Curious about Chemprop scoring function

[sdey17]

Hi, I am trying out REINVENT4 and had a question. I saw chemprop being listed in the paper (Table 2) as a scoring function. I used it earlier to build GCNN classification models and was wondering what would the best way to think about implementing such a model for reinforcement learning.

[halx]

Hi,

many thanks for your interest in REINVENT and welcome to the community!

You take your ready-made model and use it as follows. In case it is not a ChemProp 1.x model but a newer 2.x, you would need to use the ExternalProcess component and write a script of your own. I do not know how to play 2.x models.

[[stage.scoring.component]]
[stage.scoring.component.ChemProp]
    
[[stage.scoring.component.ChemProp.endpoint]]
name = "ChemProp"   # choose your preferred name
weight = 0.6  # adjust as needed
    
params.checkpoint_dir = "/home/me/model_dir"  # top level directory of the model(s)
params.rdkit_2d_normalized = true  # set to false if that does not apply to you
params.target_column = "dG"  # name of the target column from the training set
 
# transform in case model output is not in [0,1]
transform.type = "reverse_sigmoid"
transform.high = 0.0
transform.low = -50.0
transform.k = 0.4

Many thanks,
Hannes.

[sdey17]

Hi Hannes,

Thank you for explaining how to implement using chemprop models in REINVENT4. So, let's say I create a classification model to predict whether a small molecule would bind to a protein. Would running reinforcement learning with that model allow us to generate more actives/inactives depending on how I set up the "weight" parameter?

Thanks,
Souvik

[halx]

Hi Souvik,

that is basically the idea, yes. Actives/inactives with the limitation, of course, as prediced by the model.

Cheers,
Hannes.

[sdey17]

Thank you Hannes! That really helps.