SIMDDS

1
(No Transcript)
2
Major Application Finding Homologies
(C) Mark Gerstein, Yale University
bioinfo.mbb.yale.edu/mbb452a
3
AutoSimS

• Local two-sequences alignment is the basis of
  sequence analysis, and perhaps the most widely
  used tool in computational molecular biology 1
• The parameters of most popular local sequence
  alignment tools including BLAST and FASTA are set
  by
• Default set to for the average case, which
  may not be appropriate for the sequences being
  examined
• Custom the manual settings may be difficult,
  which usually require fine tuning through several
  manual trials
• AutoSimS (Automated Sequence Similarity Search)
  contains three modules
• A modified version of SIM/DDS (Similarity /
  DNA-DNA sequence) 2, 3 for finding similar
  regions
• Adaptive simulated annealing (ASA) 4 for
  optimizing parameters for SIM/DDS
• An AI decision-making system (not implemented)
  for guiding the adaptive simulated annealing

1
4
(SIM/DDS)
Similarity / DNA-DNA Sequence

• Integrates features from Smith-Waterman, BLAST,
  Fasta and Haste (Hash-Accelerated Search) 5
• Rated as one of fastest and least space
  consuming (linear space complexity) tools for
  universal sequence alignment 6
• Provides tradeoffs between sensitivity and speed
  using over a dozen of parameters
• Our modified SIM/DDS introduces more cutoffs
• Increases flexibility of control
• Sequence filtering
• Word masking
• Reduces the impact of short and exact matches
• Allows adjusting sensitivity for weak similarity

2
5
(ASA)
Adaptive Simulated Annealing

• Uses global and statistical optimization
  techniques that are able to handle complex,
  non-linear search spaces
• Several improvements over the original simulated
  annealing technique
• Computational complexity exponential
  temperature schedule for annealing
• Completeness decreases the chance to miss
  optima
• Generality more options to better fit problems
  to be solved
• Most attractive feature individual
  considerations given to parameter range,
  annealing-time-dependent sensitivities, and the
  probability density distribution for each
  parameter
• Provides up to 100 options
• Facilitates incorporation into the AutoSimS model

3
6
AutoSimS Model
User Preferences
AI Decision-Making Module (not implemented)
Sequence Data
Data Selection
Knowledge Base
Modified SIM / DDS
Parameters
Parameter Search
Set of possible parameters with exponential
probability
Parameter Evaluation
Exponential Annealing
Value of objective function
ASA
Preferred similarity regions
4
7
Summary of Model

• ASA works as a wrapper program to select
  parameters for SIM/DDS
• With properly specified search spaces, objective
  function and successor heuristics determined by
  the AI decision-making system, ASA is used to
  find the optimal parameter setting of modified
  SIM/DDS program. This leads to finding better
  similar regions
• Even though the above mentioned information to
  be given manually to ASA, we find it easier to do
  so and let ASA tune the parameters for SIM/DDS
  than to manually tune SIM/DDSs parameters
• Adding the AI decision-making module will make
  AutoSimS nearly autonomous by automatically
  providing most of the information ASA needs

5
8
Results

• AHSC (Average of High-Scoring Chain Scores) may
  be used as an ASA objective function to find
  parameters yielding highly similar regions
• We find close-to-optimal parameter settings are
  difficult to find manually, and that there are
  many different parameter settings that yield
  close-to-optimal search results
• An automatic search for parameters may be
  effective
• Adaptive simulated annealing may be a preferred
  search technique

Three runs of our modified SIM/DDS program using
parameters selected by adaptive simulated
annealing for a 100 and 200 letter pair of DNA
sequences yield similar results, but with
different parameter settings. ASA settings
Annealing schedule T 20 exp(-0.005t) if t lt
100 and 0 otherwise Acceptance function exp(
?E / T )
6
9
Future Work

• Implement the AI decision-making system,
  including the decision analysis and knowledge
  base system
• Experiment on a large number of different types
  of molecular biological sequences to determine
  the proper annealing temperature schedules and
  successor heuristics and/or their parameters
• Parallelize AutoSimS
• Incorporate core ideas of more efficient very
  large-scale sequence comparison techniques, such
  as LSH (Locality-Sensitive Hashing) 7
• Generate statistical estimates for the local
  alignment score distributions 1, which will be
  used in AutoSimSs decision-making system
• Explore different ASA objective functions, which
  may improve results

7
10
Conclusion

• ASAs ability to fit complex functions, i.e.
  nonlinear search spaces and multiple variables,
  allows it to find a suitable set of parameters
  for SIM/DDS
• The incorporation of AI decision-making system
  to our ASA-SIM/DDS program should enhance our
  ability to achieve almost autonomous two-sequence
  similarity analysis with high volume throughput
  and acceptable performance
• Our use of simulated annealing to find a
  suitable set of parameter can be adapted to other
  bioinformatics analysis programs, such as
  alignment and clustering

8
11
References
1 Altschul, S. F., Bundschuh, R., Olsen, R. and
Hwa, T., The Estimation of Statistical Parameters
for Local Alignment Score Distributions. Nucleic
Acids Research, Vol. 29, No. 2, 351361, 2001
2 Jiang, T., Xu, Y. and Zhang, M.Q., Current
Topics in Computational Molecular Biology. MIT
Press, 2002 3 Huang, X. and Miller, W., A
Time-Efficient, Linear-Space Local Similarity
Algorithm. Advances in Applied Mathematics 12,
337357, 1991 4 Ingber, L., Simulated
Annealing Practice versus Theory. Mathl. Comput.
Modelling, Vol.18, No.11, 2957, 1993 5
Borkowski, J.A., Smith, C.P. and Huang, X., PFPA
Flexible Integrated Filtering and Masking Tool,
Paracel Inc., Pasadena, CA 6 Tech Topics,
Michigan Technological University, Nov. 3, 1995,
Vol. XXVIII, No.9 7 Buhler, J., Efficient
Large-Scale Sequence Comparison by
Locality-Sensitive Hashing. Bioinformatics 17(5)
419428, 2001
9