Comparing biomarkers and proteomic fingerprints for classification studies

Abstract

Early disease detection is extremely important in the treatment and prognosis of many diseases, especially cancer. Often, proteomic fingerprints and a pattern recognition algorithm are used to classify the pathological condition of a given individual. It has been argued that accurate classification of the existing data implies an underlying biological significance. Two fingerprint-based classifiers, decision tree and medoid classification algorithm, and a biomarker-based classifier were examined using a published dataset of mass spectral peaks from 81 healthy individuals and 78 individuals with benign prostate hyperplasia (BPH). For all three methods, classifiers were constructed using the original data and the data after permuting the labels of the samples (BPH and healthy). The fingerprint-based classifiers produced accurate results for the original data, though the peaks used in a given classifier depended upon which samples were placed in the training set. Accurate results were also obtained for the dataset with permuted labels. In contrast, only three unique peaks were identified as putative biomarkers, producing a small number of reasonably accurate biomarker-based classifiers. The dataset with permuted labels was poorly classified. Since fingerprint-based classifiers accurately classified the dataset with permuted labels, the argument for biological significance from a fingerprint-based classifier must be questioned.