GenePC and Aspic

1
GenePC and Aspic
Integrating gene predictions with EST alignments
to predict alternative transcripts

• Tyler Alioto
• Center for Genomic Regulation
• Barcelona, Spain

2
The Splicing Code
3
Problem with Combiners

• Combiners generally look for consensus, not
  alternatives
• SOLUTION
• Provide mutually incompatible constraints which
  are known to be of high quality and run
  prediction iteratively or in parallel with each
  set of constraints.

4
Problem with EST-based transcript predictors

• Quality of predicted transcripts dependent on
  often poor EST sequence quality
• Transcripts can be incomplete
• SOLUTION
• Fill in with ab initio exon predictions

5
Advantages of Aspic-GenePC combo

• Aspic performs high-quality multiple alignment of
  ESTs to a genomic locus, reducing number of
  artifacts
• GenePC is well suited for incorporating introns
  as evidence
• Aspic introns are incorporated if and only if
  they are compatible with the gene prediction
  evidence

6
(No Transcript)
7
The Aspic algorithm

• The MEFC problem
• Minimum EST Factorization Compatible with a
  genomic sequence
• Optimal solution
• Minimize the number of distinct pseudo-exons in
  the gene-factorization of the genomic sequence

8
Aspic optimization
9
Intron Boundary Refinement
Example of intron detection in the human ATP1B1
(UGHs.291196) gene without (A) or with (B) the
refinement of exon-intron boundaries. The first
row shows the genomic sequence aligned to the EST
sequences (below). In (A) four different introns
are detected (A, B, C, D) that can be merged to
only two (A, D) in B. Absolute coordinate (NCBI
35 assembly) are shown for each intron and
acceptor/donor splice sites are in
black-background. Bonizzoni et al. BMC
Bioinformatics 2005 6244   doi10.1186/1471-2105-
6-244
10
GenePC uses GeneID architecture

• GeneID follows a hierarchical structure
• signal ? exon ? gene
• GenePC exon scoring replaces GeneID exon scoring
• Dynamic programming algorithm
• max score of assembled exons ? assembled gene

predicted exons
Aspic introns
11
GenePC weighted linear combination mode
12
Combining gene predictions

• Method 1
• Ad hoc linear combination of 3 factors

Normalized self- reported scores
Sum of distances between programs
Performance of programs
13
GenePC weighted linear combination mode
14
Combining gene predictions

• Method 2 combining exon probabilities
• c self-reported confidence
• SPe exon-level specificity
• rij (SNe SPe)/2 bewtween prediction sets i
  and j
• Correlation matrix R treated as a distance matrix
  for UPGMA tree.
• Terminal branches x and y with max r are
  collapsed
• R is updated with the r of resulting branch with
  each remaining branch z
• recalculated as average of rxz and ryz
• Calculate p for node
• Repeat until only one branch left

15
Final Score

• Expressed as log-likelihood ratio

16
Results ad hoc method

• 5 point gain in average sensitivity and
  specificity at the transcript level

17
GenePC Example VPS25
18
Results ad hoc method

• Increased sensitivity offset by loss in
  specificity
• Probably due in part to introns in non-coding
  regions
• Intron sensitivity is higher
• Driven by prediction of more transcripts per
  locus, not refinement of transcript predictions

19
Acknowledgments

• GENEID TEAM
• Roderic Guigó
• Enrique Blanco
• Genís Parra
• Francisco Camara

• ASPIC TEAM
• Graziano Pesole
• Ernesto Picardi
• Paola Bonizzoni
• Raffaella Rizzi