How to use scikit classification model (.pkl format) as a staged component in Reinvent4

[Jahan-I]

I am trying to generate molecules using a trained QSAR model as a scoring function. However, I am unsure if the approach in the code below correctly utilizes the QSAR model with the supported fingerprints (ECFP6 in this case). Running the code results in multiple errors, and I am unsure how to resolve them.

[[stage.scoring.component]]
[stage.scoring.component.SAScore]

[[stage.scoring.component.SAScore.endpoint]]
name = "SAScore"
weight = 0.1
transform.type = "sigmoid"
transform.high = 0.1
transform.low = 6
transform.k = 0.5

[[stage.scoring.component]]
[stage.scoring.component.ExternalProcess]

[stage.scoring.component.ExternalProcess.endpoint]
component_type = "predictive_property"
name = "tmem16f"
weight = 5

[stage.scoring.component.specific_parameters]
model_path = "opt_rf_model_ecfp4.pkl"
scikit = "classification"
descriptor_type = "ecfp_counts"
size = 2048
radius = 3
use_counts = true
use_features = true

[stage.scoring.component.specific_parameters.transformation]
transformation_type = "no_transformation"

Many thanks for the support : )

[halx]

Hi,

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

Have a look into the example scripts in support e.g. run-chemprop.py. Any parameters for your external script must either be supplied via the command line or hard-coded into your script.

I would not recommend SAScore for any serious work as it is primarily a measure for molecular complexity and a rather poor approximation to synthesisability.

Cheers,
Hannes.

[Jahan-I]

Hi Hannes,

Thank you for your prompt response, i have made a py script which imports the .pkl model, calculate necessary features (fingerprints) and feed in the .toml file for library generation. I am actually getting started with this awesome package and right now just playing around and learning to utilize to its potential. To get the best results with more relevant structures (chemistry and biological) i was thinking of incorporating a more complex scoring systems along with ligand based exploitation.

I have an small additional query, is there any other recommended package which give fair estimate of synthesizability, and guide the agent to a chemical space which has relevant synthesizable chemistry.

Thanks again :)

[halx]

There are various attempts out there to deal with synthesizability (SA) but for practical use you probably will need actual route planning. We offer AiZynthFinder for this but you need to be aware that this can be rather expensive (a minuter or more per compound). There are other solutions, of course. Simple SA scores like RAScore (trained on AiZynthFinder with ChEMBL data) and RetroGNN (ultimately trained on USPTO data) are fast but maybe not that reliable. We are currently experimenting with those but have not yet correlated their SA score with the one from AiZynthFinder. The respective papers have more information.

If you work with a fixed scaffold or a fixed set of scaffolds, you could try our Libinvent generator with the reaction filter scoring component. You can define RDKit reaction SMILES for reaction classes that you want. Obviously, this all means enumeration.

[Jahan-I]

HI Hannes,

Thanks for the giving me a good starting point to test the SA scoring systems, i will try to use faster scoring functions as i intend to generate as many compounds as possible. Secondly, the objectives of projects requires exploration aspect, such that i intend to generate novel scaffolds for my target of interest. So, at the moment i am not looking to generate congenric series and its sar exploration.

Again many thanks for answering my queries :)