difference in other variables(P>0.05).2.TagSNPs screeningIn our study,34 tagSNPs related to 5-HT signaling pathway were detected.The above 30 tagSNPs genotypes and alleles were tested by Hardy-Weinberg genetic balance(H-W balance)test.All the above 30 sites complied with the H-W test(P>0.05).The remaining 4 tagSNPs were not detected gene polymorphism including rs2059336 in TT genotype,rs143117860 in CC genotype,rs2551926 in GG or CC genotype,rs889895 in TT genotype.3.Comparison of the frequency distribution of tagSNPs alleles and genotypes between SSRIs-R group and SSRIs-NR groupComparison of the frequency distribution of 30 tagSNPs alleles and genotypes between SSRIs-R group and SSRIs-NR group,we found that the genotypes of 4tagSNPs,including CREB1rs2551645?rs4675690;BDNFrs18035210?rs7124442,were significantly different between the two groups(P<0.05).There was no significant difference in genotype and allele frequency distribution of the remaining26 tagSNPs between the two groups(P>0.05).4.Parameters optimization302 patients with SSRIs-R and 304 patients with SSRIs-NR were mixed and divided into training samples and test samples according to the scale of 5:1.There were 505 training samples,including 254 SSRIs-R patients(254/505,50.3%).There were 101 test samples,including 48 SSRIs-R patients(48/101,47.5%).In this study,we applied multiple cross-validation and grid search to kernel parameters C and?.The range of kernel parameters was in the region of=-3~15,=-15~13.Accuracy of cross-validation was 59.60%to 90.38%.5.SSRIs-R Predictive Model ScreeningAfter the 12 primary predictive variables were randomly combined 11 queues and4083 combinations were formed.According to prediction accuracy,sensitivity and specificity of each combination in each queue,optimal prediction models were filtered gradually.Using the filtering criteria in which the accuracy,sensitivity and specificity were all greater than 60%,10 predicting models were finally selected,named SSRIs-R-PM 1 to 10 respectively.6.The influence of tagSNPs on SSRIs-R predictive modelsIn our study,4 tagSNPs(CREB1rs2551645,rs4675690;BDNFrs10835210,rs7124442)of two target genes CREB1 and BDNF,which may be related to the SSRIs antidepressant treatment outcome for recurrent major depressive disorder were included.We found that mutations of cAMP responsive element binding protein 1(CREB1)and brain-derived neurotrophic factor(BDNF)in tagSNPs increased the accuracy of SSRI-R predictive models to a certain degree,accuracy of SSRI-R predictive models 8 could reach 87.5%.Conclusions1.Differences in clinical features between SSRIs-R and SSRIs-NR groups suggest that individuals may be heterogeneous in etiology,and these clinical features may also indicate SSRIs-R to some extent in the early stage.2.The genetic polymorphisms of CREB1 and BDNF may be associated with SSRIs-R in patients with recurrent depressive disorder.The combination of CREB1and BDNF mutations increases the risk of SSRIs resistance and suggests that 5-HT may be affected by gene polymorphism at the level of the second signal transduction pathway,leading to differences in clinical efficacy of SSRIs.3.The accuracy of the prediction model trained by machine learning method can reach 87.5%.It suggests that machine learning method can be used to establish mathematical models to predict whether an individual is resistant to SSRIs early.The SSRIs resistance of an individual can be predicted from a clinical or biological point of view,which provides the basis for individual treatment,and it can be verified the clinical significance of biomarker.