Thank you for your interest in DeepPurpose! As more and more models are coming up, we want to include as much as the models and their pretrained models in our framework. Here we provide step-by-step instructions to do that:
For any dataset, we expect each drug is associated with SMILES and each protein with amino acid sequence. However, as different encoders expect different input to the model (e.g., MPNN expects mol graph), we need to first transform it to the expected format. To do that, in the utils.py file, define a new function smiles2xxx or target2xxx which taks a input SMILES/sequence and outputs the encoding format for that single input.
Then, in the encode_drug or encode_protein functions, include a elif statement to transform all of the data points in the input dataframe using just defined smiles2xxx or target2xxx.
For special input formats such as further transformation on the fly, please add a elif statement to the data_process_loader, data_process_DDI_loader, data_process_PPI_loader, data_process_loader_Protein_Prediction, data_process_loader_Protein_Prediction. You can refer to the examples for CNN in these functions.
Now, in the generate_config file, add an elif statement to include all important encoder parameters (e.g. input dimension, model dim and etc.). If your encoder has new parameters that you want the users to specify in the model_initialize function, you should also add in the function parameter space. If so, please specify the default values.
In the encoders.py, define the encoder models. The input of the __init__ in default should contain encoding, which is either 'drug' or 'protein', and **config, which includes all the model parameters defined by users. For the forward function, we expect to input one feature matrix and output the hidden embedding.
Finally, we need to modify the training wrappers. Every file has similar structures so we will talk about one file and the rest should follow. In the main class __init__ function, include an elif statement to define the model based on the definitions in encoders.py.
That's it! You have successfully included your model in DeepPurpose!
Before you create a pull request, please also test it locally and send kexinhuang@hsph.harvard.edu a test case. Then, you are good to go!