Assessing the utility of SELDI-TOF and model averaging for serum proteomic biomarker discovery

The SELDI-TOF technique was used to profile serum proteins from Type 1 diabetes (T1D) patients and healthy autoantibody-negative (AbN) controls. Univariate and multivariate analyses were performed to identify putative biomarkers for T1D and to assess the reproducibility of the SELDI technique. We found 146 protein/peptide peaks (581 total peaks discovered) in human serum showing statistical differences in expression levels between T1D patients and controls, with 84% of these peaks showing technical replication. Because individual proteins did not offer great power for disease prediction, we used our model averaging approach that combines the information from multiple multivariate models to accurately classify T1D and control subjects (88.9% specificity and 90.0% sensitivity). Analyses of a test subset of the data showed less accuracy (82.8% specificity and 76.2% sensitivity), although the results are still positive. Unfortunately, no multivariate model could be replicated using the same samples. This first attempt of high throughput analyses of the human serum proteome in T1D patients suggests that model averaging is a viable method for developing biomarkers; however, the reproducibility of SELDI-TOF is currently not sufficient to be used for classification of complex diseases like T1D.