Predicting on new set of proteins with DPFunc trained model

[sophiachen1]

Hi @CSUBioGroup ,

Thank you so much for creating this excellent tool. I have a list of my own protein sequences which is about thousands of entries and I would like to predict on them using DPFunc trained model. I have 2 questions. First, I have downloaded the save_models from https://drive.google.com/file/d/1V0VTFTiB29ilbAIOZn0okBQWPlbOI3wN/view?usp=drive_link, is this model trained with PDB dataset or large-scale dataset? Second, my understanding is that I can provide my own protein data as the test set to predict, is that right? So do I need to create pid_list_file, pid_go_file, pid_pdb_file and interpro_file listed in yaml file for my proteins? Shall I use this notebook https://github.com/CSUBioGroup/DPFunc/blob/main/DataProcess/Process_data.ipynb to process my proteins?

Thank you so much for your help!
Best,
Sophia

[CSUBioGroup]

Sry for late response. The saved models are trained on the large-scale datasets. It is true that you can provide your own protein data as the test set to predict and you also need to generate corresponding pid_list_file, pid_pdb_file, and interpro_file. As you would like to predict, the pid_go_file can be empty and ignore the output evaluation metrics of our models.

[sophiachen1]

Thanks for your reply @CSUBioGroup. Can you provide guidance in generating file interpro_list_26203.pkl and how do you run interproScan to get the output of all_protein_interpros.pkl for users' protein sets? I really wish I can try out this wonderful tool. However, I cannot find script or illustrations on how the all_protein_interpros.pkl file is generated. It's very much appreciated if you can provide the instructions in a timely manner. Thanks ahead!

[sophiachen1]

Hi @CSUBioGroup, by checking on the all_protein_interpros.pkl file, are the values all InterPro annotation accession numbers associated with the protein from InterProScan output? and interpro_list_26203.pkl is the whole collection of all accesions?

[sophiachen1]

Hi @CSUBioGroup,

I hit an error when running DPFunc_predict.py on my own protein sets:

raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for combine_inter_model:
size mismatch for inter_embedding.embedding_layer.weight: copying a param with shape torch.Size([22369, 1280]) from checkpoint, the shape in current model is torch.Size([1873, 1280]).

I got 1873 interproscan IDs for my test protein set. and I find there are new IDs predicted for my proteins by InterProScan that are not present in the 22369 list (I suppose these are the interpro IDs you used in training?). In this case, will DPFunc predict on my proteins?

Thank you,
Sophia

[goes0n]

Hi Sophia,

It is true that 22369 interpro items are used for training our model. So, if you would like to test your proteins on our trained models, you need ignore some items which dose not exists in 22360 items. I hope this response is helpful for you.

Best,
Wenkang

[goes0n]

The interpro_list_26203 file contains all related interpro items from our training/validation/test proteins.