Remove RDKit down-constraint from sire conda package

[j-wags]

We recently found that the root cause of a user report was due to the sire 2025.1.0 conda package down-constraining rdkit (constraint rdkit >=2024.3.5,<2024.3.6.0a0 which caused an error due to a change in some sanitization behavior that OpenFF NAGL relies on). Would it be possible to remove this down-constraint from the sire package?

[lohedges]

This is tricky. We have support for conversion to RDKit objects via the C++ API, hence the need to pin. (We link against a specific version of libRDKit.) However, we don't actually specify the version of RDKit that we use so it is resolved by the needs of our other dependencies for self-consistency. We are in the following situation:

• Unpin the runtime constraint and Sire RDKit conversion will fail if libRDKit is incompatible.
• Try to pin RDKit to a newer version that doesn't have the sanitisation bug and our build environment can't be resolved.

[lohedges]

A possible solution could be to do all our RDKit interconversion via the Python API. However, I'm not sure if there are any limitations of doing so and there probably isn't the bandwidth right now.

Also, how was the environment even resolvable if you require a newer RDKit in the first place? Perhaps you haven't introduced a pin yet since the sanitisation issue has only recently been found?

[lohedges]

Another option would be to try adding openff-nagl to the host requirements section in our recipe to see if that resolves a working version of RDKit. I'll try this morning.

[lohedges]

No, this won't work since the openff-nagl-base package has no pin on the runime rdkit version.

You state that:

We recently found that the root cause of a openforcefield/openff-toolkit#2089 (comment) was due to the sire 2025.1.0 conda package down-constraining rdkit (constraint rdkit >=2024.3.5,<2024.3.6.0a0 which caused an error due to a change in some sanitization behavior that OpenFF NAGL relies on). Would it be possible to remove this down-constraint from the sire package?

But there is no pin on rdkit. If you require specific sanitization behaviour, then I suggest adding a minimum version constraint.

As such, our solutions are:

1. Work our which of our dependencies are constraining the rdkit version and see if we can drop them, or massage a solution. This would likely be a massive pain and I doubt that the same solution would work on all platforms and python variants.
2. Re-write the Sire-to-RDKit conversion code using the rdkit Python API. This would free us from linking against libRDKit so we wouldn't enforce a runtime pin. However, environment resolution would still be at the mercy of the rdkit version that satisfies the constraints of our dependencies, hence the need for packages to pin if there are issues, like the one that you mention above.

[lohedges]

It looks like this was a result of us not noticing the change in the Boost conda package naming scheme. This meant that we could only resolve older versions of RDKit, that had also been built against the old Boost package. With the new scheme we can resolve the latest RDKit without requiring any pin.

[j-wags]

Oh wow, thanks so much for the deep dive and quick action on this, and sorry I was incommunicado for so long!

I think it's worth us considering up-pinning RDKit on the OpenFF side, we'd initially only been peripherally aware of this since the sanitization error only pops up on specific chemistries. Either way we really appreciate the package rebuild with the latest RDKit. I do notice that the latest dev-label sire conda packages still have the rdkit version constrained (rdkit >=2025.3.4,<2025.3.5.0a0), but the version number that they support resolves our current issue, and presumably future releases will continue to update to more recent versions. Again, thanks for being so quick and on top of this!