Kins comments

Results 8 comments of


                                            Kins

Low correlation in GFP dataset

That's true! I have tried the code and got the same conclusion with yours.

How to use with sequence-only dataset

Sorry, our model has to take both structure tokens and residue tokens as input. Maybe you should utilize AlphaFold to predict the structures.

Thank you for the work you have done. This question pertains to downstream tasks.

Mean Pooling and Attention Pooling both work well in our test. (https://huggingface.co/AI4Protein/ProSST-2048/blob/main/modeling_prosst.py#L221)

Structural tokenizer (`PdbQuantizer`) is too slow at processing long proteins

Hi, thanks for concerning our work. We will release an accelerated PdbQuantizer with multi-thread parallel processing next month.

Structural tokenizer (`PdbQuantizer`) is too slow at processing long proteins

Accelerated version speed: | Protein Name (Uniprot_ID) | Length (Local structures) | Splitting to local structure | Encoding | | --- | --- | --- | --- | | CCDB_ECOLI_Adkar_2012...

Structural tokenizer (`PdbQuantizer`) is too slow at processing long proteins

Hi, please check the new quantizer.py. (Make sure that you have installed pathos in your python env.[pip install pathos])

Issues when create environment

Sorry, ginnm==0.0.2 is a personal package and it is not used in this project. You can refer to https://github.com/tyang816/ProtSSN/blob/master/README.md and create a conda environment.

Is the MLP and training script available in the future?

We will release our training code and data for model fine-tuning, but you can see the model arc for fine-tuning at: [https://huggingface.co/AI4Protein/ProSST-2048/blob/main/modeling_prosst.py](https://huggingface.co/AI4Protein/ProSST-2048/blob/main/modeling_prosst.py) (line 1223).