feat: add tautobase

[lwehrhan]

Pre-review checklist for PR author

PR author must check the checkboxes below when creating the PR.

• I've confirmed the contribution guidelines.

Summary

The tautobase is a dataset of 1,391 tautomer pairs of organic small molecules. The geometries were refined at xtb level of theory. QM single point energies (reference for this benchmark) were re-computed by InstaDeep in-house at ωB97M-D3(BJ)/def2-TZVPPD level of theory.

Linked issue

Resolves #641

Progress

• Calculations
• Analysis
• Application
• Documentation

Testing

New decorators/callbacks

[joehart2001]

Hey @lwehrhan thanks for the pr! can you attach the data for us to upload please?

• have a look at the element filtering?

docs: https://ddmms.github.io/ml-peg/developer_guide/filter.html

[joehart2001]

I've added similar changes to the folmsbee benchmark, and apart from the comments/conflicts i think we're basically good to go. @ElliottKasoar could you take a look?

[ElliottKasoar]

This looks great! Just a couple of minor suggestions/questions, plus if you could rebase/resolve conflicts with due to merging #429?

[ElliottKasoar]

Thanks for the quick changes, @lwehrhan!

I've rebase and resolved a couple of conflicts, but nothing else should be changed. Just to note that for this benchmark, the list saved is (I think) the union of all of the elements in all pairs , not a set per structure.

As we discussed, this is fine, but if we wanted to try implementing the "full" filtering in future, we'd need to change the construction very slightly, e.g.

elements = [
    sorted(
        {
            symbol
            for symbols in pair.atom_symbols
            for symbol in symbols
        }
    )
    for pair in benchmark._tautomers_data.values()
]

This would need to match the sorting of the data/labels, which is why the example for Folmsbee looks a bit more complicated, with a zip etc.

@joehart2001 I think I'm happy with this, so if you can do a final check, we can probably merge. I'll wait for @lwehrhan to be back before making any data live, but we can start collecting it at least.

[joehart2001]

Thanks for the quick changes, @lwehrhan!

I've rebase and resolved a couple of conflicts, but nothing else should be changed. Just to note that for this benchmark, the list saved is (I think) the union of all of the elements in all pairs , not a set per structure.

As we discussed, this is fine, but if we wanted to try implementing the "full" filtering in future, we'd need to change the construction very slightly, e.g.

elements = [
    sorted(
        {
            symbol
            for symbols in pair.atom_symbols
            for symbol in symbols
        }
    )
    for pair in benchmark._tautomers_data.values()
]

This would need to match the sorting of the data/labels, which is why the example for Folmsbee looks a bit more complicated, with a zip etc.

@joehart2001 I think I'm happy with this, so if you can do a final check, we can probably merge. I'll wait for @lwehrhan to be back before making any data live, but we can start collecting it at least.

tested and all good from my side!