Gene prediction in flies ● Background ● Gene prediction pipeline ● Resources

Background

Genome quality

Genes in Drosophila melanogaster ● high gene density ● at least 20% with alternative transripts ● can be nested  on the same strand  on different strands ● di-cistronic ● involve trans-splicing  exons from a different strand

Gene prediction pipeline ● Gene prediction by homology  no ab-initio predictions  not using genomic alignments ● TBLASTN/Genewise process  quick genome scan to find putative gene containing regions  aligning peptide sequence to genomic fragment using a gene model ● cds ● introns ● splice-sites

Sensitivity – Selectivity - Speed ● Genome scan  strict trade-off between ● sensitivity versus memory/time ● Transcript prediction  t = O(MN) ● N: length of peptide sequence = quite short ● M: length of DNA sequence = large  you want to minimize ● the length of the genomic sequence to search ● the number of fragments you align

Solutions ● ENSEMBL: Minigenes  cut out putative introns ● My pipeline:  priority lists  gene structure conservation

Difficulties ● Terminal exons  short and thus alignment signal is weak ● Spindly genes  there is no length penalty on introns

Concepts ● Predict in three passes 1)Predict clear cut cases 2)Predict dubious cases  only if they don't overlap with a previous prediction 3)Predict alternative transcripts ● Iteratively search for duplications ● Accept a prediction with conserved exon boundaries

Conservation of gene structure Query Prediction Conserved Query Prediction Partially conserved Query Prediction Single exon Query Prediction Retrotransposed Query Prediction Unconserved (exon boundaries of query/prediction mapped on query protein)

Quality control ● Classify predictions into categories  Full length or fragment  Gene or pseudogene  Conserved or not conserved gene structure ● Heuristically remove predictions  that are redundant  that are in conflict ● nested genes ● good predictions take precedence over bad predictions

Results ●

Number of predicted genes

Orthology assignments Genes in D. melanogaster with ortholgs

Technical details ● Hardware:  28 dual CPU nodes with 2Gb memory  sun grid engine (SGE) ● Pipeline logic  gmake ● Tasks  Python scripts (and Perl scripts)  Bash/awk scripts ● Database  Postgres

Downstream analysis ● Pairwise orthology assignment  PhyOP Pipeline (Leo Goodstadt (2006)) ● Multiple orthology assignment  My own concoction based on graph clustering with some consistency criteria ● Multiple alignment of cds  Dialign (<50 sequences)  Muscle (<500 sequences)

Phylogenetic analysis ● 14,000 GBlocks cleaned multiple alignments ● Calculation of ka and ks with PAML ● Phylogenetic trees  Genome trees  Gene trees  built with Fitch/Kitsch

Odds and bits ● Mapping of Pdb -> Uniprot -> dmel proteins ● Mapping of Interpro domains onto predictions  not up-to-date ● Codon bias analysis  ENC, CAI, information theoretic measures  GC3, GC3_4D

Comparison of measures Experimental CAI Computational CAI ENC GC3 Encoding | bias Encoding | unbiased Encoding | uniform Ribosomal CAI

Other groups ● see ● Gene predictions by others  Don Gilbert: SNAP  Lior Pachter: GeneMapper (genomic alignments)  Eisen Lab : TBLastN + Genewise/Exonerate, GeneMapper  Batzoglou Lab: CONTRAST  Brent Lab: N-Scan  Guigo: geneid and SGP2

summaries/genepredictions.html

Consensus predictions ● Gbrowser comparison of all gene predictions  ● Mike Eisen's group: GLEAN consensus set ● Don Gilbert: ● Other resources  tRNA predictions  genome alignments