Prediction of Protein Pairs Sharing Common Active Ligands Using Protein Sequence, Structure, and Ligand Similarity

We benchmarked the
ability of comparative computational approaches
to correctly discriminate protein pairs sharing a common active ligand
(positive protein pairs) from protein pairs with no common active
ligands (negative protein pairs). Since the target and the off-targets
of a drug share at least a common ligand, i.e., the drug itself, the
prediction of positive protein pairs may help identify off-targets.
We evaluated representative protein-centric and ligand-centric approaches,
including (1) 2D and 3D ligand similarity, (2) several measures of
protein sequence similarity in conjunction with different sequence
sources (e.g., full protein sequence versus binding site residues),
and (3) a newly described pocket shape similarity and alignment program
called SiteHopper. While the sequence-based alignment of pocket residues
achieved the best overall performance, SiteHopper outperformed sequence-based
approaches for unrelated proteins with only 20–30% pocket residue
identity. Analogously, among ligand-centric approaches, path-based
fingerprints achieved the best overall performance, but ROCS-based
ligand shape similarity outperformed path-based fingerprints for structurally
dissimilar ligands (Tanimoto 25%–40%). A significant drop in
recognition performance was observed for ligand-centric approaches
when PDB ligands were used instead of ChEMBL ligands. Finally, we
analyzed the relationship between pocket shape and ligand shape in
our data set and found that similar ligands tend to bind to similar
pockets while similar pockets may accept a range of different-shaped
ligands.