Value of organismspecific gene finders

1
Value of organism-specific gene finders
2
The Combiner

• Inputs any 3 or more gene finders
• GENSCAN (Burge Karlin, 1997)
• GeneMark.hmm (Lukashin Borodovsky, 1998)
• GlimmerHMM
• others...

3
Candidate Prediction Programs

• GENSCAN (Burge Karlin, 1997)
• GeneMark.hmm (Lukashin Borodovsky, 1998)
• GlimmerA (variant of GlimmerM Salzberg, Pertea,
  Gardner Tettelin, 1999)

Remember Your vote counts as much as those
people who actually know who the candidates are!
4
Statistics On The Dataset

• The gene set
• 1131 genes
• 5242 exons
• 4112 introns
• Average length of the genes/exons/introns
• 1048 nt / gene
• 226 nt / exon
• 161 nt / intron
• Maximum number of introns 33 for gene At2g34680

5
Evaluating Performance
6
Metrics

• Sensitivity (Sn) true positives / actual coding
• Precision (Pr) true positives / reported as
  coding
• (SnPr)/2
• Percentage of genes exactly correct, end to end

7
Correctly Predicted Exons
Exons
8
Oracles

• The GeneFinder Switch Oracle picks the best gene
  prediction for each sequence
• The Maximum Precision Oracle picks from the
  exons predicted by all gene finders exactly those
  that are correct

9
Limits on any combiner (Arabidopsis)
Precision
Exon Level Sensitivity
10
Distribution of Correct Exon Predictions among
the Genefinders
Exons
Exons
GA-GM
GS-GA
GM-GS
GM-GA
GA-GS
GS-GM
Legend GS GENSCAN GM GeneMark.hmm GA
GlimmerA
11
Majority Voting Algorithm

• 1 vote for each gene finder
• Establish a label for each base by voting for it
  as coding or noncoding
• Labels comprise a mosaic prediction (not a valid
  gene model)
• Choose the prediction that is closest to the
  voted mosaic

12
Combiner Performance
Precision
Exon Level Sensitivity
13
Correct exons and whole genes