Mining metabolic networks for optimal drug target identification

1
Mining metabolic networks for optimal drug target
identification
2
An Example Targets for Affecting Central Nervous
System
Drug Phenylbutazone (Therapeutic category 1144)
3
Drug Discovery Process
4
Goal

• Given a set of target compounds, find the set of
  enzymes whose inhibition stops the production of
  the target compounds with minimum side-effects.

Side-effect Coming soon
5
Directed Graph Model
Edges
Vertices
Catalyzes
Enzyme
Reaction
Produces
Compound
Consumes
Target compound
6
Simple Metabolic Network
E2
R3
C5
E3
R4
R1
C2
C3
R2
E1
C4
7
Traditional Approach Inhibit E1
E2
R3
C5
E3
R4
R1
C2
C3
R2
E1
C4
8
Inhibit E2 or/and E3
E2
R3
C5
E3
R4
R1
C2
C3
R2
E1
C4
Damage (E1) 3
Damage (E2) 0 Damage (E3) 0

• What is the best enzyme combination?
• Number of combinations is exponential !

Damage (E2, E3) 1
9
OPMET An Optimal Method

• Basic branch-and-bound search strategy
• Enzyme ordering
• Filtering

10
Basic Search Strategy
E1,E2,E3,E4
4-Enzyme network
0000, 0, 0, F
n0
D INF
E1 1
E1 0
n1
n1
n2
0000, 1, 0, F
1000, 1, 5, T
D 5
E2 1
0100, 2, 1, F
n3
E3 0
E3 1
n4
n5
n4
0100, 3, 1, F
D 2
0110, 3, 2, T
E4 1
Continue Search
n6
0101, 4, 5, F
Order of the enzymes is important
11
Ordering the Enzymes

• Static priority ordering
• Dynamic priority ordering

12
Cost Model Basis of Ordering

• Takes both observed and potential damage into
  cost computation
• Weights of Enzymes, Reactions and Compounds
• W(Ei) 0, Ei is inhibited W(Ei) 1,
  otherwise
• Intuition
• A Reaction takes place only if all inputs are
  present
• A Compound vanishes only if all input reactions
  are stopped

13
Cost Computation

• Impact Vector of enzyme Ei
• I(Ei) Wi(C1), Wi(C2),..., Wi(Cn), for a
  network with n compounds.
• Normalization Vector
• V v1, v2, , vn, for a network with n
  compounds
• where
• C1, C2,, Ck Target compounds
• Ck1, Ck2,,Cn Non-target compounds

Cost of enzyme Ei
14
Static ordering
Compute cost of each enzyme individually.
Sort enzymes in ascending cost order.
Branch-and-bound search.
15
Dynamic ordering
Compute cost of each remaining enzyme
individually.
Pick the next best enzyme.
Branch-and-bound search.
16
Dynamic enzyme selection
17
Filtering the search space
18
Filtering Strategies (1/2)

• Target Filter
• Eliminate bulk of search space where there is no
  combination of enzymes that can stop production
  of all target compounds

Filter subtree
0101000000, 6, 2, F
n4
0101001111, 10, ?, F
19
Filtering Strategies (1/2)

• Non-Target Filter
• Eliminate sub trees which do not have any
  solution with damage lt best damage observed so
  far.

Filter subtree
0101000000, 6, 2, F
n4
010100XXXX, 10, D, T
D gt Dbest
20
Experiments Datasets
21
Qualitative Results
22
Benchmark Benoxaprofen (D03080) 1/2

• Inhibits arachidonate 5-lipoxygenase
  (E1.13.11.34)
• Eliminates three target compounds
• LTB4 (C02165), LTC4 (C02166), and LTE4 (C05952).
• Can not entirely eliminate
• LTD4 (C05951)
• Damage 4 compounds
• OPMET, with all the four target compounds, finds
  two different enzymes
• LTA4H (E3.3.2.6) and LTC4 synthase (E4.4.1.20)
• Damage 1 compound

23
Benoxaprofen (D03080) 2/2
OPMET
Benoxaprofen
Eliminated by benoxaprofen
Eliminated by OPMET targets
24
Benchmark Rasagiline (D02562)

• Anti-Parkinsonian
• Inhibits amine oxidase (E.1.4.3.4)
• Eliminates compounds
• Methylimidazole acetaldehyde (C05827)
• Methylimidazoleacetic acid (C05828)
• Has correlation with severity of Parkinson's
  disease
• OPMET, with the same two compounds, finds the
  same enzyme

25
Two more benchmarks

• Ozagrel (D01683)
• Inhibits thromboxane synthase (E.5.3.99.5)
• Eliminates compound thromboxane (C02198).
• Erythromycin acistrate (D02523).
• Inhibits microsomal monooxygenase (E.1.14.14.1)
• Eliminates compound diethylphosphoric acid
  (C006608).
• OPMET, with the same compound, finds the same
  enzyme

26
Quantitative Results
27
Ordering strategies (1/2)
28
Ordering strategies (2/2)
29
Filtering Strategies (1/2)
30
Filtering Strategies (2/2)
31
Iterative search
32
Filtering Strategies (2/2)
33
Initialization
34
Initialization Enzymes
E2
R3
C5
E3
R4
R1
C2
C3
R2
E1
C4
T, 3
E1 E2 E3
F, 0 F, 0

T True F False
35
Initialization Reactions
E2
R3
C5
E3
R4
R1
C2
C3
R2
E1
C4
T, 3
E1 E2 E3
E1, T, 3
R1 R2 R3 R4
F, 0 F, 0

E1, T, 3

E2, F, 0
E3, F, 0
T True F False
36
Initialization Compounds
E2
R3
C5
E3
R4
R1
C2
C3
R2
E1
C4
T, 3
E1 E2 E3
E1, T, 3
C1 C2 C3 C4 C5
E1, T, 3
R1 R2 R3 R4
F, 0 F, 0

E1, T, 3 E1, T, 3 E1, T, 3
E1, T, 3


E2, F, 0
E3, F, 0
T True F False
E2, E3, T, 1
37
Iterations
38
Iterations Reactions
E2
R3
R1 minR1, C5 min3, 1
C5
E3
R4
R1
C2
E2, E3, T, 1
R1 R2 R3 R4

C3
R2
E1
C4
E1, T, 3
C1 C2 C3 C4 C5
T, 3
E1, T, 3 E1, T, 3 E1, T, 3
E1 E2 E3
F, 0 F, 0


E2, E3, T, 1
39
Iterations Compounds
E2
R3
C1 minC1, R1 min3, 1
C5
E3
R4
R1
C2
C3
R2
E1
C4
E1, T, 3
C1 C2 C3 C4 C5
T, 3
E1, T, 3 E1, T, 3 E1, T, 3
E1 E2 E3
F, 0 F, 0


E2, E3, T, 1
40
How many iterations?
Number of iterations is at most the number of
reactions on the longest path that traverses each
node at most one
41
Experiments Datasets
42
Experiments Accuracy

• Average damage for one, two, and four randomly
  selected target compounds
• 10 10 10 runs for each network

43
Experiments Running Time
44
Experiments Number of Iterations