2024-07-17 BW/LW Check-in

General updates and discussion on projects

• Torsion fitting

  • fit finished – hpc3 is down today, hopefully will have benchmarking tomorrow

• smee

  • Got TDs loaded, starting step 2

  • Probably expecting better results compared to just training to opt dataset, but worse than SPICE.

  • Around 13k (just 5k with opt)

• lipids

  • ready to put through QCA submission/s – 72 molecules

    • TD – use 15 ° for now

  • Have the SMILES, just finishing up TD

  • One of the alkane parameters has 1.x M in chembl

  • Lipid MAPS produced ~47.5k molecules, ~450k fragments

    • Has 48k molecules

    • XFF 20% dataset had ~10k molecules, 128k conformers

    • Try to cluster 10-20% down to ~50k fragments?

    • Also do an entire molecule dataset

  • Happy to do training

• benchmarking infrastructure repo?

  • “combinations” of QCRecord IDs

  •  

• infrastructure work --

  • going well, working on QCSubmit

  • Caching PortalClient not threadsafe

  • Moving fragment database code into own repo, now technically installable, has a few tests

  •  

• Next week: review of roadmap plan and where we are

  • Coming month

    • Focus on finishing up torsion multiplicity

      • Know we’ll be done if benchmarks aren’t worse

        • Benchmarking to OpenFF Industry 1.1 dataset

          • Dataset name: 'OpenFF Industry Benchmark Season 1 v1.1'

          • JSON file, probably the full dataset

        • Collect record IDs where connectivity changes, and QM is just wrong

          • Pass record IDs in a text file

          • With “bad” record IDs, need to remove them from sqlite database to fix up ddE calculations

        • Benchmark additional torsion optimizationdataset as well

        •