The absorption comes from impurities in the GG475 filters and the first and third quadrant are most strongly affected (see plot - the black curve shows the spectrum of a featureless star observed in Q1, the green shows the same star observed in Q3). For MOS data, you can correct these features to a large extent during the flux-calibration of your data using the parameter --resp_use_flat_sed=true (default for MR data) when determining the response curve.