Gene predictions for eukaryotes attgccagtacgtagctagctacacgtatgctattacggatctgtagcttagcgtatct gtatgctgttagctgtacgtacgtatttttctagagcttcgtagtctatggctagtcgt.

Gene predictions for eukaryotes

attgccagtacgtagctagctacacgtatgctattacggatctgtagcttagcgtatctgtatgctgttagctgtacgtacgtatttttctagagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctagagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcttagtcgtgtagtcttgatctacgtacgtatttttctagagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctagagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgtacgtacgtatttttctagagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctagagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatggtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctag

Gene predictions for eukaryotes

attgccagtacgtagctagctacacgtatgctattacggatctgtagcttagcgtatctgtatgctgttagctgtacgtacgtatttttctagagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctagagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcttagtcgtgtagtcttgatctacgtacgtatttttctagagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctagagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgtacgtacgtatttttctagagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctagagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctagtcgtagtcgtagtcgttagcatctgtatggtcgtagtcgttagcatctgtatgctgttagctgtacgtacgtatttttctaggggagcttcgtagtctatggctag

Gene predictions for eukaryotes

Gene predictions for eukaryotes

Three different approaches to computational gene-finding:

Intrinsic: use statistical information about known genes (Hidden Markov Models)

Extrinsic: compare genomic sequence with known proteins / genes

Cross-species sequence comparison: search for similarities among genomes

Hidden-Markov-Models (HMM) for gene prediction

3 5 6 6 6 4 6 5 1 6 5 1 2 s

B F F U U U U U F F F F F F E φ

For sequence s and parse φ:

P(φ) probability of φ P(φ,s) joint probability of φ and s = P(φ) * P(s|φ) P(φ|s) a-posteriori probability of φ

Hidden-Markov-Models (HMM) for gene prediction

3 5 6 6 6 4 6 5 1 6 5 1 2

B F F U U U U U F F F F F F E

Goal: find path φ with maximum a-posteriori probability P(φ|s)

Equivalent: find path that maximizes joint probability P(φ,s)

Optimal path calculated by dynamic programming (Viterbi algorithm)

Hidden-Markov-Models (HMM) for gene prediction

3 5 6 6 6 4 6 5 1 6 5 1 2

B F F U U U U U F F F F F F E

Program parameters learned from training data

Hidden-Markov-Models (HMM) for gene prediction

Application to gene prediction:

A T A A T G C C T A G T C s (DNA) Z Z Z E E E E E E I I I I φ (parse)

Introns, exons etc modeled as states in GHMM („generalized HMM“)

Given sequence s, find parse that maximizes P(φ|s)

(S. Karlin and C. Burge, 1997)

AUGUSTUS

Basic model for GHMM-based intrinsic gene finding comparable to GenScan (M. Stanke)

AUGUSTUS

AUGUSTUS

AUGUSTUS

Features of AUGUSTUS:

Intron length model Initial pattern for exons Similarity-based weighting for splice sites Interpolated HMM Internal 3’ content model

Hidden-Markov-Models (HMM) for gene prediction

A T A A T G C C T A G T C s (DNA) Z Z Z E E E E I I I I φ (parse)

Explicit intron length model computationally expensive.

AUGUSTUS

Intron length model:

• Explicit length distribution for short introns• Geometric tail for long introns

Intron (fixed)

Exon

Intron (expl.)

Exon

Intron (geo.)

AUGUSTUS

AUGUSTUS+

Extension of AUGUSTUS using include extrinsic information:

Protein sequences EST sequences Syntenic genomic sequences User-defined constraints

Gene prediction by phylogenetic footprinting

Comparison of genomic sequences

(human and mouse)

Gene prediction by phylogenetic footprinting

AUGUSTUS+

Extended GHMM using extrinsic information

Additional input data: collection h of `hints’ about possible gene structure φ for sequence s

Consider s, φ and h result of random process. Define probability P(s,h,φ)

Find parse φ that maximizes P(φ|s,h) for given s and h.

AUGUSTUS+

Hints created using

Alignments to EST sequences Alignments to protein sequences Combined EST and protein alignment (EST

alignments supported by protein alignments) Alignments of genomic sequences User-defined hints

AUGUSTUS+

Alignment to EST: hint to (partial) exon

EST

G1

AUGUSTUS+

EST alignment supported by protein: hint to exon (part), start codon

EST

G1

Protein

AUGUSTUS+

Alignment to ESTs, Proteins: hints to introns, exons

ESTs, Protein

G1

AUGUSTUS+

Alignment of genomic sequences: hint to (partial) exon

G2

G1

AUGUSTUS+

Consider different types of hints:

type of hints: start, stop, dss, ass, exonpart, exon, introns

Hint associated with position i in s (exons etc. associated with right end position) max. one hint of each type allowed per position in s Each hint associated with a grade g that indicates its source.

AUGUSTUS+

hi,t = information about hint of type t at position i

hi,t = [grade, strand, (length, reading frame)] if hint available

(hints created by protein alignments contain information about reading frame)

hi,t = $ if no hint of type t available at i

AUGUSTUS+

Standard program version, without hints

A T A A T G C C T A G T C s (sequence) Z Z Z E E E E E E I I I I φ (parse)

Find parse that maximizes P(φ|s)

AUGUSTUS+

AUGUSTUS+ using hints

A T A A T G C C T A G T C s (sequence) $ $ $ $ $ $ $ X $ $ $ $ $ h (type 1) $ $ $ $ $ $ $ $ $ $ $ $ $ h (type 2) $ $ $ $ X $ $ $ $ $ $ $ $ h (type 3) . . . .

Z Z Z E E E E E E I I I I φ (parse)

Find parse that maximizes P(φ|s,h)

AUGUSTUS+

As in standard HMM theory: maximize joint probability P(φ,s,h)

How to calculate P(φ,s,h) ?

AUGUSTUS+

Simplifying assumption: Hints of different types t and at different positions i independent of each other (for redundant hints: ignore „weaker“ types).

AUGUSTUS+

Simplifying assumption: Hints of different types t and at different positions i independent of each other (for redundant hints: ignore „weaker“ types).

),|(),(),,( shPsPhsP

AUGUSTUS+

Simplifying assumption: Hints of different types t and at different positions i independent of each other (for redundant hints: ignore „weaker“ types).

),|(),(),,( shPsPhsP

ti

ti shPsP,

, ),|(),(

AUGUSTUS+

Results:

Gene (sub-)structures supported by hints receive bonus compared to non-supported structures

Gene (sub-)structures not supported by hints receive malus

(M. Stanke et al. 2006, BMC Bioinformatics)

AUGUSTUS+

AUGUSTUS+

Using hints from DIALIGN alignments:

1. Obtain large human/mouse sequence pairs (up to 50kb) from UCSC

2. Run CHAOS to find anchor points3. Run DIALIGN using CHAOS anchor points4. Create hints h from DIALIGN fragments5. Run AUGUSTUS with hints

AUGUSTUS+

Hints from DIALIGN fragments:

Consider fragments with score ≥ 20

Distinguish high scores (≥ 45) from low scores Consider reading frame given by DIALIGN Consider strand given by DIALIGN

=> 2*2*2 = 8 grades

AUGUSTUS+

EGASP competition to evaulate and compare gene-prediction methods (Sanger Center, 2005)

AUGUSTUS best ab-initio method at EGASP

EGASP test results

AUGUSTUS

GENSCAN

geneid GeneMark.hmm

Genezilla

0

10

20

30

40

50

60

70

80

90

100 Nukleotid Level

Sensitivität

Spezifität

EGASP test results

AUGUSTUS

GENSCAN

geneid GeneMark.hmm

Genezilla

0

10

20

30

40

50

60

70

80

90

100 Exon Level

Sensitivität

Spezifität

EGASP test results

AUGUSTUS

GENSCAN

geneid GeneMark.hmm

Genezilla

0

2,5

5

7,5

10

12,5

15

17,5

20

22,5

25

27,5

30 Transkript Level

Sensitivität

Spezifität

EGASP test results

AUGUSTUS

GENSCAN

geneid GeneMark.hmm

Genezilla

0

2,5

5

7,5

10

12,5

15

17,5

20

22,5

25

27,5

30 Gen Level

Sensitivität

Spezifität

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

Sn Sp Sn Sp Sn Sp Sn Sp

Base Exon Transcript Gene

Ac

cu

rac

y

AUGUSTUS

AUGUSTUS+DIALIGN

DOGFISH-C

SGP2

TWINSCAN

TWINSCAN-MARS

N-SCAN

EGASP test results

Application of AUGUSTUS in genome projects

Brugia malayi (TIGR)

Aedes aegypti (TIGR)

Schistosoma mansoni (TIGR)

Tetrahymena thermophilia (TIGR)

Galdieria Sulphuraria (Michigan State Univ.)

Coprinus cinereus (Univ. Göttingen)

Tribolium castaneum (Univ. Göttingen)