Bioinformatics t8-go-hmm v2014

FBW
2-12-2014
Wim Van Criekinge

Gene Prediction, HMM & ncRNA
What to do with an unknown
sequence ?
Gene Ontologies
Gene Prediction
Composite Gene Prediction
Non-coding RNA
HMM

UNKNOWN PROTEIN SEQUENCE
LOOK FOR:
• Similar sequences in databases ((PSI)
BLAST)
• Distinctive patterns/domains associated
with function
• Functionally important residues
• Secondary and tertiary structure
• Physical properties (hydrophobicity, IEP
etc)

BASIC INFORMATION COMES FROM SEQUENCE
• One sequence- can get some information eg
amino acid properties
• More than one sequence- get more info on
conserved residues, fold and function
• Multiple alignments of related sequences-can
build up consensus sequences of known
families, domains, motifs or sites.
• Sequence alignments can give information
on loops, families and function from
conserved regions

Additional analysis of protein sequences
• transmembrane
regions
• signal sequences
• localisation
signals
• targeting
sequences
• GPI anchors
• glycosylation sites
• hydrophobicity
• amino acid
composition
• molecular weight
• solvent accessibility
• antigenicity

FINDING CONSERVED PATTERNS IN PROTEIN SEQUENCES
• Pattern - short, simplest, but limited
• Motif - conserved element of a sequence
alignment, usually predictive of structural or
functional region
To get more information across whole
alignment:
• Profile
• HMM

PATTERNS
• Small, highly conserved regions
• Shown as regular expressions
Example:
[AG]-x-V-x(2)-x-{YW}
– [] shows either amino acid
– X is any amino acid
– X(2) any amino acid in the next 2 positions
– {} shows any amino acid except these
BUT- limited to near exact match in small
region

PROFILES
• Table or matrix containing comparison
information for aligned sequences
• Used to find sequences similar to
alignment rather than one sequence
• Contains same number of rows as
positions in sequences
• Row contains score for alignment of
position with each residue

HIDDEN MARKOV MODELS (HMM)
HMM
• An HMM is a large-scale profile with gaps,
insertions and deletions allowed in the
alignments, and built around probabilities
• Package used HMMER (http://hmmer.wusd.edu/)
• Start with one sequence or alignment -HMMbuild,
then calibrate with HMMcalibrate, search
database with HMM
• E-value- number of false matches expected with
a certain score
• Assume extreme value distribution for noise,
calibrate by searching random seq with HMM
build up curve of noise (EVD)

sequence ?
Gene Ontologies
Gene Prediction
HMM
Non-coding RNA

What is an ontology?
• An ontology is an explicit
specification of a conceptualization.
• A conceptualization is an abstract,
simplified view of the world that we
want to represent.
• If the specification medium is a
formal representation, the ontology
defines the vocabulary.

Why Create Ontologies?
• to enable data exchange among
programs
• to simplify unification (or translation)
of disparate representations
• to employ knowledge-based services
• to embody the representation of a
theory
• to facilitate communication among
people

Summary
• Ontologies are what they do:
artifacts to help people and their
programs communicate, coordinate,
collaborate.
• Ontologies are essential elements in
the technological infrastructure of
the Knowledge Age
• http://www.geneontology.org/

The Three Ontologies
•Molecular Function — elemental activity or task
nuclease, DNA binding, transcription factor
•Biological Process — broad objective or goal
mitosis, signal transduction, metabolism
•Cellular Component — location or complex
nucleus, ribosome, origin recognition complex

DAG Structure
Directed acyclic graph: each child
may have one or more parents

GO: Applications
• Eg. chip-data analysis: Overrepresented item
can provide functional clues
• Overrepresentation check: contingency table
– Chi-square test (or Fisher is frequency < 5)

sequence ?
Web applications
Gene Ontologies
Gene Prediction
HMM
Non-coding RNA

Computational Gene Finding
Problem:
Given a very long DNA sequence, identify coding
regions (including intron splice sites) and their
predicted protein sequences

Eukaryotic gene structure

• There is no (yet known) perfect method
for finding genes. All approaches rely on
combining various “weak signals”
together
• Find elements of a gene
– coding sequences (exons)
– promoters and start signals
– poly-A tails and downstream signals
• Assemble into a consistent gene model

GENE STRUCTURE INFORMATION - POSITION ON PHYSICAL MAP
This gene structure corresponds to the position on the physical map

The Active Zone limits the extent of
analysis, genefinder & fasta dumps
A blue line within the yellow box
indicates regions outside of the active
zone
The active zone is set by entering
coordinates in the active zone (yellow
box)
GENE STRUCTURE INFORMATION - ACTIVE ZONE
This gene structure shows the Active Zone

GENE STRUCTURE INFORMATION - POSITION
This gene structure relates to the Position:
Change origin of
this scale by
entering a
number in the
green 'origin'
box

Boxes are Exons,
thin lines (or
springs) are Introns
GENE STRUCTURE INFORMATION - PREDICTED GENE STRUCTURE
This gene structure relates to the predicted gene structures

Find the open reading frames
The triplet, non-punctuated nature of the genetic code helps us out
64 potential codons
61 true codons
3 stop codons (TGA, TAA, TAG)
Random distribution app. 1/21 codons will be a stop
Any sequence has 3 potential reading frames (+1, +2, +3)
Its complement also has three potential reading frames (-1, -2, -3)
6 possible reading frames
GAAAAAGCTCCTGCCCAATCTGAAATGGTTAGCCTATCTTTCCACCGT
E K A P A Q S E M V S L S F H R
K K L L P N L K W L A Y L S T
K S S C P I * N G * P I F P P

There is one column
for each frame
Small horizontal
lines represent stop
codons
GENE STRUCTURE INFORMATION - OPEN READING FRAMES
This gene structure relates to Open reading Frames

They have one
column for each
frame
The size indicates
relative score for the
particular start site
GENE STRUCTURE INFORMATION - START CODONS
This gene structure represents Start Codons

Computational Gene Finding: Hexanucleotide frequencies
• Amino acid distributions are biased
e.g. p(A) > p(C)
• Pairwise distributions also biased
e.g. p(AT)/[p(A)*p(T)] > p(AC)/[p(A)*p(C)]
• Nucleotides that code for preferred amino
acids (and AA pairs) occur more frequently in
coding regions than in non-coding regions.
• Codon biases (per amino acid)
• Hexanucleotide distributions that reflect those
biases indicate coding regions.

Gene prediction
Generation of datasets (Ensmart@Ensembl):
Dataset 1 (http://biobix.ugent.be/txt/coding.txt) consists of >900
coding regions (DNA):
Dataset 2 (http://biobix.ugent.be/txt/noncoding.txt) consists of
>900 non-coding regions
Distance Array: Calculate for every base all the distances (in
bp) to the same nucleotide (focus on the first 1000 bp of the
coding region and limit the distance array to a window of
1000 bp)
Do you see a difference in this “distance array” between coding
and noncoding sequence ?
Could it be used to predict genes ?
Write a program to predict genes in the following genomic
sequence (http://biobix.ugent.be/txt/genomic.txt)
What else could help in finding genes in raw genomic
sequences ?

The grey boxes indicate
regions where the codon
frequencies match those of
known C. elegans genes.
the larger the grey box the
more this region resembles a
C. elegans coding element
GENE STRUCTURE INFORMATION - CODING POTENTIAL
This gene structure corresponds to the Coding Potential

blastn (EST)
For raw DNA sequence analysis blastx is
extremely useful
Will probe your DNA sequence against the protein database
A match (homolog) gives you some ideas regarding function
One problem are all of the genome sequences
Will get matches to genome databases that are strictly identified by
sequence homology – often you need some experimental evidence

The blue boxes indicate
regions of sequence which
when translated have
similarity to previously
characterised proteins.
To view the alignment,
select the right mouse
button whilst over the blue
box.
GENE STRUCTURE INFORMATION - SEQUENCE SIMILARITY
This feature shows protein sequence similarity

The yellow boxes represent
DNA matches (Blast) to C.
elegans Expressed Sequence
Tags (ESTS)
To view the alignment use the
right mouse button whilst
over the yellow box to invoke
Blixem
GENE STRUCTURE INFORMATION - EST MATCHES
This gene structure relates to Est Matches

New generation of programs to predict gene coding
sequences based on a non-random repeat pattern
(eg. Glimmer, GeneMark) – actually pretty good
Borodovsky et al., 1999, Organization of the Prokaryotic Genome (Charlebois, ed) pp. 11-34

• CpG islands are regions of sequence that
have a high proportion of CG dinucleotide
pairs (p is a phoshodiester bond linking
them)
– CpG islands are present in the promoter and
exonic regions of approximately 40% of
mammalian genes
– Other regions of the mammalian genome contain
few CpG dinucleotides and these are largely
methylated
• Definition: sequences of >500 bp with
– G+C > 55%
– Observed(CpG)/Expected(CpG) > 0.65

This column shows
matches to members of a
number of repeat families
Currently a hidden markov
model is used to detect
these
GENE STRUCTURE INFORMATION - REPEAT FAMILIES
This gene structure corresponds to Repeat Families

This column shows regions
of localised repeats both
tandem and inverted
Clicking on the boxes will
show the complete repeat
information in the blue line
at the top end of the screen
GENE STRUCTURE INFORMATION - REPEATS
This gene structure relates to Repeats

Computational Gene Finding: Splice junctions
• Most Eukaryotic introns have a
consensus splice signal: GU at the
beginning (“donor”), AG at the end
(“acceptor”).
• Variation does occur in the splice sites
• Many AGs and GTs are not splice sites.
• Database of experimentally validated
human splice sites:
http://www.ebi.ac.uk/~thanaraj/splice.h
tml

The Splice Sites are shown
'Hooked'
The Hook points in the
direction of splicing, therefore
3' splice sites point up and 5'
Splice sites point down
The colour of the Splice Site
indicates the position at which
it interrupts the Codon
The height of the Splices is
proportional to the Genefinder
score of the Splice Site
GENE STRUCTURE INFORMATION - PUTATIVE SPLICE SITES
This gene structure shows putative splice sites

Towards profiles (PSSM) with indels – insertions and/or deletions
• Recall that profiles are matrices that
identify the probability of seeing an
amino acid at a particular location in a
motif.
• What about motifs that allow insertions
or deletions (together, called indels)?
• Patterns and regular expressions can
handle these easily, but profiles are
more flexible.
• Can indels be integrated into profiles?

Hidden Markov Models: Graphical models of sequences
• Need a representation that allows
specification of the probability of
introducing (and/or extending) a gap in
the profile.
A .1
C .05
D .2
E .08
F .01
continue
Gap A .04
C .1
D .01
E .2
F .02
Gap A .2
C .01
D .05
E .1
F .06
delete

Hidden Markov Chain
• A sequence is said to be Markovian if the
probability of the occurrence of an element in
a particular position depends only on the
previous elements in the sequence.
• Order of a Markov chain depends on how
many previous elements influence probability
– 0th order: uniform probability at every position
– 1st order: probability depends only on immediately
previous position.
• 1st order Markov chains are good for proteins.

Markov chain with begin and end

Markov Models: Graphical models of sequences
• Consists of states (boxes) and transitions
(arcs) labeled with probabilities
• States have probability(s) of “emitting” an
element of a sequence (or nothing).
• Arcs have probability of moving from one
state to another.
– Sum of probabilities of all out arcs must be 1
– Self-loops (e.g. gap extend) are OK.

Markov Models
• Simplest example: Each state emits (or,
equivalently, recognizes) a particular
element with probability 1, and each
transition is equally likely.
Begi
n
Emit 1
Emit 2
Emit 4
Emit 3
End
Example sequences: 1234 234 14 121214 2123334

Hidden Markov Models: Probabilistic Markov Models
• Now, add probabilities to each transition (let
emission remain a single element)
0.5
Begi 1.0
n
0.5
0.25
0.9
• We can calculate the probability of any sequence given this
model by multiplying
0.75
0.1
0.2
0.8
Emit 1
Emit 2
Emit 4
Emit 3
End
p(1234) = 0.5 * 0.1 * 0.75 * 0.8 = 0.03
p(14) = 0.5 * 0.9 = 0.45
p(2334)= 0.5 * 0.75 * 0.2 * 0.8 = 0.06

• If we let the states define a set of emission
probabilities for elements, we can no longer be
sure which state we are in given a particular
element of a sequence
0.9
Begi 1.0
n
BCCD or BCCD ?
0.5
0.5
0.25
0.75
0.1
0.2
0.8
A (0.8) B(0.2)
B (0.7) C(0.3)
C (0.1) D (0.9)
C (0.6) A(0.4)
End
Hidden Markov Models: Probablistic Emmision

• Emission uncertainty means the sequence doesn't
identify a unique path. The states are “hidden”
0.5
Begi 1.0
n
0.5
0.9
• Probability of a sequence is sum of all paths that can
produce it:
0.25
0.75
0.1
0.2
0.8
A (0.8) B(0.2)
B (0.7) C(0.3)
C (0.1) D (0.9)
C (0.6) A(0.4)
End
p(bccd) = 0.5 * 0.2 * 0.1 * 0.3 * 0.75 * 0.6 * 0.8 * 0.9
+ 0.5 * 0.7 * 0.75 * 0.6 * 0.2 * 0.6 * 0.8 * 0.9
= 0.000972 + 0.013608 = 0.01458
Hidden Markov Models

Hidden Markov Models: The occasionally dishonest casino

Use of Hidden Markov Models
• The HMM must first be “trained” using a training set
– Eg. database of known genes.
– Consensus sequences for all signal sensors are needed.
– Compositional rules (i.e., emission probabilities) and
length distributions are necessary for content sensors.
• Transition probabilities between all connected
states must be estimated.
• Estimate the probability of sequence s, given model
m, P(s|m)
– Multiply probabilities along most likely path
(or add logs – less numeric error)

Applications of Hidden Markov Models
• HMMs are effectively profiles with gaps, and
have applications throughout Bioinformatics
• Protein sequence applications:
– MSAs and identifying distant homologs
E.g. Pfam uses HMMs to define its MSAs
– Domain definitions
– Used for fold recognition in protein structure
prediction
• Nucleotide sequence applications:
– Models of exons, genes, etc. for gene
recognition.

Hidden Markov Models Resources
• UC Santa Cruz (David Haussler group)
– SAM-02 server. Returns alignments, secondary
structure predictions, HMM parameters, etc. etc.
– SAM HMM building program
(requires free academic license)
• Washington U. St. Louis (Sean Eddy group)
– Pfam. Large database of precomputed HMM-based
alignments of proteins
– HMMer, program for building HMMs
• Gene finders and other HMMs (more later)

Example TMHMM
Beyond Kyte-Doolitlle …

HMM in protein analysis
• http://www.cse.ucsc.edu/research/compbio/is
mb99.handouts/KK185FP.html

Hidden Markov model for gene structure
Signals (blue nodes):
• begin sequence (B)
• start translation (S)
• donor splice site (D)
• acceptor splice site (A)
• stop translation (T)
• end sequence (F)
Contents (red arcs):
• 5’ UTR (J5’)
• initial exon (EI)
• exon (E)
• intron (I)
• final exon (EF)
• single exon (ES)
• 3’ UTR (J3’)
• A representation of the linguistic rules for what features might follow
what other features when parsing a sequence consisting of a multiple
exon gene.
• A candidate gene structure is created by tracing a path from B to F.
• A hidden Markov model (or hidden semi-Markov model) is defined by
attaching stochastic models to each of the arcs and nodes.

Classic Programs for gene finding
Some of the best programs are HMM based:
• GenScan – http://genes.mit.edu/GENSCAN.html
• GeneMark – http://opal.biology.gatech.edu/GeneMark/
Other programs
• AAT, EcoParse, Fexeh, Fgeneh, Fgenes, Finex, GeneHacker, GeneID-3,
GeneParser 2, GeneScope, Genie, GenLang, Glimmer, GlimmerM, Grail
II, HMMgene, Morgan, MZEF, Procrustes, SORFind, Veil, Xpound

Hidden Markov Models: Gene Finding Software
• A Semi-Markov Model
GENSCAN
not to be confused with GeneScan, a commercial product
– Explicit model of how long
to stay in a state (rather
than just self-loops, which
must be exponentially
decaying)
• Tracks “phase” of exon or
intron (0 coincides with codon
boundary, or 1 or 2)
• Tracks strand (and direction)

Conservation of Gene Features
100%
95%
90%
85%
80%
75%
70%
65%
60%
55%
50%
aligning identity
Conservation pattern across 3165 mappings of human
RefSeq mRNAs to the genome. A program sampled 200
evenly spaced bases across 500 bases upstream of
transcription, the 5’ UTR, the first coding exon, introns,
middle coding exons, introns, the 3’ UTR and 500 bases
after polyadenylatoin. There are peaks of conservation at the
transition from one region to another.

Composite Approaches
• Use EST info to constrain HMMs (Genie)
• Use protein homology info on top of HMMs
(fgenesh++, GenomeScan)
• Use cross species genomic alignments on top
of HMMs (twinscan, fgenesh2, SLAM, SGP)

Gene Prediction: more complex …
1. Species specific
2. Splicing enhancers found in coding regions
3. Trans-splicing
4. …

Length preference
5’ ss intcomp branch 3’ ss

Contents-Schedule
RNA genes
Besides the 6000 protein coding-genes, there is:
140 ribosomal RNA genes
275 transfer RNA gnes
40 small nuclear RNA genes
>100 small nucleolar genes
?
pRNA in 29 rotary packaging motor (Simpson
et el. Nature 408:745-750,2000)
Cartilage-hair hypoplasmia mapped to an RNA
(Ridanpoa et al. Cell 104:195-203,2001)
The human Prader-Willi ciritical region (Cavaille
et al. PNAS 97:14035-7, 2000)

RNA genes can be hard to detects
UGAGGUAGUAGGUUGUAUAGU
C.elegans let-27; 21 nt
(Pasquinelli et al. Nature 408:86-89,2000)
Often small
Sometimes multicopy and redundant
Often not polyadenylated
(not represented in ESTs)
Immune to frameshift and nonsense mutations
No open reading frame, no codon bias
Often evolving rapidly in primary sequence
miRNA genes

Lin-4
• Lin-4 identified in a screen for mutations that affect timing and
sequence of postembryonic development in C.elegans. Mutants re-iterate
L1 instead of later stages of development
• Gene positionally cloned by isolating a 693-bp DNA fragment that
can rescue the phenotype of mutant animals
• No protein found but 61-nucleotide precursor RNA with stem-loop
structure which is processed to 22-mer ncRNA
• Genetically lin-4 acts as negative regulator of lin-14 and lin-28
• The 3’ UTR of the target genes have short stretches of
complementarity to lin-4
• Deletion of these lin-4 target seq causes unregulated gof phenotype
• Lin-4 RNA inhibits accumulation of LIN-14 and LIN-28 proteins
although the target mRNA

Let-7
(Pasquinelli et al. Nature 408:86-89,2000)
Let-7 (lethal-7) was also mapped to a ncRNA gene with a 21-
nucleotide product
The small let-7 RNA is also thought to be a post-transcriptional
negative regulator for lin-41 and lin-42
100% conserved in all bilaterally symmetrical animals (not
jellyfish and sponges)
Sometimes called stRNAs, small temporal RNAs

Two computational analysis problems
• Similarity search (eg BLAST), I give you a query,
you find sequences in a database that look like the
query (note: SW/Blat)
– For RNA, you want to take the secondary structure of
the query into account
• Genefinding. Based solely on a priori knowledge
of what a “gene” looks like, find genes in a
genome sequence
– For RNA, with no open reading frame and no codon
bias, what do you look for ?

Context-free grammers
Basic CFG
“production rules”
S -> aS
S -> Sa
S -> aSu
S -> SS
A CFG “derivation”
S -> aS

Basic CFG
S -> aS
S -> Sa
S -> aSu
S -> SS
S -> aS
S -> aaS

Basic CFG
S -> aS
S -> Sa
S -> aSu
S -> SS
S -> aS
S -> aaS
S -> aaSS

Basic CFG
S -> aS
S -> Sa
S -> aSu
S -> SS
S -> aS
S -> aaS
S -> aaSS
S -> aagScuS

Basic CFG
S -> aS
S -> Sa
S -> aSu
S -> SS
S -> aS
S -> aaS
S -> aaSS
S -> aagScuS
S -> aagaSucugSc

Basic CFG
S -> aS
S -> Sa
S -> aSu
S -> SS
S -> aS
S -> aaS
S -> aaSS
S -> aagScuS
S -> aagaSucugSc
S -> aagaSaucuggScc
S -> aagacSgaucuggcgSccc

Basic CFG
S -> aS
S -> Sa
S -> aSu
S -> SS
S -> aS
S -> aaS
S -> aaSS
S -> aagScuS
S -> aagaSucugSc
S -> aagaSaucuggScc
S -> aagacuSgaucuggcgSccc
S -> aagacuuSgaucuggcgaSccc
S -> aagacuucSgaucuggcgacSccc
S -> aagacuucgSgaucuggcgacaSccc
S -> aagacuucggaucuggcgacaccc

Basic CFG
S -> aS
S -> Sa
S -> aSu
S -> SS
G
C
S -> aS
A
S -> aaS
S -> aaSS
S -> aagScuS
S -> aagaSucugSc
S -> aagaSaucuggScc
S -> aagacuSgaucuggcgSccc
S -> aagacuuSgaucuggcgaSccc
S -> aagacuucSgaucuggcgacSccc
S -> aagacuucgSgaucuggcgacaSccc
S -> aagacuucggaucuggcgacaccc
C
G
U
*
A
A
A
A
A
G
G G G
C
C
C
C C C
U
U
U
*
*
* * *

The power of comparative analysis
• Comparative genome analysis is an indispensable means of
inferring whether a locus produces a ncRNA as opposed to
encoding a protein.
• For a small gene to be called a protein-coding gene, one
excellent line of evidence is that the ORF is significantly
conserved in another related species.
• It is more difficult to positively corroborate a ncRNA by
comparative analysis but, in at least some cases, a ncRNA
might conserve an intramolecular secondary structure and
comparative analysis can show compensatory base
substitutions.
• With comparative genome sequence data now
accumulating in the public domain for most if not all
important genetic systems, comparative analysis can (and
should) become routine.

Compensatory substitutions
that maintain the structure
U U
C G
U A
A U
G C
A UCGAC 3’
G C
5’

Evolutionary conservation of RNA molecules can be revealed
by identification of compensatory substitutions

• Manual annotation of 60,770 full-length mouse complementary
DNA sequences, clustered into 33,409 ‘transcriptional units’,
contributing 90.1% of a newly established mouse transcriptome
database.
• Of these transcriptional units, 4,258 are new protein-coding and
11,665 are new non-coding messages, indicating that non-coding
RNA is a major component of the transcriptome.

Therapeutic Applications
• Shooting millions of tiny RNA molecules into a
mouse’s bloodstream can protect its liver from the
ravages of hepatitis, a new study shows. In this
case, they blunt the liver’s selfdestructive
inflammatory response, which can be triggered by
agents such as the hepatitis B or C viruses.
(Harvard University immunologists Judy
Lieberman and Premlata Shankar)
• In a series of experiments published online this
week by Nature Medicine, Lieberman’s team gave
mice injections of siRNAs designed to shut down a
gene called Fas. When overactivated during an
inflammatory response, it induces liver cells to
self-destruct. The next day, the animals were given
an antibody that sends Fas into hyperdrive. Control
mice died of acute liver failure within a few days,
but 82% of the siRNA-treated mice remained free
of serious disease and survived. Between 80% and
90% of their liver cells had incorporated the
siRNAs.

Bioinformatics t8-go-hmm v2014

More Related Content

What's hot (13)

Similar to Bioinformatics t8-go-hmm v2014 (20)

More from Prof. Wim Van Criekinge (20)

Recently uploaded (20)

Bioinformatics t8-go-hmm v2014