This depends on which program you used to generate the Ramachandran plot. Different programs have different distributions and an outlier in one could be okay in another. I like MolProbity to find outliers and WHAT_CHECK for an overall score. 4% outliers seems a lot, so you should try to fix these by peptide flipping and such.
Perhaps you can give a more context. Is this an X-ray structure model or from another experimental or computational technique? If you model comes from X-ray crystallography, you can try the PDB_REDO server to help you on your way to a better Ramachandran plot.
Thank you for your response. I am using the RAMPAGE server from Cambridge to get my plot. My protein is from a X-Ray from the RCSB protein database and I am cleaning up the model by using Modloop.
Are you aware of any papers that have modeled proteins of this size?
Not really, I guess you could start with papers about model assessment in the CASP competitions. By the way, you can try to see if the PDB_REDO version of the PDB entry is better at http://www.cmbi.ru.nl/pdb_redo/
The Ramachandran Z-score for the model is the score to look at. Higher is better. If it is not good I can have a look at this entry if you give me the PDBid (off-list).