Basic bioinformatics concepts, databases and tools

Module 3

Sequence analysisJoachim Jacob

http://www.bits.vib.be

Updated Feb 2012http://dl.dropbox.com/u/18352887/BITS_training_material/Link%20to%20mod3-intro_H1_2012_SeqAn.pdf

In this third module, we will discuss the possible analyses of sequences

Module 1

Sequence databases and keyword searching

Module 2

Sequence similarity

Module 3

Sequence analysis: types, interpretation, results

In this third module, we will discuss the possible analyses of sequences

Sequence analysis tries to read sequences to infer biological properties

AGCTACTACGGACTACTAGCAGCTACCTCTCTG

- is this coding sequence?

- can this sequence bind a certain TF?

- what is the melting temperature?

- what is the GC content?

- does it fold into a stable secondary structure?

…

Tools that can predict a biological feature are trained with examples

Automatic annotation

vs.

experimentally verified annotations

- Training dataset of sequences (← exp. verified)

- An algorithm defines parameters used for prediction

- The algorithm determines/classifies whether the sequence(s) contains the feature (→ automatic annotation)

The assumption to being able to read biological function is the central paradigm

DNA → protein sequence → structure → activity (binding, enzymatic activity, regulatory,...)

So the premise to do analysis: biological function can be read from the (DNA) sequence.

Predictions always serve as a basis for further experiments.

Protein− Metrics (e.g. how many alanines in my seq)

− Modifications and other predictions

− Domains and motifs

DNA/RNA− Metrics (e.g. how many GC)

− Predicting Gene prediction Promotor Structure

Analysis can be as simple as measuring properties or predicting features

One might be interested in: pI (isoelectric point) prediction Composition metrics Hydrophobicity calculation Reverse translation (protein → dna) Occurrence of simple patterns (e.g. does

KDL occurs and how many times) ...

Simple protein sequence analysis

http://www.sigmaaldrich.com/life-science/metabolomics/learning-center/amino-acid-reference-chart.htmlhttp://en.wikipedia.org/wiki/Hydrophobicity_scales

Protein sequence analysis tools are gathered on Expasy

http://www.expasy.org/tools (SIB)

Others: http://www.ebi.ac.uk/Tools/protein.html http://bioweb.pasteur.fr/protein/intro-en.html SMS2

Never trust a tool's output blindly Interpreting depends on the kind of output

When a prediction result is obtained, the question arises 'Is it true?' (in biological sense)

Programs giving a 'binary' result: 1 or 0, a hit or a miss.

Approach: You should comparing different prediction programs for higher confidence.

E.g. SignalP for signal peptide prediction.

Programs giving score/P-value result: the chance that the 'result' is 'not real' → the lower, the better

Approach: asses the p-value

E.g. ScanProsite for a motif

The basis for the prediction of features is nearly always a sequence alignment

Based on experimentally verified sequence annotations, a multiple sequence alignment is constructed

Different methods exist to capture the information gained from this multiple sequence alignment

Alignment reveals similar residues which can indicate identical structure

Most protein pairs with more than 25-30 out of 100 identical residues were found to be structurally similar.Also proteins with <10% identity can have similar structure.http://peds.oxfordjournals.org/content/12/2/85.long

Same structure, hence most likely same function

Chances are that the structure is not the same

The structure of a protein sequence determines his biological function

Primary = AA chain

Feb 2012: ~ 535 000 in Swissprot

Secondary = structural entities

(helix, beta-strands, beta-sheets, loops)

Tertiary = 3D

Nov 2011: ~ 80 000 in PDB

Quaternary = interactions

Number of Reportedstructures

http://en.wikipedia.org/wiki/Protein_structure

Degree of similarity with other sequences varies over the length

Homologous Histone H1 protein sequences

More conserved

Protein sequences can consist of structurally different parts

Domain

part of the tertiary structure of a protein that can exist, function and evolve independently of the rest, linked to a certain biological function

Motif

part (not necessarily contiguous) of the primary structure of a protein that corresponds to the signature of a biological function. Can be associated with a domain.

Feature

part of the sequence for which some annotation has been added. Some features correspond to domain or motif assignments.

Based on motifs and domains, proteins are assigned to families

Nearly synonymous with gene family

Evolutionary related proteins

Significant structural similarity of domains is reflected in sequence similarity, and is due to a common ancestral sequence part, resulting in domain families.

Domains and motifs are represented by simple and complex methods

Motif/domain in silico can be represented by

1. Regular expression / pattern

2. Frequency matrix / profile

3. Machine learning techniques : Hidden Markov Model

Gapped alignment

domain

http://bioinfo.uncc.edu/zhx/binf8312/lecture-7-SequenceAnalyses.pdf

Regular expressions / patterns are the simplest way to represent motifs

A representation of all residues with equal probability.

123456ATPKAEKKPKAAAKPKAKTKPKPAAKPKT-AKPAAKKLPKADAKPKAA

Consensus: AKPKAA

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

[AKT] [AKLT] P [AK] [APT] [ADEKT-]

V V V V V V

V V V X V V

V V V V V V

Position:

? Does this sequence match: AKPKTE

? And this sequence: KKPETE

? And what about this one: TLPATE

For every position the mostFrequently occurring residue

Frequency matrices or profiles include the chance of observing the residues

For every position of a motif, a list of all amino acids is made with their frequency. Position-specific weight/scoring matrix or profile. More sensitive way.

123456ATPKAEKKPKAAAKPKAKTKPKPAAKPKT-AKPAAKKLPKADAKPKAA

Consensus: AKPKA-

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

A 0.625 0 0 1/8 6/8 3/8D 0 0 0 0 0 1/8E 0 0 0 0 0 1/8K 0.25 6/8 0 7/8 0 2/8L 0 1/8 0 0 0 0P 0 0 1 0 1/8 0T 1/8 1/8 0 0 1/8 0- 0 0 0 0 0 1/8Sum 1 1 1 1 1 1

Position:

? Query: AKPKTE

? Query: KKPETE

? Query: TLPATE

http://prosite.expasy.org/prosuser.html#meth2Example: http://expasy.org/prosite/PS51092

Profile

How good a sequence matches a profile is reported with a score

123456ATPKAEKKPKAAAKPKAKTKPKPAAKPKT-AKPAAKKLPKADAKPKAA

Consensus: AKPKA-

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

A 2.377 -2.358 -2.358 0.257 2.631 1.676D -2.358 -2.358 -2.358 -2.358 -2.358 0.257E -2.358 -2.358 -2.358 -2.358 -2.358 0.257K 1.134 2.631 -2.358 2.847 -2.358 1.134L -2.358 0.257 -2.358 -2.358 -2.358 -2.358P -2.358 -2.358 0.257 -2.358 0.257 -2.358T 0.257 0.257 -2.358 -2.358 0.257 -2.358

Position:

? Query: AKPKTE Score = 11.4

? Query: KKPETE Score = 5.0

? Query: TLPATE Score = 4.3

PSWM: scores

http://prosite.expasy.org/prosuser.html#meth2

A hidden Markov Model takes also into account the gaps in an alignment

The schematic representation of a HMM

http://www.myoops.org/twocw/mit/NR/rdonlyres/Electrical-Engineering-and-Computer-Science/6-895Fall-2005/E096327C-7C77-4D23-BEBA-C28B087A9280/0/lecture6.pdf

Building a HMM from a multiple sequence alignment

Use HMMER to very sensitively search protein database with a HMM

You can search with a profile in a sequence database

Some profile adjustments to the BLAST protocol exist for particular purposes

PSI-BLAST to identify distantly related proteins

PSI-BLAST (position specific iterated)

After a search result, a profile is made of the similar sequences, and this is used again to search a database

PHI-BLAST protein with matching of a pattern

PHI-BLAST (pattern hit initiated): you provide a pattern, which all BLAST results should satisfy.

CSI-BLAST is more sensitive than PSI-BLAST in identifying distantly related proteins

PSI BLAST http://www.ncbi.nlm.nih.gov/blast/Blast.cgi?CMD=Web&PAGE=Proteins&PROGRAM=blastp&RUN_PSIBLAST=on PHI BLAST http://www.ncbi.nlm.nih.gov/blast/Blast.cgi?CMD=Web&PAGE=Proteins&PROGRAM=blastp&RUN_PSIBLAST=on CSI BLAST http://toolkit.tuebingen.mpg.de/cs_blast

Many databases exist that keep patterns, profiles or models related to function

Motif / domain databases (see NCBI bookshelf for good overview)

http://www.ebi.ac.uk/interpro/ - integrated db

http://expasy.org/prosite/ (motifs)

PFAM – hidden markov profiles (domains)

CDD (Conserved domains database) (NCBI - integrated)

Prodom (domain) (automatic extraction)

SMART (domain)

PRINTS (motif) sets of local alignments without gaps, used as frequency matrices, made by searching manually made "seed alignments" against UniProt sequences

Prosite is a database gathering patterns from sequence alignments

ScanProsite tool : search the prosite database for a pattern ( present or not )

Example : [DE](2)-H-S-{P}-x(2)-P-x(2,4)-C>

You can retrieve sequences which correspond to a pattern, you made up yourself, observed in an alignment or an known one. The syntax is specific, but not difficult: see link below!

http://prosite.expasy.org/scanprosite/scanprosite-doc.html#pattern_syntax

Interpro classifies the protein data into families based on the domain and motifs

Interpro takes all existing motif and domains databases as input ('signatures'), and aligns them to create protein domain families. This reduces redundancy. Each domain is than given an identifier IPRxxxxxxx.

Uneven size of motifs and families between families are handled by 'relations' :

parent - child and contains - found in

Families,... Regions, domains, ...

http://www.ebi.ac.uk/interpro/user_manual.html#type

Interpro summarizes domains and motifs from a dozen of domain databases

ftp://ftp.ebi.ac.uk/pub/software/unix/iprscan/README.html#2

http://www.ebi.ac.uk/interpro/databases.html

InterPro entries are grouped in types

Family

Entries span complete sequence

Domain

Biologically functional units

Repeat

Region

Conserved site

Active site

Binding site

PTM site

InterPro entries are grouped in types

You can search your sequence for known domains on InterProScan

Interproscan http://www.ebi.ac.uk/Tools/pfa/iprscan/

A sequence logo provides a visual summary of a motif

Creating a sequence logo

Create a nicely looking logo of a motif sequence: size of letters indicated frequency.

Weblogo - a basic web application to create colorful logo's

IceLogo - a powerful web application to create customized logo's

A sequence logo provides a visual summary of a motif

iceLogo

123456ATPKAEKKPKAAAKPKAKTKPKPAAKPKT-AKPAAKKLPKADAKPKAA

Consensus: AKPKA-

http://www.bits.vib.be/wiki/index.php/Exercises_on_multiple_sequence_alignment#Sequence_logo

True negatives

True positives

ScoreThreshold

Number of

matches

Ideal situation

True negatives

True positives

False positivesFalse negatives

ScoreThreshold

Number of

matches

Reality of the databases

There is always a chance that a prediction of a feature by a tool is false

Assessing the performance of categorizing tools with sensitivity and specificity

Sequence contains feature

Sequence does NOT contain feature

Feature ispredicted

Feature is NOT predicted

False positive“Type II error”

FalseNegatives

“Type I error”

Truepositive

Truenegative

“Confusion matrix”

TRUTH

PREDICTION

Assessing the performance of categorizing tools with sensitivity and specificity

Sequence contains feature

Sequence does NOT contain feature

Feature ispredicted

Feature is NOT predicted

False positive

Truenegative

SensitivityTrue positives/(TP + FN)

“Confusion matrix”

TRUTH

PREDICTION

Assessing the performance of categorizing tools with sensitivity and specificity

Sequence contains feature

Sequence does NOT contain feature

Feature ispredicted

Feature is NOT predicted

Selectivity or SpecificityTN/(FP + TN)

“Confusion matrix”

PREDICTION

TRUTH

Assessing the performance of categorizing tools with sensitivity and specificity

Sequence contains feature

Sequence does NOT contain feature

Feature ispredicted

Feature is NOT predicted

error rate*FP+FN/total

“Confusion matrix”

* misclassification rate

TRUTH

PREDICTION

Assessing the performance of categorizing tools with sensitivity and specificity

Sequence contains feature

Sequence does NOT contain feature

Feature ispredicted

Feature is NOT predicted

AccuracyTP+TN/total

“Confusion matrix”

PREDICTION

TRUTH

Protein sequences can be searched for potential modifications

http://www.expasy.org/tools/ e.g. modification (phosphorylation, acetylation,...)

To deal with the confidence in the results, try different tools, and make a graph (venn diagram) to compare the results

E.g. predict secreted proteins by signalP and RPSP, combine results in Venn

− http://bioinformatics.psb.ugent.be/webtools/Venn/

− http://www.cmbi.ru.nl/cdd/biovenn/

Overview SignalPeptide prediction tools: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2788353/

Protein sequences can be searched for secondary structural elements

Based on know structures, machine learning models of secondary structure elements are made and can be searched for.

See http://bioinf.cs.ucl.ac.uk/psipred/

Better

In case of multiple analyses on multiple sequences, mark instead of filter

WorseStarting set of sequences

Analysis filter 1

Analysis filter 2

Analysis filter 3

Analysis filter 1

Analysis filter 2

Analysis filter 3

After performing all analyses on all sequences, different filters on the results can be applied (e.g. secreted sequence, phosphorylated and containing a motif)!

NA sequences

NA sequence analyses

GC% http://mobyle.pasteur.fr/cgi-bin/portal.py?#forms::geecee

Melting temperature

For primer development, such as with Primer3

Structure

Codon usage

Codon usage table with cusp

Codon adaptation index calculation with cai

...

A lot of tools can be found at the Mobyle Portal:

Profiles and models are being used to model biological function in NA seqs

To detect Transcription factor binding sitesTRANSFAC : commercial (BIOBASE, Wolfenbüttel, Germany), started as

work of Edgard Wingender, contains eukaryotic binding sites as consensus sequences and as PSSMs. Also TRANSCompel with modules of binding sites.

ooTFD : commercial (IFTI, Pittsburgh PA, USA), started as work of David Gosh, contains prokaryotic and eukaroytic binding sites as consensus sequences and as PSSMs.

JASPAR : open access, only representative sets of higher eukaryote binding sites as PSSMs. Can be searched against sequence or sequence pair at Consite.

OregAnno : open access, collection of individual eukaryotic binding sites with their localization in the genome

PAZAR : collection of open access TF databanks

Sequence logos can give an insight in the important residues of binding sites DNA: an entry from JASPAR: tata box

A [ 61 16 352 3 354 268 360 222 155 56 83 82 82 68 77 ]C [145 46 0 10 0 0 3 2 44 135 147 127 118 107 101 ]G [152 18 2 2 5 0 10 44 157 150 128 128 128 139 140 ]T [ 31 309 35 374 30 121 6 121 33 48 31 52 61 75 71 ]

The RNA world has the Vienna servers

http://rna.tbi.univie.ac.at/

− secondary structure prediction of ribosomal sequences

− siRNA design

RNA families can be modeled by conserved bases and structure

RNA motifs (http://rfam.sanger.ac.uk/search)

Rfam is a databank of RNA motifs and families. It is made at the Sanger Centre (Hinxton, UK), from a subset of EMBL (well-annotated standard sequences excluding synthetic sequences + the WGS) using the INFERNAL suite of Soan Eddy. It contains local alignments with gaps with included secondary structure annotation + CMs.

Some interesting links

Nucleic acid structure

Unafold - Program accessible through webinterface

After designing primers, you might want to check whether the primer product does (not) adapt a stable secondary structure.

Some collections of links− Good overview at http://www.imb-jena.de/RNA.html

− European Ribosomale RNA database (VIB PSB)

Prediction of genes in genomes rely on the integration of multiple signals

Signals surrounding the gene (transcription factor binding sites, promoters, transcription terminators, splice sites, polyA sites, ribosome binding sites,...)

→ profile matching

Differences in composition between coding and noncoding DNA (codon preference), the presence of an Open Reading Frame (ORF)

→ compositional analyses

Similarity with known genes, aligning ESTs and (in translation) similarity with known proteins and the presence of protein motifs

→ similarity searches

Signals

Composition

Similarity

e.g. potential methylation sites (profiles)

GC

Alignment of ESTs

Prediction of genes in genomes rely on the integration of multiple signals

Software for prediction genes

EMBOSS

− simple software under EMBOSS : syco (codon frequency), wobble (%GC 3rd base), tcode (Ficketstatistic : correlation between bases at distance 3)

Examples of software using HMM model of gene :

Wise2 : using also similarity with known proteins http://www.ebi.ac.uk/Tools/Wise2

GENSCAN : commercial (Chris Burge, Stanford U.) but free for academics, has models for human/A. thaliana/maize, used at EBI and NCBI for genome annotation http://mobyle.pasteur.fr/cgi-bin/portal.py?#forms::genscan

GeneMark : commercial (GeneProbe, Atlanta GA, USA) but free for academic users, developed by Mark Borodovsky, has models for many prokaryotic and eukaryotic organisms http://exon.gatech.edu

Tutorial on gene prediction http://www.embl.de/~seqanal/courses/spring00/GenePred.00.html

Short addendum about downloading files

FTP, e.g. ftp://ftp.ebi.ac.uk/pub/databases/interpro/

– 'file transfer protocol'

– Most browsers have integrated ftp 'client'

– Free, easy to download files, possibility to resume after fails

HTTP, e.g. http://www.ncbi.nlm.nih.gov/entrez

Standard protocol for internet traffic,

Slowest method

Aspera – for large datasets (>10GB) downloads

In use in the short read archive (SRA)

Fastest method available currently

Conclusion

Prediction vs. experimental verified Different algorithms need to be compared Predictions need to be validated by independent

method

Software <-> Databases Questions? Get social!

→ www.seqanswers.com

→ http://biostar.stackexchange.com Always only basis for further wet-lab research

Summary In this third module, we will discuss the possible analyses of sequences

Sequence analysis tries to read sequences to infer biological properties

Tools that can predict a biological feature are trained with examples

The assumption to being able to read biological function is the central paradigm

Analysis can be as simple as measuring properties or predicting features

Protein sequence analysis tools are gathered on Expasy

Never trust a tool's output blindly

The basis for the prediction of features is nearly always a sequence alignment

Alignment reveals similar residues which can indicate identical structure

The structure of a protein sequence determines his biological function

Degree of similarity with other sequences varies over the length

Protein sequences can consist of structurally different parts

Based on motifs and domains, proteins are assigned to families

Domains and motifs are represented by simple and complex methods

Regular expressions / patterns are the simplest way to represent motifs

Frequency matrices or profiles include the chance of observing the residues

How good a sequence matches a profile is reported with a score

A hidden Markov Model takes also into account the gaps in an alignment

Use HMMER to very sensitively search protein database with a HMM

Some profile adjustments to the BLAST protocol exist for particular purposes

Many databases exist that keep patterns, profiles or models related to function

Prosite is a database gathering patterns from sequence alignments

Interpro classifies the protein data into families based on the domain and motifs

Interpro summarizes domains and motifs from a dozen of domain databases

You can search your sequence for known domains on InterProScan

A sequence logo can provide a visual summary of a motif

Protein sequences can be searched for potential modifications

Protein sequences can be searched for secondary structural elements

In case of multiple analyses on multiple sequences, mark instead of filter

Profiles and models are being used to model biological function in NA seqs

Sequence logos can give an insight in the important residues of binding sites

The RNA world has the Vienna servers

RNA families can be modeled by conserved bases and structure

Prediction of genes in genomes rely on the integration of multiple signals