You Can Learn a Lot by Lookin': Insights into Affymetrix Data

• Activity Number: 415
• Type: Contributed
• Date/Time: Thursday, August 12, 2004 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Terry M. Therneau*+ and Karla Ballman and Ann Oberg
• Companies: Mayo Clinic and Mayo Clinic and Mayo Clinic
• Keywords: microarray ; Affymetrix ; backround binding

Abstract

The above advice from Yogi Berra has proven useful to our group in many a medical research study. Feller suggests that nearly all dose response curves are well fit by a logistic curve with x=log(dose) and y=log(response). Guided by this, we looked at the individual data for each probe from three publicly available spike-in experiments: (a) the Gene Logic U95A data containing 3 subexperiments with 200, 220, and 220 probes, respectively; (b) the Affymetrix U95A data with 256 probes; and (c) the Affymetrix U133A data with 498 total probes. This generated 1,394 plots on 90 pages (one page/gene), using the raw data from each experiment (no normalization or backround correction). To each plot we fit a simple model y= a + b*f(c(x-d)) where f is the logistic function. Separate curves for PM and MM are on each plot, but differ only in the location of their inflection point d. (1) The logistic fits were excellent, with an R^2 > 0.95. (2) Residuals are equivariant across the range of x. (3) Both PM and MM probes show strong specific binding. (4) 10% of the probes show evidence of possible cross-binding. (5) 5-10% of the probes show no activity above backround.