Here are some next steps we discussed in the meeting today.
Please add additional items and feel free to add details and questions, etc.
Could also tag things that you are working on so we know who is working on what.
Thanks
—Rick
1. Combine the SAVI generated synthetically accessible structures with the generated
structures from Songhao’s generator to construct a balanced training set that includes
roughly half that are synthetically accessible and half where we have no information on
synthetic accessibility and see if we can train a model to discriminate between them.
2. Run the SAVI structures through the synethsizability rating tool that Fangfang found
to see if the tools estimates are consistent with the SAVI expectations.
3. Write a script to sweep through features in both Gene Expression and Drug Descriptors to
see which columns can be predicted from the remaining columns.
4. Look into adding dropout based UQ to UnoMT pytorch implementation
5. Look into a quantized inference optimized version of UnoMT to accelerate inference use case
and assuming progress on #4 construct a inference optimized version of UnoMT with UQ support.
6. Gather in one place a large-scale chemical compound library of SMILE strings from
the various sets to use for training and testing.
7. Get a pipeline working to go from SMILES to Mordred descriptors (2D and 3D) so that we
can generate descriptors for the library in #6
8. Construct a variety of training sets for the experiment to see if we can learn a model to
invert the SMILES —> Descriptor function.
9. Get everyone accounts on Theta and Washington.
10. Gather relevant papers together in one Repo
