PTM Dephosphorylation Predictation Tool

[haibkhn]

This pull request introduces a new tool for predicting dephosphorylation. Key features include:

• Two modes of operation:

  1. Manual hyperparameter selection
  2. Automated hyperparameter search using Optuna

• Support for 3 protein language model variations:

  • ProtT5-XL-UniRef50
  • ESM
  • ProtT5-XL-BFD

• Additional hyperparameter search option:

  • SMAC functionality included, but note that the latest SMAC version (2.2.0) is not yet updated in Anaconda

@anuprulez Please review these changes and let me know if any modifications are needed.

[anuprulez]

@haibkhn thanks for the PR. I will have a look and provide my feedback

[anuprulez]

Here are a few obvious comments to start with:

• The tool does not have automatic tests. Please see here: https://github.com/bgruening/galaxytools/blob/master/tools/bioimaging/bioimage_inference.xml how to add tool tests
• CI's listing test is failing. Please fix those. All the automatic CI tests should pass.

ping @haibkhn
Maybe @bgruening also has more comments. I will look into more details probably end of this week or next week.

[bgruening]

please add a .shed.yml file

[bgruening]

Suggested change

	<tool id="hyperparameter_finetune" name="Hyperparameter Search for Finetuning model" version="1.0.0">
	<tool id="hyperparameter_finetune" name="Hyperparameter Search for Finetuning model" version="1.0.0" profile="23.0">

[anuprulez]

I think while testing, the script tries to download the LLM (ProtT5-XL-UniRef50) having a size > 10 GB. It could be the reason CI/CD throws an error related to the memory

It could be possible to provide a remote link to the model but I think HuggingFace does not allow the model from remote. The models should be hosted at HuggingFace, correct? @haibkhn

Moving to a container-based tool might help but not sure. Can the HuggingFace table used with the Flux tool might help?
I am looking at #1496 where we store the names of HF models but not the models themselves

ping @bgruening @arash77

Thanks!

[arash77]

I don't think that it is possible to test the tool completely with the models in GitHub if the model is big. unless there is a smaller version of the model available.
But for production, we can use two methods, one that I have used is to use HF_HOME environment variable to store all hugging face models or you can use cache_dir in from_pretrained function