The storage and retrieval of chemical graphical datatypes such as structures and reactions in relational database systems is a common technique used in academia and industry alike. Due to the computationally intensive algorithms used for (sub)graphisomorphism detection, such systems commonly use faster screening mechanisms in order to reduce the set of potentional match positives before applying aforementioned algorithms.
Widely used screening mechanisms are based on numerical and binary vectors, called
fingerprints, with a clear dominance of binary fingerprints due to the raw speed
advantage of bitwise operations and compactness in storage. The two most commonly
used types of binary fingerprints are path-generated and substructure-generated, both
of which have specific shortcomings, especially blind spots.
To overcome this shortcomings, the Pgchem::Tigress chemistry extension to the
PostgreSQL object-relational database management system uses a hybrid binary
fingerprint, consisting of an invariant path-generated part and an substructure-generated
part which is externally configurable through a dictionary of substructure patterns.
This thesis presents a novel approach of using dynamic discrete optimization to find an
optimized dictionary configuration for the substructure-generated part of the fingerprint
for arbitrary sets of structural data.
By means of applying the method developed in this thesis, the computational power
neccessary to run a chemical information system can be reduced by 42 percent on
average. By improving the query throughput, upgrading the server hardware to the
next level of computational power can be avoided and thus opportunity revenues of
the operating costs are realized.
Update: It is now readable online.