Combine ions to salt

[dswigh]

Combining ions to their respective salts will mean that there will be fewer components per reaction, which may be helpful for ML model prediction.

NB: "[Na+].[OH-]" becomes "[Na]O" because non-organic atoms must have square brackets around them for the string to be a valid smiles string.

[dswigh]

This turned out to be more complicated than initially expected.

We were planning to simply look at the lists of reactants, agents, and products, and in case there's a single cation and a single anion, merge them to the corresponding salt. After thinking more carefully, I realized the following issues with this:

• Case 1: There's only one cation/anion (e.g. only [OH-] is listed, with no cation). This can't be merged, so the dataset would both have hydroxide ions and also salts
• Case 2: There are multiple different anions and/or cations. How do we match up the ions?
• Case 3: The anion is a reactant, while the cation is an agent/product. How do we match up the ions in this case?

I think it's quite likely that when there's multiple (formal) charges (especially within the same molecule) that adding such a function could lead to unexpected and unintended consequences. As an example, consider the following 2 reactions:

1. CC1N=CC2C(C=1)=C(N+=O)C=CC=2.[Cl:15][C:16]1[CH:25]=[CH:24]C:23=[C:22]2[C:17]=1[CH:18]=[CH:19][N:20]=[CH:21]2.Cl.CC1N=CC2C(C=1)=C(N+=O)C=CC=2.[IH:44]>>[IH:44].[Cl:15][C:16]1[CH:25]=[CH:24]C:23=[C:22]2[C:17]=1[CH:18]=[CH:19][N:20]=[CH:21]2
2. S:1([O-:4])(=[O:3])=[O:2].[NH4+:6].[NH4+]>O>S:1(=[O:2])([OH:5])[O-:4].[NH4+:6].S:1([O-:4])(=[O:3])=[O:2].[NH4+:6].[NH4+:6]

In addition to these reasons, consider the following:

1. When we put trust_labelling = False, we take list(set()) of the molecules lists, which (purposely) breaks stoichiometry in favour of removing duplicates, however, by merging ions to salts we would need to selectively reintroduce stoichiometry (to match formal charges) - which sounds like a bad idea.
2. It's tempting to add manual merging rules (e.g. when you see [Na+] and [OH-] in the same reaction, merge them to [Na]O, OR when you see [Pd] and C in the same reaction, merge to [Pd]C, but adding a rule-based system should only be done with extreme caution, since there are so many edge cases in chemistry - keep in mind we're already using a rule-based system for solvent identification, and the set of industrially relevant solvents must be much smaller than the set of molecules that can carry a formal charge