Preliminary Investigation Of Artificial Neural Network Model For Renal Cell Carcinoma Diagnosis Based On Metabonomics

Objective:To explore the potential diagnosis metabolites,a nuclear magnetic resonance(~1HNMR)based metabonomics technology was employed to analyze the body fluids(blood and urine)from the renal cell carcinoma patients and healthy volunteers.Then,the potential biomarkers associated with renal cell carcinoma diagnosis were identified and confirmed..On this basis,the contents of some potential biomarkers in biofluild were measured by a novelly developed method.After optimizing the artifical neural network(ANN)parameters,the ANN model for renal cell carcinoma diagnosis was established as the quantivative contents of potential biomarkes as input parameters.Methods:1.The ~1HNMR spectra of blood and urine from patients in renal cell carcinoma and controls were obtainedand analyzed by some multivariate statistical analysis methods(SIMCA-P).The important peaks in ~1HNMR spectra for classification were determined by variable importance value(VIP)of orthogonal partial least-squares discriminant analysis(OPLS-DA)model and the results of variance analysis.Then,the potential biomarkers associated with diagnosis of renal cell carcinoma were identified and confirmed by TCOSY,HSQC,STCOSY and standards of metabolites.2.The contents of metabolites in uric acid pathway was determined by a novelly developed HPLC-UV method.The analytes in biofluild were separated by a ZIC-HILIC column.The cetonitrile-5m M ammonium dihydrogen phosphate and0.1%phosphoric acid was the mobile phase with a gradient elution process.Flow speed was 0.8ml/min.The column wasmaintained at 38?C and the injection volume of samples was 20 ul.The detection wavelength was set at 215 nm for Cr,270 nm for UA and hypoxanthine,and 250 nm for xanthine.3.In order to eliminate the differences in sampling,the total samples for establishing a ANN model waer divided randomly more than 5000 times into training samples(90%)and predictivesamples(10%).For example,t 82 blood samples were randomly assigned to two groups,one group(74 samples)as training samples and the other aspredictive samples.It means that each model was calculated more than5000 times for different sample combinations.The predictive accuracy of intra-group was expressed by the average percentages of correct classification by the ANN models for the training samples(90%of total smaples).The predictive accuracy of inter-group was that for the predictive samples(10%of total smaples).The average predictive accuracy of the model whichwas calculated by the mean of the predictive accuracy of intra-group and that of inter-group was used as evaluation index to optimize parameters of ANN model and input parameters.Receiver operating characteristic curve(ROC curve)was also employed to analyze the data.the average prediction accuracy ofResults:1.The results showed that different metabolic phenotypes of ~1HNMR spectra in biofluids are significantly different from patients in renal cell carcinoma and controls.The related metabolites causing metabolic phenotype differences were identified and confirmed.Compared to the control group,increasedblood concentrations of leucine,isoleucine,valine,alanine,acetic acid,succinic acid,choline,betaine,taurine,creatinine,α-glucose,xanthine and decreased that of lipid were observed.In urine samples,the levels of 3-hydroxybutyric acid,acetic acid,α-glucoseand creatine appears to be elevated,as well as those of lysine,citric acid,hypoxanthine,hippuric acid,xanthine,trimethylamine oxide reduced.2.We developed a HILIC-HPLC-UV method for simultaneous determination of xanthine(XAT),hypoxanthine(HYP),uric acid(UA)and Creatinine(Cr)in human bio-fluid.When compared with the normal group,our results showed notable reductions of uric acid concentration and UA/HYP,XAT/HYP,UA/Cr ratio in blood of renal cancer group.Xanthine,uric acid and creatinine concentration and UA/HYP,XAT/HYP,UA/Cr ratios in the random urine samples of renal cell carcinoma patients was lower.3.ANN model was established for metabolite combinationsobtained from~1HNMR based metabonomics such as 3-hydroxy butyric acid and glutamate and HILIC-HPLC-UV method such as xanthine,uric acid and creatinine.The results showed the average predictive accuracy was above 95%by developed ANN model with BP algorithm when combination of the blood concentratons of 3-hydroxy butyric acid and glutamate was the input parameter.The average predictive accuracy of the model was up to 90%when combination of HYP/Cr and UA/Cr as input parameters..After data evaluation through ROC curve,it was found that the diagnostic value for quantitative data in blood from ~1HNMR and in uirine from uric acid metabolic metabolite is higher.Conclusion:1.Our resultsshowed that the differences between renal cell carcinoma and normal group were mainly related to energy metabolism,uric acid metabolism and amino acid metabolism pathway based on ~1HNMR based metabonomics approach.2.Our results showed that lower concentrations of uric both in blood and urine in renal cell carcinoma patients,whearas,the concentration of hypoxanthine appeared to be increased in blood.Therefore,the activity of purine oxidase in liver and kidney was possibly inhibited,which resulted in reduction of uric acid synthesis.3.ANN models for diagnosis of renal cell carcinoma were preliminarily established by optimizing input parameters in this study.The results of ROC curve evaluation verified the good prediction performance of ANN model.