Automatic nanodesign using evolutionary techniques

1
Automatic nanodesign using evolutionary techniques

• Al Globus, MRJ Technology Solutions, Inc. at NASA
  Ames Research Center
• John Lawton, U.C. at Santa Cruz
• Todd Wipke, U.C. at Santa Cruz

2
Problem

• Many molecules must be designed
• Often only requirements known
• Existing design techniques generally manpower
  intensive
• Genetic software can automatically design systems
• Molecular graph representation rarely supported
  by genetic software

3
Hypothesis
Crossover alone, operating on graphs, can evolve
any possible molecule given an appropriate
fitness function and a population containing both
rings and chains. This hypothesis is supported
but not proven.
4
Genetic software

• Randomly generate a set of molecules
• Many times
• Select parent molecules at random with bias
  towards better performance
• Randomly rip copies of each parent in two
• Mate opposite halves
• Replace random molecules with bias towards worse
  performance
• Repeat until satisfied

5
Algorithm properties

• Stochastic
• Embarrassingly parallel
• Robust to failure
• No guaranteed outcome
• Fitness function is crucial and non-trivial
• Performs well as cycle-scavenger
• 300-400 NAS user workstations
• 2.5 million idle workstation hours/year
• Condor, University of Wisconsin,
  http//www.cs.wisc.edu/condor

6
Previous work

• Patent US5434796, David Weininger, Daylight
  Chemical Information Systems, Inc. 1995
• two parameter crossover fragments
• some commercial success
• Robert B. Nachbar, Merck Research Laboratories,
  1998
• tree representation
• no crossover within rings
• Astro Teller, CMU, 1998?
• Neural Programming

7
(No Transcript)
8
Crossover rip in half

• Choose random bond
• Find the shortest path
• Remove and remember random path bond
• Repeat until cut set found

9
Crossover mate halves

• Select a random cut bond
• If cut bond in other half exist
• choose one at random
• merge cut bonds, respect valence
• else
• flip coin
• heads -- attach cut bond to random atom in other
  half respecting valence
• tails -- discard cut bond
• repeat until all cut bonds processed

10
Fitness function

• Given two molecules, decides which is better
• Must operate on any molecule, including very bad
  ones
• Must provide routes for evolution to reach good
  molecules
• Must make fine distinctions

11
Molecular target

• All-pairs-shortest-path (APSP) distance
• Assign extended types to each atom (element plus
  bond pattern)
• Find shortest path between each pair of atoms
• Create bag where each element is the extended
  types and length of each shortest path
• Tanimoto distance between bags
• intersection / union

12
(No Transcript)
13
(No Transcript)
14
Future work

• Drug discovery (preliminary results)
• Chemical stability fitness function
• Circuits
• Improvements in cycle-scavenging
• Data to Java server instead of file system
• Generic Java checkpointing
• problem stack not consistent format

15
Summary

• Genetic software techniques applied to molecular
  graphs
• Molecular design works but needs improvement
• Robust cycle-stealing execution
• Technique should generalize to other graph
  problems

16
Acknowledgments

• John Lawton, UCSC
• Todd Wipke, UCSC
• Creon Levit, NASA Ames
• Jason Lohn, Caelum, Inc. at NASA Ames
• Rich McClellan, UCSC
• Subash Saini, NASA Ames
• Meyya Meyyapan, NASA Ames