[Rdkit-devel] lapack dependency finally gone

[Greg L. <gre...@gm...>]

Dear all,

This morning I finally figured out how to get rid of the lapack
dependencies in the RDKit in a reasonable (and backwards compatible)
manner.

Lapack was being used to generate eigenvalues of the distance matrix
for molecules and fragments of molecules. These eigenvalues are used,
along with the Balaban J index, as invariants to distinguish molecular
subgraphs from each other in the Subgraphs and FragCatalog code. I've
replaced the invariant calculation completely and now use a hashing
approach with the Morgan algorithm similar to that used in the
MorganFingerprint code. The new, simpler, code is in
$RDBASE/Code/GraphMol/Subgraphs/SubgraphUtils.cpp. There's not much in
the way of documentation in the new code and it could definitely stand
to be cleaned up, but the big step is taken.

A particularly pleasant side effect of all this is that the
findUniqueSubgraphs family of functions are now substantially faster
(about a factor of three). The fragment catalog stuff should also be a
lot faster, but I haven't done the benchmarking there yet.

The last remaining lapack dependency is in the regression tests for
the power eigenvalue solver; I'll get rid of this in the next few
days, clean up the build system, and remove all traces of lapack from
the code; that's easy stuff.

This one has been nagging for a while... it's nice to have it done.

-greg