whether to redesign low confidence aas

[18217265596]

I tried to use mpnn to design sequence for rf-difussion generated backbones
and i noticed there a lot of A's in loops and E/K in non-interface free area
i understand the reason why
but such sequences were scored by AF2 to result in high plddt and very low pae
which means it should fold and maybe bind to my target
in this situation should i redesign those A's and K/E area to a more "reasonable" sequence (only to me,not to af2)
such sequence maybe like DLYHPSVAL and suddenly KKEEEEKE or suddenly AAAVAAAAD