Dear all,

When I made the changes to the handling of stereochemistry in the Q3
2008 release of the RDKit, I attempted to get the canonicalization of
ring stereochemistry specifications (i.e. things like the molecule
encoded by the SMILES "c...@h]1cc[c@@H](C)CC1") right. I knew that it
was unlikely that I had actually done so, so I added a warning to the
code that you probably have seen:
   Warning: ring stereochemistry detected. This may not be handled correctly.
Based on the tests I did before the release, I believed that the
current status of the code was correct a large enough fraction of the
time that this warning was sufficient.

After some large-scale testing done by a colleague, it turns out that
this assumption was very bad: the handling of ring stereochemistry was
incorrect much, much too frequently. To be explicit: the SMILES that
were being generated for molecules that fell into this class did not
have the correct stereochemistry; i.e.  the molecules were wrong. This
is far from acceptable: it's ok to generate output that's
non-canonical, but not output that's incorrect.

Here's what I have done in order to address this:
 - The stereochemistry specifications on ring atoms are no longer canonicalized.
 - The warning about ring stereochemistry has been moved into the
SMILES writer and now indicates clearly that the output is not
canonical
 - Detection of cases that can have ring stereochemistry has been
improved; spiro centers, symmetrically substituted centers, and
bridged systems no longer trigger the warning.

These changes are all in svn; I encourage you to update if you are
building from source. I plan to do the Q4 release sometime in the next
couple of weeks; the changes will also be in that release.

Once I've assembled a sufficiently large set of test molecules I will
revisit the canonicalization of these systems sometime next year and
try again to get this right.

This is definitely the fun part of an open-source project: everyone
sees when you screw up. :-S

-greg

Reply via email to