Here, the tar file contains only the LMDB files of the training dataset and the checkpoints for HuDiff-Ab and HuDiff-Nb. For the code and further details, please visit the HuDiff GitHub repository. Below are simple code templates for utilizing the checkpoint to humanize the sequences.
HuDiff-Ab Humanization
There are two ways to humanize an antibody, either by using a complex fasta file containing antigen-antibody sequences or by providing individual heavy and light chain sequences.
python antibody_scripts/sample_for_anti_cdr.py
--ckpt checkpoints/antibody/hudiffab.pt
--anti_complex_fasta data/fasta_file/7k9i.fasta
Keep in mind that you need to replace 7k9i.fasta.
HuDiff-Nb Humanization
Using the fasta file of the nanobody, our model can humanize it. If you require a higher degree of humanization, consider increasing the batch size or the number of samplings.
python nanobody_scripts/sample_for_nano_cdr.py
--ckpt checkpoints/nanobody/hudiffnb.pt
--nano_complex_fasta data/fasta_file/7x2l.fasta
--model finetune_vh
--inpaint_sample True