Generative model from team TCS

[rarijit]

MurD_generative_modelling_TCS.pdf We have designed few molecules against allosteric site of MurD ligase protein. To design molecules specific to the allosteric site of MurD ligase protein, the conditional variational autoencoder (CVAE) model developed in our recent study (Krishnan et al., 2021) on structure-based drug design using deep learning was utilized. The model takes as input a residue-level binding site graph of the target protein of interest and can generate novel small molecules which can bind to the target binding site, based on a reinforcement learning framework.

[mattodd]

That's great @rarijit , thanks to you and the team for the interesting suggestions! I see the molecules have very nice logP predictions, which will help in the assay.

A quick check:

1) Your method. Can it be described as generative (as opposed to minimisation/docking of commercial libraries)? 2) Did you already check for commercial availability of these compounds? Otherwise @edwintse @danielgedder could you please begin that search? 3) The predicted molecules are binding in the same region as the starting point fragments?

[rarijit]

Thanks @mattodd . Please find below the answers to your questions.

1. Your method. Can it be described as generative (as opposed to minimisation/docking of commercial libraries)?

Ans: No minimization/docking of commercial libraries was used. We used a structure-based generative model as described in our recent publication (https://pubs.acs.org/doi/10.1021/acs.jcim.1c01319 ). The multi-parameter optimization approach was also taken from our recently accepted manuscript (https://pubs.acs.org/doi/abs/10.1021/acs.jcim.2c00462 ).

2. Did you already check for commercial availability of these compounds? Otherwise @edwintse @danielgedder could you please begin that search?

Ans: No. These are novel molecules, but some similar molecules/part of the designed molecules may be available from commercial libraries. It would be great if @edwintse @danielgedder can help. We checked the synthesis accessibility score of the molecules using two methods, RAscore and SAS. We tried to check the synthesis route prediction using IBM Reaxys server but it is not working at the moment.

3. The predicted molecules are binding in the same region as the starting point fragments?

Ans: Yes. The predicted molecules are binding in the same region as the starting point fragments, and this was shown in the figure 1 of the document.

[edwintse]

@mattodd @danielgedder Here are all the structures. I've done a quick scifinder search and the 5 in orange have some literature routes for their syntheses. If any others look appealing we can look into potential synthesis further.

Mur Ligase Entry 4.xlsx

[edwintse]

Here are the routes for the 5 orange compounds above. I've purchased the purple reagents for the middle 3 compounds. I'm giving the first compound a skip for now since it's quite long, and the last one seems less favourable as it's quite restricted and flat with the 3 fused rings.

[drc007]

@edwintse I've implemented the AZ AI-assisted retrosynthesis tool (https://github.com/MolecularAI/aizynthfinder) on my local machine, I'm happy to run compounds to find possible synthetic routes. I just need the structures in SMILES format.

[mattodd]

I'd agree with your analysis @edwintse - great, let's go for the middle three. @drc007 Thanks. If the AZ tool spits out something better, let us know (Ed, can you provide SMILES or the chemdraw?). Always useful to check, though I like the look of @edwintse's analysis.

[edwintse]

@drc007 Thanks! I've attached the chemdraw of the schemes below. I was actually trying to see if I could get the Az retrosynthesis tool running on my Mac but was struggling a bit. Do you have any tips to get it working? The official instructions were a bit confusing for me.

Mur Ligase Entry 4 TCS Rxns.zip

[drc007]

@edwintse Here are the results. retrosyn.pptx

See what you think of the results, if you want to try more I can try and help get it installed on your Mac.

[jhjensen2]

I can also recommend https://postera.ai/manifold/ for retrosynthesis prediction