Urine was collected from twenty consecutive patients with SLE at the time of renal biopsy. There were five Caucasians, 15 African-Americans, two males and 18 females. Mean age was 27 ± 8 years. ISN/RPS class of lupus nephritis and indices of activity and chronicity were determined by a renal pathologist. Nine patients had histologic characteristics of more than one class of lupus nephritis and were assigned to more than one ISN/RPS class.
A representative example of a two-dimensional gel of urine proteins from a patient with class V lupus nephritis is shown in . We aligned 213 protein spots across the gels. Analysis by artificial neural networks was done on protein abundance normalized by quantile. Output values were the binary value for each ISN/RPS class II, III, IV, and V and an ordinal value for activity and chronicity. ROC curves were derived for ISN/RPS classes. The area under the curve (AUC) for the ROC curve was greater than 0.85 for all classes. The r value for activity index was a relatively poor 0.77, but the r value for the chronicity index was much better at 0.87 (). ROC AUC and correlation coefficient are not reported for class II since there were only two positive diagnoses.
Two-dimensional gel separation of proteins from a patient with class V lupus nephritis
Statistics of sample classification by artificial neural networks
The original data set was analyzed by the trained artificial neural networks to determine its ability to predict the disease. For each gel, a prediction of presence or absence of each class of lupus nephritis was given. Sensitivity was 86% or greater for all classes. The sensitivity was lowest in class III disease, in which six of the seven cases were correctly identified. The specificity was 92% or greater for all classes except V, where only four of the eight patients who were negative for class V were correctly identified. Interestingly, all the false positive identifications of class V had class III or IV lupus present and were correctly identified as such. Clinically, a patient with both proliferative (class III or IV) and membranous (class V) would be treated for the more aggressive proliferative lesion, so the false positive identification of class V would not have affected the treatment of the patient.
In addition to being useful for identifying the ISN/RPS class of lupus nephritis, urine markers could be used to predict the duration and amount of renal injury from the disease. We have trained the artificial neural networks to correlate with the histologic score for chronicity and activity. A high degree of correlation was obtained for chronicity (r
= 0.87) (), and a lesser degree of correlation was obtained for the activity index (r
= 0.77). It is worth recalling that these values are cross-validated and that the artifical neural network classifiers selected correspond to the median performer of a set of artificial neural network models that rely on different resampled subsets of the available data [1
]. Therefore, the results obtained are representative to the same extent that the data set itself is sufficiently representative.
Predicted vs. observed values from a trained artificial neural network for chronicity index
Classification of patients was based on patterns of protein abundance. In order to derive a clinically useful test to predict class, activity, and chronicity of lupus nephritis, the identity of the proteins that provide the most sensitivity in the trained network needs to be determined. Both the amount of sensitivity for an individual protein in a given gel and the overall amount of sensitivity for the analysis was determined. Analysis of sensitivity for each of the outcomes was performed. Interestingly, most of the sensitivity was derived by a limited set of spots. lists in order the ten spot numbers or demographic factors that provided the most sensitivity. Spot numbers 5, 77, and 44 were near the top for amount of sensitivity provided for most of the analyses. None of the spots alone could differentiate between classes. Race, gender, or age were important in the analysis for several diagnoses. Total amount of sensitivity provided by the top ten variables for each diagnosis is shown at the bottom of the table. Using matrix-assisted laser desorption-ionization tandem mass spectrometry (MALDI-TOF-TOF) and informatic tools we have identified the following proteins that provide the highest sensitivity: spot 5, α-1 acid glycoprotein; spot 44, α1 microglobulin; spot 52, zinc α-2 glycoprotein; spot 53, zinc α-2 glycoprotein; spot 75, IgG κ light chain; and spot 77, α1 microglobulin.
Ranked sensitivities for diagnosis of lupus class, activity, and chronicity