Development and validation of a 3D similarity method for virtual screening

A predictive 3D similarity workflow approach has been developed using a set of modular
Java computer programs that implement algorithms that aim to capture the key
components of a 3D similarity search and aim to incorporate methods that address both
the similar property principle and molecular recognition paradigms. This approach will
expect as input a single query molecule conformation (at least one conformer is required
per molecule) and will identify molecules that are similar to it when compared with a target
database of 3D conformations.

This workflow is achieved by first mapping each of the molecular conformation’s geometric
coordinates, together with atomic property data, to abstract representative models
referred to as fuzzy pharmacophore objects. A geometric partitioning approach maps full
geometric atomic coordinates to a reduced point representation for a molecule in order to
capture the overall global shape of the molecule in relatively few points. This sort of
“reduced points” approach for molecular representation was first suggested by (Glick et
al., 2002) in the context of Protein active site identification. Pharmacophore classifications
are applied to the molecular fragments via mapping of internal constituent group atoms
and their properties in order to assign the amount of potential interaction type present.
The classifications are Hydrophobic, Aromatic, Acceptor, Donor and Hydrophilic and each
atom can be mapped to several of these type definitions. Thus we have assigned a
biologically relevant code to each of the fragments. These fuzzy pharmacophore object
abstract representations will naturally provide a summary level description of a whole
molecule in a relatively small number of geometric points.
Two such objects are then aligned to minimise the RMSD between points and the volume
and properties overlap is evaluated in order to derive global 3D similarity scores for each
alignment. One alignment method is to systematically align representations and is in
essence a triangle and tetrahedron matching search technique. The second alignment
method is based on graph theory and parameterised maximal common substructure or
clique detection is applied to a correspondence graph constructed using two
representations, followed by minimal RMSD alignment of the evaluated Bron-Kerbosch
cliques with the Kabsch rotation algorithm. This provides an alternative and more efficient
approach to systematic alignment since the systematic approach is limited to aligning four points maximum.

A volume and property overlap scoring function is used to compare two
such fuzzy pharmacophore objects and the resultant Tanimoto coefficient is used for
ranking. Initially representations of similar size and with equivalent numbers of points
(typically three to six points) are compared and are considered shape searches.
Subsequently, objects of different scales and representations are compared in a sub-shape
search sense, whereby a smaller object could feasibly be searched for within a larger
object. The graph theoretical approach to alignment and clique detection facilitates shape
and sub-shape search automatically by including the entire representation or just the
cliques in scoring.

In principle there are many potential ways to overlay two molecules and the sub-shapes or
fragments contained within each molecule. Each alignment can score differently and
certain alignment orientations will maximise or minimise certain aspects of the scoring
criteria. Hence, several key alignments are feasible between two conformations which may
define some or all of each molecule that is biologically active in a given context. An
alignment and associated maximal volume and properties overlap score is used to rank
order the molecules by normalised similarity. When applied to a target database evaluated
similarity measures are used to order the list for proposed biological activity. The overall
workflow is thus described as a hybrid shape / properties comparison and fragment based
biosteric similarity search. The volume distribution and by implication shape, as well as
mass derived pharmacophore feature density overlap scores, are determined and thus this
aims to capture both shape and pharmacophore search.