1. Eukaryotic Gene Finding
Adapted in part from

2. Prokaryotic vs. Eukaryotic Genes
Prokaryotes
small genomes
high gene density
no introns (or splicing)
no RNA processing
similar promoters
overlapping genes
Eukaryotes
large genomes
low gene density
introns (splicing)
RNA processing
heterogeneous promoters
polyadenylation

4. Pre-mRNA Splicing ... ... U 1 s n R N P 2 intronic repressor 5 ’
splice signal U 2 A F 6 5 3 1 s n R N P
SR proteins
intron definition
exon definition
exonic enhancers 5 ’
splice signal 3
polyY
branch signal
intronic enhancers
exonic repressor
...
(assembly of
spliceosome,
catalysis)
...

6. Some Statistics On average, a vertebrate gene is about 30KB long
Coding region takes about 1KB
Exon sizes can vary from double digit numbers to kilobases
An average 5’ UTR is about 750 bp
An average 3’UTR is about 450 bp but both can be much longer.

7. Human Splice Signal Motifs

10. Semi-Markov HMM Model

11. GHMM A finite Set Q of states Initial state distribution Π
Transition probabilities Ti,j for
Length distribution f of the states (fq is the length distribution of state q)
Probability model for each state

12. GHMM – contin.
A parse Ф of a sequence S of length L is an ordered sequence of states (q1, , qt) with an associated duration di to each state
The most probable pass Фopt can be computed as in Veterbi algorithm

13. Genscan HSMM

14. GenScan States N - intergenic region P - promoter
F - 5’ untranslated region
Esngl – single exon (intronless) (translation start -> stop codon)
Einit – initial exon (translation start -> donor splice site)
Ek – phase k internal exon (acceptor splice site -> donor splice site)
Eterm – terminal exon (acceptor splice site -> stop codon)
Ik – phase k intron: 0 – between codons; 1 – after the first base of a codon; 2 – after the second base of a codon

15. GenScan features Model both strands at once
Each state may output a string of symbols (according to some probability distribution).
Explicit intron/exon length modeling
Advanced splice site modeling
Complete intron/exon annotation for sequence
Able to predict multiple genes and partial/whole genes
Parameters learned from annotated genes
Separate parameter training for different CpG content groups (< 43%, 43-51%, 51-57%,>57% CG content)

16. Various parameters in GENSCAN

18. GenScan Signal Modeling
PSSM: P(S) = P1(S1)•P2(S2) •…•Pn(Sn)
PolyA signal
Translation initiation/termination signal
Promoters
WAM: P(S) = P1(S1) •P2(S2|S1)•…•Pn(Sn|Sn-1)
5’ and 3’ splice sites

19. GENSCAN Performance
> 80% correct exon predictions, and > 90% correct coding/non coding predictions by bp.
BUT - the ability to predict the whole gene correctly is much lower

20. HMM-based Gene Finding
GENSCAN (Burge 1997)
FGENESH (Solovyev 1997)
HMMgene (Krogh 1997)
GENIE (Kulp 1996)
GENMARK (Borodovsky & McIninch 1993)
VEIL (Henderson, Salzberg, & Fasman 1997)

21. Using Sequence Similarity for Gene Finding
Compare genomic sequence with expressed sequence tags (ESTs) (e.g. by BLASTN), to identify regions corresponding to processed mRNA
Compare genomic sequence to Protein DB (e.g. by BLASTX), to identify probably coding regions
“Spliced Alignment” of genomic sequence of a complete gene with a homologous protein sequence (e.g. by PROCRUSTES) may enable exon/intron reconstruction
Compare predicted peptides (e.g. by GENSCAN) with protein DB to assign confidence to predictions and functional annotations
Compare Genomic sequence with homologous from close organisms/species (e.g. by BLAST, CLASTW), to identify conserved regions which might correspond to coding regions and DNA signals
“Each of these methods can provide useful information about gene locations, as well as clues to gene function, although similarity based methods are currently (1998) able to identify only about hald of all human genes, and this proportion is increasing rather slowly. It should be kept in mind that similarity bsed mehtods are only as reliable as the DB that are searched, and apparent homology can be misleading at times…” (from Burge review, 98)

22. GenomeScan proteins are available.
Idea: We can enhance our gene prediction by using external information: DNA regions with homology to known proteins are more likely to be coding exons.
Combine probabilistic ‘extrinsic’ information (BLAST hits) with a probabilistic model of gene structure/composition (GenScan)
Focus on ‘typical case’ when homologous but not identical
proteins are available.

25. GeneWise [Birney, Amitai]
Motivation: Use good DB of protein world (PFAM) to help us annotate genomic DNA
GeneWise algorithm aligns a profile HMM directly to the DNA

26. Sample GeneWise Output

27. Developing GeneWise Model
Start with a PFAM domain HMM
Replace AA emissions with codon emissions
Allow for sequencing errors (deletions/insertions)
Add a 3-state intron model

28. GeneWise Model

29. GeneWise Intron Model
PY tract
central
spacer
5’ site
3’ site

30. GeneWise Model
Viterbi algorithm -> “best” alignment of DNA to protein domain
Alignment gives exact exon-intron boundaries
Parameters learned from species-specific statistics

31. GeneWise problems
Only provides partial prediction, and only where the homology lies
Does not find “more” genes
Pseudogenes, Retrotransposons picked up
CPU intensive
Solution: Pre-filter with BLAST
Retrotransposons are explained in p in Genetic Analysis, 5th edition:
Basically parts of the genome which are copied multiple times into the DNA genome by means of reverese transcription (from mRNA back into the DNA).
Good example is the ~200bp long human Alu sequnence, which we have hunderds of thousends of copies of, making ~5% of our genome.
This is an example of SINES ( short interspersed elements).
There are also LINES ( long interspersed elements), 1-5kb long, 20k-40k copies of them in the human genome.
Elements of this class have ORFs that potentially code for enyzmes used in transposition.
Many are the result of RNA virus that replicate through a DNA stage which can integrate into host chromosomes.
Pseudogenes are explained in p.480 of the same book: basically these were once copies of a gene that stopped being functional (transcribed), and therefore is now under no selective pressure  it is still very similar to “real” genes, but with many more mutations.

32. Other Sequence Usage
Search translated genomic sequences for the occurrences of the shot peptide motifs that are characteristic of common protein families (e.g. zinc finger, ATP/GTP binding modifs etc.)
Identify sequences which are probably NON coding: identify known classes of interspersed repeates (e.g LINE SINE) in none coding regions. Can be essential to remove these before simple BLAST is done against EST’s.

33. Summary
Genes are complex structures which are difficult to predict with the required level of accuracy/confidence
Different approaches to gene finding:
Ab Initio : GenScan
Ab Initio modified by BLAST homologies: GenomeScan
Homology guided: GeneWise

34. Future Directions
Find genes not for proteins (tRNA, rRNA, smRNA) – hard !
Deal better with overlapping genes, multiple genes in a single sequence
Alternative splicing/transcription/translation – a whole separate issue
The mechanisms governing it, the signals
predicting the various genes
Very important !