Bayesian Methods for Transcript Level Estimation from Noisy Array Measurements

Gene arrays demonstrate a promising ability to characterize expression levels across the entire genome, but they suffer from significant levels of measurement noise. We present a statistical technique to estimate transcript levels or transcript level ratios from one or more gene array experiments, incorporating a model of measurement noise and prior information about biological expression levels. The Bayesian Estimation of Array Measurements (BEAM) technique provides a principled method to handle issues currently addressed through rough and varying heuristics, such as identification of changes in expression level, combination of repeated measurements, and rectification of negative measurements. More importantly, the BEAM technique produces associated measures of estimation uncertainty (e.g., p-values) which serve to determine the statistical significance of reported results. While our method applies to any gene array technology, we illustrate it using a detailed noise model that we develop for Affymetrix yeast chips. The BEAM technique allows for the design of experiments that maximize the useful information derived from a minimum number of chips. Further, it can be used to extract additional and more statistically rigorous conclusions from existing data.