Direct use of the python package and runmodes for library variations

[DAlvGar]

Hi all,

I would like to create some friendly and automatized interface so my colleagues could easily create a library from a reference structure with vartiatons on the structure (decorations but also some scaffold changes allowed). To do so, I have a few questions:

1. On the method itself: What run mode/modes would be better for this purpose? I was checking Mol2Mol but any advice is very welcome.
2. Regarding mol2mol, should I run it in transfer learning mode with my molecule as reference and then sample, or should it be run in reinforcement learning mode?
3. For easier programming: Are there any examples on how to use the package directly from python without using the CLI or going through toml/json config files? I stuggle to understand how the code is structured to do so.

Thanks a lot!
Daniel

[halx]

Hi,

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

1. That's basically a nearest neighbour search. Since you are happy we changing scaffolds Mol2Mol is the right way to go. If you wanted to hard-constrain the scaffold you would use the Libinvent generator.
2. The idea of TL is to focus your model towards a set, typically rather small, such that it produces those structures (sequences really) with higher probabliity. RL is used to impose restraints through scores onto the generated molecules.
3. The main entry point is Reinvent.py but each runmode has a separate run_*.py which is basically setup. You will need a configuration (in the form of a dict) in any case at some point.

HTH,
Hannes.