Well, two options:
- Your channel doesn't only add noise, but has some nonlinearity, or the noise is not just additive, so: your model is wrong, or
- your denoiser is not perfect and doesn't only remove noise, but maybe adds some interference at some frequencies. So: your denoiser can't denoise perfectly, and has some unexpected side effects.
Option 2 seems more likely. Also note that it is mathematically in general impossible to remove all noise. From an information-theory point of view, whenever you add noise with a PDF that has support wider than half the minimum distance between signal points, you simply eradicate some information. It's mathematically impossible to perfectly denoise.
Note that learned denoisers are trained towards one objective only – minimizing some objective function ("loss function"). I'd expect that function to be something like a euclidean distance between the denoised and the transmit sequence at the symbol instants – but that doesn't mean it's the only possible loss function for this task, nor does it mean your denoiser actually converged against a maximum estimator that's uncorrelated to your data.
First of all, unless you can prove the opposite, you have converged (best case) to a local optimum, not to the global optimum, and:
Denoising might, depending on your signal level, simply have learned to e.g. set specific samples to specific values (e.g. in a filtered/shaped pulse train, you'd expect "0" in the middle between symbol instants, on average. But learning to always set that value to zero, even when the filter doesn't fulfill Nyquist criteria, doesn't make the loss function I described above any worse, and might train a bit of a zero-forcing equalizer for the transmit filter – which might be amplifying your signal harmonics. So, I'd even call this a bit expected!