This is a repository with Jupyter notebooks describing how we pre-trained the model, and how we apply the model for fine-tuning.
-
What dataset did you use? This is described in our paper, but briefly, we used a section of the Observed Antibody Space (OAS) database (Kovaltsuk et al., 2018) for pre-training, and a snapshot of SAbDab (Dunbar et al., 2014) as of 26 August, 2021. We've included small snippets of the OAS database that we used for pre-training, and the paratope prediction datasets under
assets
. -
Why HuggingFace? We felt that the maturity of the library and its straight-forward API were key advantages. Not to mention it fits really well with cloud compute architectures like AWS.