We only disclose the inference models. TransformerCPI2.0 is based on TransformerCPI whose codes are all released. The details of TransformerCPI2.0 are described in our paper https://doi.org/10.1038/s41467-023-39856-w which is now published on Nature communications. Trained models are available at present.

environment.yaml is the conda environment of this project.

predict.py makes the inference, the input are protein sequence and compound SMILES. featurizer.py tokenizes and encodes the protein sequence and compounds. mutation_analysis.py conducts drug mutation analysis to predict binding sites. substitution_analysis.py conducts substitution analysis.

Trained models is now available freely at https://drive.google.com/drive/folders/1X7i1eO-EykCQcvqMeWeB7QXT3E9eLG08?usp=sharing. The current open source version only aims to reproduce the results reported in the article, so the inference speed is limited.

python = 3.8.8

pytorch = 1.9

tape-proteins = 0.5

rdkit = 2021.03.5

numpy = 1.19.5

scikit-learn = 0.24.1