Gaussian process assumes the target data to be distributed as a multivariate Gaussian, with Gaussian noise.
This is often an unreasonable assumption to make for the raw data
Standard practice is to apply a preprocessing transformation in an ad-hoc manner preceding the modelling with the GP