translation between representation

[kjappelbaum]

needs a bit of thinking for the best design

but, in principle, we can make many pre-training tasks by translating between representations

could also be the question: are X and Y the same molecule? Where X and Y are in different representations (or randomized SMILES)

[Arkhymadhe]

Need more clarity here. Are you talking of featurizer that converts from one representation to another (e.g., SELFIES to SMILES) or talking of a featurizer that takes in two molecular representations and tells if they represent the same molecule?

[Arkhymadhe]

could also be the question: are X and Y the same molecule? Where X and Y are in different representations (or randomized SMILES)

I think the Comparator API may be of use here.

[kjappelbaum]

yeah, so the basic thing might be solved with a comparator. I think the question here is rather where the sampling will be implemented.

[Arkhymadhe]

yeah, so the basic thing might be solved with a comparator. I think the question here is rather where the sampling will be implemented.

Could you provide more expatiate on this?

[kjappelbaum]

Could you provide more expatiate on this?

you will only get a meaningful signal if you run this on a set of molecules. And then you have the questions:

• do you run it on all pairs
• if not, how do you select on which pairs to run it
• do you run all possible translations on all pairs or do you just use one or a few?

[kjappelbaum]

could we build something based on the Comparator API?

[Arkhymadhe]

could we build something based on the Comparator API?

Yes, I'm sure we can. The Comparator exists already, so the harder part is done. All I'd need to figure out is the basis of the comparison.

[Arkhymadhe]

I think I have a basic solution worked out for this. Expect a PR tomorrow.