| Summary: | <p>Small molecules can be designed to interact with a specific protein or proteins, and, by
extension, to modulate the biological pathways in which the respective protein is involved. Small molecules can thus influence a disease phenotype on a cellular, tissue, or
organism level. However, some compounds, so-called ‘promiscuous’ compounds, modulate the phenotype by binding to a wide range of on- and off-targets. In the case of
drugs, this binding to off-targets can be responsible for side effects, as the perturbation
of other biological pathways can lead to severe or even lethal side effects. ‘Selective’
compounds target only one protein or a small set of closely related targets, the ‘ontargets’, at a given bioactivity range. Contrary to promiscuous small molecules, they
should have a low number of off-targets.
</p>
<p>The selectivity or promiscuity of a compound is dependent on both protein and ligand
structural features. I am the first to explore the inclusion of individual residue movement
as a descriptor in random forest models. Using simple 2D flexibility descriptors, I
show that I can identify flexible regions, depending on the data set. However these
flexibility values are too uniform and coarse to be used as descriptors in the random
forest algorithm to identify up discriminating features.
I also show that the current 2D random forest models and the available kinase data are
sufficient to correctly predict the activity of selective compound-target pairs depending
on previous data.</p>
<p>Lastly, I investigated the use of 3D protein-ligand interaction features in filtering out
selective (and promiscuous) compounds and the use of interaction grids to classify candidate compounds in different binding affinity ranges. Both promiscuous ligands and
selective compounds active against closely related targets tend to have a conserved binding mode across their kinase targets. Additionally, binding features cannot be used to
separate strong or weak binders for a given target ; however, a separation based on a
3D grid might be promising.</p>
<p>However, from a biological perspective, the often unknown involvement of a target in
other pathways (‘pathway crosstalk’) might add confounding biological readouts and
thus unexpected side effects, even if the compound is selective against the target of
interest. Hence, I explored drug targets from a pathway-centric perspective. I show
that the overall number of pathways associated with all drug targets of a drug does
not correlate with its success or its withdrawal. Lastly, I show that - depending on
target class - drug targets are generally associated with a wider range of pathways
than ‘undrugged’ proteins, highlighting either a research bias towards drug targets or
possibly that drug targets are chosen due to a higher involvement in pathway crosstalk.
I also found that pathways could be targeted for the same disease area using multiple
associated targets, thus highlighting a potential for finding novel drug targets outside
the traditional ‘target class’-centric paradigm.</p>
<p>In conclusion: in order to study selectivity, I have focused on a widely studied target
class, namely kinases. My results indicate promise in using 3D-based selectivity models
to derive kinase inhibitors, especially combined with the use of a standardized nomenclature across kinase binding pockets. Secondly, I also show the necessity of using a
more ‘pathway-centric’ exploration of drug targets instead of the current, traditional
‘target class’-centric approaches.</p>
|
|---|