The relatively low signal-to-noise ratio prevalent in expression technology, typically in the range 2-5 when “signal” is defined as the median level of gene expression, makes quality metrics such as confidence intervals and statistical significance essential in interpreting the data. We present models for the several sources of noise which affect expression measurements, including cross-hybridisation and chip-to-chip variation, and discuss how the model parameters may be obtained on-the-fly, either from single-scan data or from replicate experiments. These noise models have been integrated into the PFOLD algorithm: this is a method, grounded in a Bayesian framework, which not only estimates fold-changes, but also assigns a statistical significance to the change through a P-value. We have found that the ability to make selections in the two-dimensional space of fold-changes/P-values (rather than fold-change alone) can greatly enhance the selectivity of detection. In turn, the PFOLD algorithm has been fully integrated into our suite of analysis tools called GECKO (Gene Expression Computation and Knowledge Organisation). Turning to the analysis of entire expression profiles, as obtained for instance from time-courses, we shall present results on cross-technology comparisons (e.g. Affymetrix vs. MD) and also on the biological reproducibility of expression profiles, including protocols for estimating false-positive rates in the presence of complex queries. We shall finally present some graphical methods for the display of expression profiles with moderate number of samples, including what we have called the “Gene Planet.”