jourmore / NanoBERTa-ASP

Predicting Nanobody Binding Epitopes Based on a Pretrained RoBERTa-based Model

NanoBERTa-ASP

Predicting Nanobody Binding Epitopes Based on a Pretrained RoBERTa Model

Model Description

The NanoBERTa-ASP is based on the RoBERTa architecture.

Usage

Download data

-To download the required data from the respective databases:
  ·Download unpaired heavy chain data specific to human sources from OAS for pretraining.
  *https://opig.stats.ox.ac.uk/webapps/oas/oas_unpaired/
  ·Download PDB data(<3.0Å) from SAbDab for model fine-tuning.
  *https://opig.stats.ox.ac.uk/webapps/sabdab-sabpred/sabdab/search/

Data preprocessing

-To preprocessing the data，you can:
  ·Pretrain:
    Filter sequences that meet the criteria, example code:data-process/pretrain-data.py.
  ·Fine-tuning:
    Calculate the binding sites, example code:data-process/finetuning-data.py.

Training

  tokenizer：model/tokenizer
  ·Pretrain:
     example code:model/pre-train.py.
  ·Fine-tuning:
     example code:model/finetuning.py.
     The Fine-tuning dataset is provided in folder NanoBERTa-ASP/assets in parquet format, 
     you could open it by pandas package of Python.

Contact

For any questions or inquiries, please contact Shangru Li ([e-mail]) and [e-mail]