buiphuong33/graphepitope ? reverse-engineered prompt

I want a local Python tool that reproduces GraphBepi for B cell epitope prediction from proteins.

It should work in two simple ways. If I already have a protein structure file, I want to run a prediction on that and get the results saved in an output folder. If I only have a FASTA sequence, I want it to use the built in structure prediction path so I can still run the same kind of epitope prediction from the sequence alone.

Please also make sure the training workflow works, including building the dataset first and then training the model on the provided BCE_633 data with sensible defaults. I want the repo set up so someone on Linux with a GPU can follow the commands from the README without guessing what is missing, including the DSSP requirement in the mkdssp folder.

Keep the project runnable end to end, with clear commands, expected inputs and outputs, and any small fixes needed so prediction and training both actually work. If you need to, check the current docs online.