Akademik Veri Yönetim Sistemi
HIBIT 2019, İzmir, Türkiye, 17 - 19 Ekim 2019, ss.233-234
Protein kinases are integral players of cellular metabolism. The
essentiality of proper kinase functioning is challenged by mutations
occurring in kinases, which often result in severe diseases, like cancer.
Point mutations occurring within or in the vicinity of catalytically
important kinase motifs can switch the kinase conformation to
constitutively active or drug resistant state. For many years, it has been
of great interest to assess structural/kinetic impact of protein kinase
mutations. Thus, it is a requisite to compile accurate binding affinity
prediction workflows to estimate the functional impact of kinase
mutations in silico. Expanding on this, we compiled the first high-
resolution kinase:ligand benchmark to assess the field’s capability in
predicting the impact of kinase mutations on kinase:ligand binding
affinities.
To that end, we collected a set of 12 wild type kinase structures
and their 49 mutant states from Protein Data Bank. These numbers
represent the cases in which both wild type and mutant states of the
kinase of interest are bound to the same ligand. Our benchmark is
made of cytosolic and membrane-associated protein kinases, mainly
functioning in cell cycle, cell growth, DNA damage, metabolism, and
transcriptional regulation. Ligands bound to these kinases are mostly
nitrogen-rich poly-heterocyclic compounds (46% are halogenated).
Together with the structures, we have also collected experimentally
determined kinase:ligand binding data (either IC50, Kd, or Ki) from
PDBbind.
As of today, there is no tool specialized to predict the functional impact
of kinase mutations. Though, there are web-based approaches
poised to estimate protein-ligand binding affinities. Within this
context, we chose the most commonly used protein:ligand affinity
predictors (HADDOCK2.2[1](refinement interface), PRODIGY-LIG[2],
DSXonline[3], and KDEEP[4]) to estimate the binding affinities of our
wild type and mutant complexes. When run only on the wild type
complexes, PRODIGY-LIG correlated computed andexperimental
binding affinities (log(IC50)) the best (Pearson’s R2=0.76) (Figure
1). When the mutant forms were included in the data set, none of
the webservers could produce a meaningful correlation. In this case,
the best performer KDEEP could relate the calculated affinities to
experimental Ki values with a Pearson’s R2 of 0.32. We are currently
analyzing the results to understand if any particular characteristics of
the mutation, protein type, or ligand are responsible for this sharp
POSTER PRESENTATIONS234
drop in the prediction accuracy. After this, we plan to probe the
predictive capacity of these webservers on a derived benchmark set,
made of predicted and experimentally determined binding affinities
of the mutant cases normalized according to their wild type values.
Our ultimate aim is to present an optimal affinity prediction workflow
that can aid experimentalists in designing kinase mutations during
their experimental setups.
Keywords: Binding Affinity Prediction; Protein Kinase; Point Mutation;
Benchmarking