Adding more components to the target scoring function

[palazzef]

I'd like to add more components to the target scoring function. In one of the notebook I found this nice example:

maximizing drug likeness and synthesizability

target = {"qed": {"weight": 1.0},
"sa": {"weight": 1.0}}

Which is the correct syntax to add ECFP2 tanimoto similarity? Or more in general, where do I have to check to see how to add components? Thanks for the help!

[jannisborn]

Hi Ferri,

thanks for your interest, here are the supported minimzation functions:

gt4sd-core/src/gt4sd/algorithms/controlled_sampling/paccmann_gp/implementation.py

Line 51 in d6dd591

MINIMIZATION_FUNCTIONS = {

It does not contain ECFP2 similarity. But this model relies on the paccmann_gp package which has a callable minimization function that supports arbitrary callables. See here: https://github.com/PaccMann/paccmann_gp/blob/main/paccmann_gp/callable_minimization.py

You would have to define the function yourself and then pass this just like the others.
To work properly, this will require a change in the gt4sd code. In the line I linked above, it might suffice to add 'callable': 'CallableMinimization'

[jannisborn]

If this line suffices or you find another way to make this work, we would be delighted to see a PR and integrate this into the package for future releases.

[jannisborn]

@palazzef this feature will be released in the next version, see the PR I opened #182

[palazzef]

Great, thanks!

[palazzef]

Hi @jannisborn, maybe I have to expand a bit more my question.

Basically I would like to expand the conditional part of the discovery-demo example by adding more components (qed+similarity+sa) and also more generative models (AAE, VAE, ORGAN etc..).

For this reason, I am looking for a mix&match version of "target=gentrl_ddr1_smi" (similarity only) and "target={"qed": {"weight": 1.0},
"sa": {"weight": 1.0}}" (QED and SA only).

[jannisborn]

Adding more models should be easy, I recommend checking the docs, e.g., see the GuacaMol models; https://gt4sd.github.io/gt4sd-core/api/gt4sd.algorithms.conditional_generation.guacamol.html

There's also a function called gt4sd.algorithms.registry.list_available that shows all algorithms including versions.

For the mix & matching: If you want to plot the SA instead of QED, can use the SA property predictor:

>>>from gt4sd.properties import PropertyPredictorRegistry
>>>sa = PropertyPredictorRegistry.get_property_predictor("sas")
>>>sa('CCO')

If you want to have a generative model that simultaneously optimizes similarity to a seed molecule, QED and SA, then, AFAIK, PaccMannGP is the only model that supports this and the way to achieve this is the one I described above. Hope this helps!

[palazzef]

Thanks for the answer.

I thought that the models with "get_target_description()" function (like SMILESGAGenerator) were able to do MPO given a specific multiparameter target function.

[jannisborn]

For every model, that function just gives the description of the target value. But you're right for some of the Guacamol models, it should be possible to do such MPO

[palazzef]

ok, so if I can suggest a possible change (related also to #182), I will add a general "scoring function builder" that people can then easily use within several models (some of the guacamol models, reinvent, pacmannGP, and so on..). Maybe it is already like this, sorry I haven't had time to check the code in details.

[jannisborn]

This is available in a base form under gt4sd.properties.scores, see here: https://github.com/GT4SD/gt4sd-core/blob/main/src/gt4sd/properties/scores/core.py.

I'm not 100% sure that you can use the exakt same syntax for all models (GuacaMool, Reinvent, PaccMannGP), maybe @drugilsberg knows this from the top of his head.

There is even a PropertyPredictorScorer class that should allow to use all the property predictors from gt4sd.properties as scoring functions. See also: #124