“Over the past few years, dramatic claims have been made about how proteomic patterns derived from serum samples could be used to detect ovarian cancer and possibly other types as well.”

There are a few reasons. First, we looked at a diagnostic test of great clinical relevance. Second, we gave clear examples of how things could go wrong, supported by vivid pictures. Third, we suggested how these problems could be avoided in future studies. Fourth, we illustrated the need for the occasional use of "forensic bioinformatics," diagnosing features of the experiment on the basis of traces left in the raw data. Fifth, the results have been controversial, and have led to an ongoing debate. The most recent part of this debate is in Baggerly et al., J. Natl. Canc. Inst. 97:307-309, 2005.

Over the past few years, dramatic claims have been made about how proteomic patterns derived from serum samples could be used to detect ovarian cancer and possibly other types as well. These patterns are found by data mining of high-dimensional mass spectrometry profiles, and do not require the identification of the proteins involved. Since we don't have noninvasive tests at present, such assays would be of great clinical utility.

In reexamining the raw data, we found evidence that the differences found and attributed to disease status could be explained at least as well by systematic biases in sample handling and preprocessing. As an example, if cancer data are gathered on Monday and control data on Wednesday, then a drift of the values reported due to a gradual breakdown of the machine would give rise to machine differences that could be mistaken for biological differences. Our paper illustrated the types of problems that could occur, and pointed out the need for careful experimental design in the planning of such studies so that biases could be avoided (especially vital when black box approaches are used). It indicated the importance of having access to the raw data, and the need to do quality checking before data mining.