What is the difference between MolScore and DockStream?

[xuzhang5788]

Thank you for your repo. Your opinion is highly appreciated.

[MorganCThomas]

Thanks for the question. Firstly I've not tried DockStream yet so my opinion is based on my current understanding, which may not be completely correct.

There is quite some overlap between functionality. For example,

• DockStream can prepare ligands using rdkit, tautenum, ligprep, corina, openeye. While MolScore has code for rdkit+epik, ligprep, openeye omega, moka+corina and gypsum_dl. Currently, their code is more modular, while mine is specific to different docking protocols (e.g., gypsum_dl + smina). However, separating ligand preparation for modularity is on my to-do list.
• DockStream can dock ligands using rDock, Openeye Hybrid, Glide, Gold, AutoDock Vina. While MolScore only has code for Openeye Fred (untested due to lack of license), Glide and Smina (fork of Vina).

Key differences:

• DockStream in itself has automated scripts for retrospective analysis of protocol performance.
• DockStream includes protein preparation, while MolScore does not and leaves this to the user, which in my opinion is important. (Although DockStream doesn't automatically prepare proteins for Glide, so in that case, they're the same).
• It seems DockStream is parallelised using python multiprocessing (use cores from running machine/node) whereas I use Dask (which can use cores across different compute machines/nodes).

Conceptual differences:

• DockStream seems like an additional scoring component to reinvent, which is deeply integrated with reinvent. While MolScore is a scoring component manager designed to be flexible, customisable and to be used with any generative model requiring scoring (not just reinvent). Including score transformations and combinations for multi-parameter optimization.
• MolScore also focuses on gathering all generated molecules in a data frame for downstream analysis using MolEval. This also enables other benefits such as not re-running duplicated molecules that have been scored previously (especially when scoring is expensive such as docking). Such that comparison between scoring configurations / generative models has also been at the forefront of my design.
• I've written both an older dash monitor and newer streamlit monitor to monitor generative model progress 'live', which comes in very handy and has some nice functionality (I think more than tensorboard). Such as browsing generated compounds and being able to also view docked poses.
• I've written a streamlit user interface for writing the configuration file (.json file). I think this is useful as it lists and describes all the options possible for the different settings. This is done by automatically parsing the code and code documentation, therefore if a user adds their own component properly annotated, it will automatically be added to the streamlit configuration user interface.

Other than the listed differences, there is of course the matter of fact that I've not published a software paper yet, I currently use MolScore more than I develop it, unfortunately. As for writing a publication I've got at least 2 other pieces of work to write up first.

[xuzhang5788]

Thank you so much for the detailed explanation. I think the combination of different generative models and docking packages is very important. There are so many generative models, but we can't say which one is the best for every problem. The docking packages are the same. So, the freedom of choosing the different combinations and collecting all the results together is very important. I am looking forward to your new update soon. Thank you again for your great work.