Bias, randomization, and ovarian proteomic data: a reply to "producers and consumers"

Abstract

Proteomic patterns derived from mass spectrometry have recently been put forth as potential biomarkers for the early diagnosis of cancer. This approach has generated much excitement, particularly as initial results reported on SELDI profiling of serum suggested that near perfect sensitivity and specificity could be achieved in diagnosing ovarian cancer. However, more recent reports have suggested that much of the observed structure could be due to the presence of experimental bias. A rebuttal to the findings of bias, subtitled "Producers and Consumers", lists several objections. In this paper, we attempt to address these objections. While we continue to find evidence of experimental bias, we emphasize that the problems found are associated with experimental design and processing, and can be avoided in future studies.

Keywords: calibration; experimental design; mass spectrometry; proteomics; serum profiling.

Figure 1

(a) The average cancer spectra from DS2 and DS3, with the location of the maximum peak from Conrads et al. (2003) shown. The posted spectra appear offset. (b) The corresponding average spectra after using the labeled peaks in the Conrads et al. (2003) figure to recalibrate the spectra. Agreement between DS2 and DS3 is now good throughout the region bracketed by calibrants.

Figure 2

A heat map of the Qstar spectra we have available, sorted by file name, in the vicinity of m/z 8602. This m/z value is identified on the NCI/FDA website as useful for separating healthy women from ovarian cancer patients, and this separation is visible. However, roughly 80 Da below, there is a peak that serves to separate the healthy women and the first half of the ovarian cancer patients from the second half of the ovarian cancer patients.