Combinatorial Library Design Using a Multiobjective Genetic Algorithm

1
Combinatorial Library Design Using a
Multiobjective Genetic Algorithm

• Valerie J. Gillet, Wael Khatib, Peter Willett,
  Peter J. Fleming, and Darren V. S. Green. J.
  Chem. Inf. Comput. Sci. 2002, 42, 375-385

Krebs Institute for Biomolecular Research and
Department of Information Studies, University of
Sheffield, Western Bank, Sheffield S10 2TN,
United Kingdom, Department of Automatic Control
and Systems Engineering, University of Sheffield,
Western Bank, Sheffield S10 2TN, United Kingdom,
and GlaxoSmithKline, Gunnels Wood Road,
Stevenage, SG1 2NY, United Kingdom
Presented by Greg Goldgof
2
Problem How to Optimize Library Design?
Solution A Multiobjective Genetic Algorithm
Based that determines Pareto Frontiers
3
Traditionally

• Library design algorithms focused on diversity.
• Failed to deliver sufficiently improved hit
  rates.
• Generated chemicals that make undesirable lead
  compounds.

4
Single-ObjectiveSELECT
5
SELECT Extended for Multiple Objectives
6
Limitations of MO Search

• The definition of the fitness function can be
  difficult especially with noncommensurable
  objectives for example, in library design it is
  not obvious how diversity should be combined with
  cost.
• The setting of the weights is nonintuitive for
  example, in the SELECT program several
  trial-and-error experiments may be required to
  choose appropriate weights.22
• The fitness function determines the regions of
  the search space that are explored, and combining
  objectives via weights can result in some regions
  being obscured.
• The progress of the search or optimization
  process is not easy to follow since there are
  many objectives to monitor simultaneously.
• (The objectives may be coupled, thus implying
  conflict and competition, which can make it more
  difficult for the optimization process to achieve
  reasonable or acceptable results.

7
Limitations of MO Search cont.

• A single solution is found which is typically one
  among a family of solutions that are all
  equivalent in terms of the overall fitness,
  although they may have different values of the
  individual objectives.
• For example, consider a two-objective problem
  where the fitness function is defined as f(n)
  ) w1x w2ywhere x and y are hypothetical
  objectives and w1 and w2 are both set to unity.
  The solution x ) 0.4, y ) 0.5 has the same
  fitness (0.9) as the potential solution x ) 0.5,
  y ) 0.4, and thus both solutions can be
  considered as equivalent typically, however,
  only one of them will be found.

8
Pareto Frontier and Dominated versus Nondominated
9
MoSELECT
Evolutionary algorithms, however, operate with a
population of individuals and are thus
well suited to search for multiple solutions in
parallel hence they can be readily adapted to
deal with multiobjective search and optimization.
10
Charting the results of MoSELECT with two
parameters
11
(No Transcript)
12
SELECT Verses MoSELECT
13
Convergence Criteria
The second convergence criterion that was
investigated involves calculating the percentage
of nondominated solutions in the Pareto set as
the search progresses This method, however,
did not prove to be effective since there was
no clear trend to indicate what a valid threshold
should be.
14
Niche Induction

• Genetic Drift or Speciation

• If the (absolute) difference in the objectives of
  the next solution and the objectives of any
  solution that already forms the center of a niche
  is within a given threshold, for all objectives,
  the fitness (or dominance) of the current
  solution is penalized otherwise it forms the
  center of a new niche. The threshold is also
  known as the niche radius.

15
Increasing the Number of Objectives

• Diversity
• Cost
• Molecular weight (MW)
• Occurrence of rotatable bonds (RB)
• Occurrence of hydrogen bond donors (HBD)
• Occurrence of hydrogen bond acceptors (HBA)
• etc.

16
No Niching
17
With Niching
18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
Future Work

• Future work will investigate the possibility of
  interacting with the search process so that the
  relationships between objectives are explored
  during the search. This will allow the user to
  observe which objectives are relatively hard to
  improve, which are more easily optimized, and
  which objectives are in competition. The search
  process itself could then be altered to take
  account of these characteristics.

22
Discussion Questions

• MoSELECT allows the user to determine the
  importance of each parameter. Is this reasonable?
• How does the Future Work present solutions to
  this problem? Does it solve it?
• Why is there no formal analysis of runtime?