Transcript Computational Biology

V27 Cellular Drug Network
How many drug targets are there?
in 2002: ~8,000 targets of pharmacological interest, of which nearly 5,000 could be
potentially hit by traditional drug substances, nearly 2,400 by antibodies and ~800 by
protein pharmaceuticals.
Based on ligand-binding studies, 399 molecular targets were identified belonging to
130 protein families, and ~3,000 targets for small-molecule drugs were predicted to
exist by extrapolations from the number of currently identified such targets in the
human genome.
27. Lecture WS 2008/09
Bioinformatics III
1
Drug Target: Enzymes
27. Lecture WS 2008/09
Bioinformatics III
2
Drug Target: Enzymes II
27. Lecture WS 2008/09
Bioinformatics III
3
Drug Target: Enzymes III
27. Lecture WS 2008/09
Bioinformatics III
4
Drug Target: Enzymes III
27. Lecture WS 2008/09
Bioinformatics III
5
Drug Target: Receptors I
27. Lecture WS 2008/09
Bioinformatics III
6
Drug Target: Receptors II
27. Lecture WS 2008/09
Bioinformatics III
7
Drug Target: Receptors III
27. Lecture WS 2008/09
Bioinformatics III
8
Drug Target: Receptors III
27. Lecture WS 2008/09
Bioinformatics III
9
Drug Target: Ion channels
27. Lecture WS 2008/09
Bioinformatics III
10
Drug Target: Transport proteins
27. Lecture WS 2008/09
Bioinformatics III
11
Drug Target: DNA/RNA and the ribosome
27. Lecture WS 2008/09
Bioinformatics III
12
Drug Target: Targets of monoclonal antibodies
27. Lecture WS 2008/09
Bioinformatics III
13
Drug Target: Various physicochemical mechanisms
27. Lecture WS 2008/09
Bioinformatics III
14
Summary
Many successful drugs have emerged from the simplistic ‘one drug, one target,
one disease’ approach that continues to dominate pharmaceutical thinking.
However, there is an increasing readiness to challenge this paradigm in favor of
the emerging network view of targets.
However, it may be that ‘the more you know, the harder it gets’.
Targets are highly sophisticated, delicate regulatory pathways and feedback loops.
But, at present, we are still mainly designing drugs that can single out and ‘hit’
certain biochemical units — the simple definable, identifiable targets.
27. Lecture WS 2008/09
Bioinformatics III
15
Nature Biotech 25, 1119 (2007)
27. Lecture WS 2008/09
Bioinformatics III
16
27. Lecture WS 2008/09
Bioinformatics III
17
27. Lecture WS 2008/09
Bioinformatics III
18
27. Lecture WS 2008/09
Bioinformatics III
19
27. Lecture WS 2008/09
Bioinformatics III
20
27. Lecture WS 2008/09
Bioinformatics III
21
27. Lecture WS 2008/09
Bioinformatics III
22
27. Lecture WS 2008/09
Bioinformatics III
23
27. Lecture WS 2008/09
Bioinformatics III
24
Outlook
This analysis of the drug-network network suggests a need to update the single
drug–single target paradigm, just as single protein–single function relations
are somewhat limited to accurately describe the reality of cellular
processes.
Future attempts at rational drug design will eventually take into account the
‘systems’ effects of a drug on the greater network upstream and downstream of
the actual drug target, which could pave the way to more specific drugs for
diseases.
27. Lecture WS 2008/09
Bioinformatics III
25
Specific example: protein kinases
Phosphorylation of Ser, Thr, and Tyr residues is a primary mechanism for
regulating protein function in eukaryotic cells. Protein kinases, the enzymes that
catalyze these reactions, regulate essentially all cellular processes and have thus
emerged as therapeutic targets for many human diseases.
What are the uses of selective inhibitors?
- Small-molecule inhibitors of the Abelson tyrosine kinase and the epidermal growth
factor receptor have been developed into clinically useful anticancer drugs.
- Selective inhibitors can also increase our understanding of the cellular and
organismal roles of protein kinases. However, nearly all kinase inhibitors target the
adenosine triphosphate (ATP) binding site, which is well conserved even among
distantly related kinase domains.
For this reason, rational design of inhibitors that selectively target even a subset
of the 491 related human kinase domains continues to be a daunting challenge.
Cohen et al. Science 308, 1318 (2005)
27. Lecture WS 2008/09
Bioinformatics III
26
Specific example: protein kinases
Structural and mutagenesis studies: a key determinant of kinase inhibitor selectivity
is a structural filter (residue) in the ATP binding site known as the „gatekeeper“.
A compact residue at this position (such as Thr in 20% of all human kinases)
allows bulky aromatic substituents, such as those found in the Src family kinase
inhibitors, PP1 and PP2, to enter a deep hydrophobic pocket.
However, such a small gatekeeper provides only partial discrimination between
kinase active sites.
In contrast, larger residues (Met, Leu, Ile, or Phe) restrict access to this pocket.
Gleevec is a drug to treat chronic myelogenous leukemia.
It exploits a Thr gatekeeper in the Abl kinase domain.
But it also potently inhibits the distantly related tyrosine kinase, c-KIT,
as well as the platelet-derived growth factor receptor (PDGFR).
27. Lecture WS 2008/09
Bioinformatics III
27
Small molecule-kinase interaction map
Competition binding assay for measuring the
interaction between unlinked, unmodified ('free')
small molecules and kinases.
(a) Schematic overview of the assay.
Blue: the phage-tagged kinase
Green: 'free' test compound in green
Red: immobilized 'bait' ligand.
(b) Binding assay for p38 MAP kinase. The
immobilized ligand was biotinylated SB202190.
(c) Determination of quantitative binding constants.
Binding of tagged p38 to immobilized SB202190
was measured as a function of unlinked test
compound concentration.
Fabian et al. Nature Biotech 23, 329 (2005)
27. Lecture WS 2008/09
Bioinformatics III
28
Small molecule-kinase interaction map
27. Lecture WS 2008/09
Bioinformatics III
29
Small molecule-kinase interaction map
Each kinase represented in the assay
panel is marked with a red circle.
TK, nonreceptor tyrosine kinases;
RTK, receptor tyrosine kinases;
TKL, tyrosine kinase-like kinases;
CK, casein kinase family;
PKA, protein kinase A family;
CAMK, calcium/calmodulin dependent
kinases;
CDK, cyclin dependent kinases;
MAPK, mitogen-activated protein kinases;
CLK, CDK-like kinases.
27. Lecture WS 2008/09
Bioinformatics III
30
Specificity profiles of clinical kinase inhibitors
Circle size is proportional to binding
affinity (on a log10 scale).
27. Lecture WS 2008/09
Bioinformatics III
31
Distribution of binding constants
For each compound the
pKd (-log Kd) was plotted for all
targets identified.
Blue: primary targets,
Red: off-targets in red.
Staurosporine does not have a particular
primary target or targets.
The primary targets for BAY-43-9006
(RAF1) and LY-333531 (PKC ) were not
part of the assay panel.
27. Lecture WS 2008/09
Bioinformatics III
32
Hierarchical cluster analysis of specificity profiles
Lighter colors correspond to
tighter interactions.
20 kinase inhibitors profiled
against a panel of 113 different
kinases.
27. Lecture WS 2008/09
Bioinformatics III
33
Summary
Presented was a systematic small molecule−kinase interaction map for clinical kinase
inhibitors with the aim of providing a more complete understanding of the biological
consequences of inhibiting particular combinations of kinases.
In future: also integrate this information with results from cell-based or animal studies,
and ultimately with clinical observations.
Binding profiles for larger numbers of chemically diverse compounds, combined with
the phenotypes elicited by these compounds in biological systems, will help identify
kinases whose inhibition leads to adverse effects, kinases that are 'safe' to inhibit and
combinations of kinases whose inhibition can have a synergistic beneficial effect in
particular disease states.
 develop inhibitors with 'appropriate' specificity that target multiple kinases involved
in the disease process while avoiding kinases implicated in side effects.
27. Lecture WS 2008/09
Bioinformatics III
34
Small molecule-kinase interaction map
The kinase binding profiles also provide valuable information to guide structural
studies.
- In many cases kinases that tightly bind the same compound have no obvious
sequence similarity, e.g., p38 and ABL(T315I) binding to BIRB-796.
- In other cases, compounds can discriminate between kinases closely related by
sequence, such as imatinib binding to LCK but not SRC.
ABL and the imatinib-resistant ABL mutants are of particular structural interest
because some compounds bind with good affinity to all forms (e.g., ZD-6474),
whereas BIRB-796 has a strong preference for a particular mutant.
Key insights should result from an analysis of selected co-crystal structures of kinasecompound combinations identified through profiling studies. Also, this large, uniform
data set may serve as a valuable training set for computation-based inhibitor design.
27. Lecture WS 2008/09
Bioinformatics III
35
Multidrug treatments are increasingly important in medicine and for probing
biological systems. But little is known about the system properties of a full drug
interaction network.
Epistasis among mutations provides a basis for analysis of gene function.
Similarly, interactions among multiple drugs provide a means to understand their
mechanism of action.
Aim: derive a pairwise drug interaction network.
Yeh et al. Nature Genetics 38, 489 (2006)
27. Lecture WS 2008/09
Bioinformatics III
36
Different ways of drug interaction
Clustering of individual drugs
into functional classes solely on
the basis of properties of their
mutual interaction network.
Schematic illustration of
additive, synergistic and
antagonistic interactions
between drugs X and Y by
measurements of bacterial
growth under the following
conditions:
no drugs, drug X only, drug Y
only, and both drugs X and Y.
27. Lecture WS 2008/09
Additive: no interaction
Synergistic: larger-than-additive effect
Antagonistic: smaller-than-additive effect
Bioinformatics III
37
Classification of drug interactions
g , gX, gXY : growth of wild-type, with drug X,
and with drugs X and Y
W XY 
W
g XY
g
  ~ XY
 W X WY
W XY  W X W Y
~
W XY  min W X , W Y  for W XY  W X W Y and 0 otherwise
For W XY  min W X , W Y 
~ 
W XY  min W X , W Y 
1  min W X , W Y 
1
This scale maps synthetic lethal interactions to  = -1,
additive interactions are mapped to  = 0,
antagonistic buffering to  = 1,
and antagonistic suppression to  > 1.
27. Lecture WS 2008/09
Bioinformatics III
38
The Prism algorithm
27. Lecture WS 2008/09
Bioinformatics III
39
Classification
(b–d) A network (b) of synergistic interactions and antagonistic interactions between drugs can be
clustered into functional classes that interact with each other monochromatically.
(d) This classification generates a system-level perspective of the drug network.
(e,f) Two independent observations indicate whether a new drug (Z) will be clustered into a particular
drug class (dashed oval): mixed synergistic and antagonistic intraclass interactions of Z with a (e, thin
dotted green and red lines) and nonconflicting interclass interactions of Z (e, dotted thin lines) and a (e,
dotted thick lines) with all other classes. Both intra and interclass indications are depicted in e, and the
drug is clustered (black arrow) with an existing class. If drug Z has no such intra- or interclass
association with any existing drug class, the drug will be clustered in a new class (f).
black circles:
drugs
red lines:
synergistic
interactions
green lines:
antagonistic
interactions
27. Lecture WS 2008/09
Bioinformatics III
40
Tested drugs
27. Lecture WS 2008/09
Bioinformatics III
41
Experimental classification of drug interaction
Experimental classification of drug interactions into 4 types using bioluminescence measurements of
bacterial growth in the presence of sublethal concentrations of antibiotics.
(a) The pairs of antibiotics illustrate synergistic interactions.
The number of bacteria (proportional to
bioluminescence counts per second (c.p.s.)
is shown from 2 replicates, for control with
no drugs (f, solid black lines), each single
drug (X, Y; blue and magenta lines) and the
double-drug combination (X + Y, dashed
black lines).
Insets: normalized growth rates (W) with
error bars for , X, Y and X+Y, from left
to right.
The interaction of piperacillin with the
50S ribosomal subunit drug erythromycin
is clearly synergistic.
27. Lecture WS 2008/09
Bioinformatics III
42
Different modes of interaction
The pairs of antibiotics illustrate synergistic (a), additive (b), antagonistic
buffering (c) and antagonistic suppression (d) interactions
27. Lecture WS 2008/09
Bioinformatics III
43
Systematic measurements of pairwise interactions between antibiotics
(a) Growth measurements and classification of interaction for all pairwise combinations of drugs X and Y.
Within each panel, the bars represent measured growth rates for, from left to right: no drugs (f), drug X only,
drug Y only and the combination of the two drugs X and Y (see inset).
Error bars represent variability in replicate measurements.
The background color of each graph designates the
form of epistasis according to the scale in b:
synergistic (red: emax < -0.5; pink: -0.5 < emax < 0.25), antagonistic buffering (green: 0.5 < emin <
1.15; light green: 0.25 < emin < 0.5), antagonistic
suppression (blue: emin > 1.15) or additive (white: 0.25 < emax < 0.5 and -0.5 < emin < 0.25). Cases that
do not fall into any of these categories are labeled
inconclusive (gray background).
27. Lecture WS 2008/09
Bioinformatics III
44
Classification into interaction classes
Unsupervised classification of the
antibiotic network into monochromatically
interacting classes of drugs with similar
mechanisms of action.
Shown is the unclustered network of
drug-drug interactions.
red: synergistic links,
green: antagonistic buffering,
blue: antagonistic suppression
27. Lecture WS 2008/09
Bioinformatics III
45
Monochromatically interacting functional classes
Prism algorithm: classifies drugs into
monochromatically interacting functional
classes.
This unsupervised clustering shows good
agreement with known functional
mechanism of the drugs (single letter
inside each node).
Bleomycin (BLM), which is believed to
affect DNA synthesis, although its
mechanism is not well understood,
cannot be clustered monochromatically
with any other class.
The multifunctional drug nitrofurantoin
(NIT) shows non-monochromatic
interactions.
27. Lecture WS 2008/09
Bioinformatics III
46
Summary
Systems analysis of the drug-drug interaction network demonstrates that drugs
can be classified according to their action mechanism based on their
interactions with other functional drug classes.
The ability to classify drug function based solely on phenotypic measurements
and without the tools of biochemistry or microscopy can provide a simple and
powerful method for screening new drugs with multiple or novel mechanisms of
action.
Applying network approaches to drug interactions may help suggest new drug
combinations and highlight the importance of gene-environment interactions,
including, in particular, the resistance and persistence of bacteria to antibiotics
and of cancer cells to antitumor drugs.
27. Lecture WS 2008/09
Bioinformatics III
47