RNA004_HEK293T model

[yuukiiwa]

Hi @chrishendra93, I have added the RNA004_HEK293T model to m6anet. Jonathan suggested adding a --pretrained_model flag (RNA002_Hct116 would be default if a user doesn't provide the flag). I have tested this flag on arabidopsis and RNA004_HEK293T. I also added a line in the README to point users to install m6anet via conda. Thanks!!

[chrishendra93]

hi @yuukiiwa :

• Seems like the build test is failing because you have not added pretrained model option to argparser in pytest module
• In inference.py, when pretrained model is supplied, you seem to have a pre-defined read_proba_threshold depending on the choice of model, but this will not achieve the intended behavior since during inference, only args.read_proba_threshold (read_proba_threshold that is supplied by the user, or the default value) is accessed. I think you should overwrite read_proba_threshold in args but only do so if users do not specify any value for it

[chrishendra93]

hi @yuukiiwa, I've made some:

• Modify args directly since run_inference function reads args directly and variables such as read_proba_threshold is accessed as an attribute of args
• Moved the list of pretrained models to constants.py and add an exception when the wrong pretrained model is specified
• Set default model weight to be none and when it is specified, user should be warned that this model weight will overwrite the default model weight
• Change file names for norm factors and model weights so that users can tell whether they are trained on RNA 002 or RNA 004
• Add new version number to init.py, it is now version 2.0.3
• Modify quickstart.rst to reflect the new naming for Arabidopsis and HCT116 models as well as their default read proba threshold
• Add installation option via conda

I need the following things to clarify:

• Do we have norm factors for RNA004, or were you using RNA002 norm factors for training? If you have, this should be added as well and set as the default when users are using RNA004 pretrained model
• I think we should explicitly mention the fact that we have not tested the individual read probability accuracy for RNA004 model on the release note

[yuukiiwa]

Thanks so much, @chrishendra93!!

The RNA002 norm factors were used for training, so for the pretrained RNA004 model, the default one should work for both.

I have updated the release note to include that the evaluation was done on site-level prediction and that individual-read probability of the RNA004 model was not tested.

Thanks!!