In order to reduce artifacts (non-biological sample variation) from the data, we use a rescaling procedure called L1000 Invariant Set Scaling, or LISS, involving 80 control transcripts (8 each at 10 levels of low to high expression) that we empirically found to be invariant in expression across the DSGEO. The 80 genes are used to construct a calibration curve for each sample. Each curve is computed using the median expression of the 8 invariant genes at each of the 10 pre-defined invariant levels. We then loess-smooth the data and fit the following power law function using non-linear least squares regression:
