Refactor module to look 3D Similarity first

[GemmaTuron]

Currently the pipeline does the docking of all sampled smiles (10 conformers x smiles, 10 poses x conformer) and then selects the best one to check 3D shape similarity We propose to use the best conformer selected in the 3D Shape similarity for Docking only

• Do we also want to set a threshold in the 3D shape for a molecule to go to docking?
• VSFlow joins several conformers into one, so if you pass 10 conformers, you may get results for only 3 or 4. Is this correct?

[GemmaTuron]

Note: to simplify the workflow, and given the new configuration, this repository will not be used without a query smiles. An additional script will be created to do just docking without a query smiles and bypassing the 3D shape similarity

[GemmaTuron]

Update: we need a better way to find the right conformer to pass to VSFlow. I am trying the Sensaas-flex package https://academic.oup.com/bioinformatics/article/40/3/btae105/7612231

[GemmaTuron]

Update on Sensaas_Flex:

The package needs some manual installations, but I've managed to make it work. We are interested in the sensaasflex and meta-sensaasflex functionalities, which align a molecule (source) to a target molecule (our "query"). The target does not move, only the source. What I have found out:

• If you pass 2D structures, the 3D structures it generates for the alignment are really not interesting, we need therefore to generate conformers first.
• If we generate 10 conf. per molecule (using smiles-to-3d) and pass that file to sensaas_flex, it takes ~1h to run 11 molecules (total 110 scorings), but the output is quite accurate I would say. I've been using dCA as a target and in the pool of 10 sampled molecules I have dCA, which scores 1.5 on sensaa_flex (the maximum is 2). Other similar molecules score in the range of 0.6-0.8. Some post processing to select the best conformer of each molecule is needed.
• I use the output of sensaas_flex (best conformer per molecule) for vsflow - in that case, it goes very fast (only 11 molecules), and while it gives a ComboScore of 1 to dCA, the rest of the molecules score quite low (0.2-0.1) mostly due to the low FP3SimScore they get. The ShapeScore is better. I do not see an improvement of passing sensaas_flex processed molecules vs passing the original conformers obtained with the smiles-to-3d.

Next steps:

• Decide if we want to use either sensaas_score or vsflow for selecting one conformer per molecule and then do docking.
• Refactor the code accordingly