A Study of GeneWise with the Drosophila Adh Region

1
A Study of GeneWise with the Drosophila Adh Region

• Asta Gindulyte
• CMSC 838 Presentation
• Authors Yi Mo, Moira Regelson, and Mike Sievers
• Paracel Inc., Pasadena, CA

2
Motivation

• Genome annotation
• Extraction of biologically relevant knowledge
  from raw genomic sequence data
• Need faster genome annotation methods
• DNA sequences are very long (millions of
  nucleotides)
• Current methods are computationally too expensive
• Approach/Solution
• GeneMatcher2 hardware acceleration of GeneWise

3
Outline

• Motivation
• Genome annotation
• GeneMatcher2
• Design
• ASIC hardware
• Comparison
• GeneWise algorithm
• HalfWise algorithm
• Performance (time, precision)
• Observations
• Performance improvement
• Cost effectiveness

4
Approach

• Problem make GeneWise run faster
• Embarassingly parallel algorithm
• Computationally too expensive when run in
  parallel on PCs
• Paracells solution hardware acceleration
• Dont change the algorithm
• Produce an implementation on the GeneMatcher2
  supercomputer that works as much like the
  original software as possible
• 6LITE algorithm, now also in Wise2

5
GeneMatcher Architecture
6
ASIC Hardware

• ASIC application specific integration circuit
• Designed to speed up dynamic programming
  algorithms
• (could be used for Smith-Waterman)
• Each ASIC board has 3072 processors
• System has up to 9 boards
• Cost per board around 40K

7
GeneWise Algorithm

• Perform a search of genomic DNA sequence data
  using a protein HMM
• Build HMMs from protein families
• Scan genome using HMM
• Look for start codon
• GT sequence signals possible 5 splice site
• AG sequence signals possible 3 splice site
• Dynamic programming used in the scanning process
• Obtain probability of the most likely path in HMM
  generating the sequence
• Obtain alignment by backtracking

8
GeneWise model on GeneMatcher2
9
HalfWise Algorithm

• Reduce cost by running BLAST to select HMMs with
  possible hits
• Use these HMMs with GeneWise database search and
  sequence alignment algorithm
• May miss some genes due to BLAST misses

10
Evaluation

• Test data set
• A genomic DNA sequence contig of about 2.9 Mb
  from the Drosophila Adh region
• Focuss on finding all Pfam (Protein families
  database of alignments and HMMs) protein
  profile-HMMs that occur in the Adh genomic
  sequence

11
Evaluation Speed
12
Evaluation Score
13
Evaluation Sensitivity and Specificity
14
Observations

• Performance improvement
• The speedup is several orders of magnitude.
• Makes real target applications possible
• Accuracy might be improved over HalfWise
  algorithm
• Cost effectiveness
• System used costs around 500K
• 500K worth Linux PCs (500 processors at 1K
  each) would run about 10 times slower
• Weaknesses
• Cannot modify the algorithm
• Not enough data to assess scalability