Would it be possible to use chemprop as the fingerprint?

[UnixJunkie]

I don't know how yet, but I guess a pretrained chemprop model (CHEMELEON it is called) could be used for molecular encoding/vectorization.

I wonder how it would compare to classical fingerprints.

How easy is it to add a new fingerprint/encoding function that would not be calculated by rdkit?

Is there an example of a custom molecular encoding function?

[UnixJunkie]

https://arxiv.org/pdf/2506.15792

I understand that's a 2048 float fingerprint.

[UnixJunkie]

Ok, so the code allowing to do that is here:
https://github.com/JacksonBurns/chemeleon/blob/e629639e4a74f475372a878f351a326289146b18/chemeleon_fingerprint.py

[EBjerrum]

In principle It shouldn't be too difficult to create a transformer wrapper around chemeleon. The problems are dependencies and import. The Chemeleon github repository doesn't even have a pyproject.toml or setup.py install file defining the dependencies needed, neither is it structured as a proper package. I'm not a fan of dragging in big projects with many potential subdependecies, but we could add it as a contrib, and make an import test with instructions if it's not installed. But somebody (Claude code??) need to turn Chemeleon into a python package first and maybe get it on pip.