Experimental Design, Randomization, and Validation.

Abstract

Featured Article: Baggerly KA, Morris JS, Coombes KR. Reproducibility of SELDI-TOF protein patterns in serum: comparing datasets from different experiments. Bioinformatics 2004;20:777–85.2 Before our article discussed here appeared in 2004, one might have thought a golden age in proteomic diagnostics was at hand. The good news began in 2002 (1), when National Cancer Institute (NCI)3/Food and Drug Administration (FDA) researchers claimed to have processed high-throughput measurements of easy-to-get samples with machine learning algorithms to extract previously elusive diagnoses. The high-throughput measurements were from mass spectrometry, specifically surface-enhanced laser desorption and ionization (SELDI) assays. The samples were minimally processed serum. The machine learning algorithms were, broadly, black boxes taking peak intensities (nominally peptide abundances) and producing categorical calls: “disease,” “no disease,” or “something else.” The diagnosis was whether a woman had ovarian cancer. The numbers looked impressive. Starting with 216 samples (100 women with ovarian cancer, 100 healthy controls, and 16 women with …