Open TangYiChing opened 1 month ago
Thanks for trying the pipeline but unfortunately, the official implementation only support optimizing CDR3. While you can hack it to optimize other CDRs, you would have to re-train the whole system to make it optimizing all CDRs because it wasn't how the model being trained.
I see.
By the way, could you describe each checkpoint in the reproduction folder? Most of them have a descriptive name, but some are not intuitive.
Some of the checkpoints were designed for specific experiments. For example in your snippet, the "reproduction/seq_design_fixed_pos/300000.pt" was used for sequence design. All the useful ckpts are listed in the README, which you can just follow the instructions for specific usage such as sequence design and optimization. In general, the ckpt that starts with seq_design means given a backbone, designing the CDR sequence.
Hello,
Could you suggest a way to optimize multiple CDRs? I've tried two options, but I need help to get them to work.