Predicting cellular localization

1
Predicting cellular localization

• Bioe C144/C244
• Fall 2010

2
Eukaryotic protein localization
3
Why localize?

• Subcellular localization is a key functional
  characteristic of proteins.
• To co-operate for a common physiological function
  (metabolic pathway, signal transduction cascade,
  structural associate etc.), proteins must be
  localized in the same cellular compartment.

http//mendel.imp.univie.ac.at/CELL_LOC/
4
Correct Localization is required for
pathway/complex formation

• A set of many co-operating proteins is
  responsible for a physiological function
  (metabolic pathway, signal transduction cascade,
  structural associate etc.).
• Subcellular localization is an essential
  characteristic for this level.
• For proper functioning, the protein has to be
  translocated to the correct intra- or
  extracellular compartments in a soluble form or
  attached to a membrane.

http//mendel.imp.univie.ac.at/CELL_LOC/
5
Computer-Aided Approaches for the Assignment of
Subcellular Localization

• Automatic, computer-aided selection methods are
  clearly the only way to identify interesting
  attractive target proteins among the haystack of
  new gene sequence data.
• One of the helpful decision criteria is the
  probable subcellular localization of the gene
  products.
• For example, in a search for virulence factors of
  pathogenic bacteria or easily accessible entry
  points for pharmaceutical drugs, extracellular
  proteins are good candidates

http//mendel.imp.univie.ac.at/CELL_LOC/
6
Computer-Aided Approaches for the Assignment of
Subcellular Localization

• For a primary screening of gene sequences, the
  first step is a general classification into
  intracellular, extracellular, membrane-related
  (both with transmembrane regions and with lipid
  anchors) and viral proteins.
• In the case of Eukaryotes, intracellular location
  is desirable to be further detailed with respect
  to organelles (mitochondrium, chloroplast,
  endoplasmatic reticulum and Golgi apparatus,
  nucleus).

http//mendel.imp.univie.ac.at/CELL_LOC/
7
Predicting subcellular localization by homology
with characterized proteins

• Subcellular localization can often be assigned by
  searching for homologous sequences.
• This is an easy task for a few new proteins but
  very difficult for thousands of sequences
  contained in new genomes.
• Even with the most advanced retrieval systems and
  relying on the well-annotated SWISS-PROT, it is
  impossible to get exhaustive classifications with
  respect to subcellular localization.

http//mendel.imp.univie.ac.at/CELL_LOC/
8
Prediction method 2 analysis of sequence
properties

• First attempts to classify proteins with respect
  to cellular localization based on amino acid
  sequence properties Nishikawa and Ooi (J.Biochem.
  1982)
• amino acid composition, disulphide bonds, the
  secondary structural class related to function
  and localization
• Early results were promising, but based on a
  small sample.

http//mendel.imp.univie.ac.at/CELL_LOC/
9
Prediction by signal peptide detection

• Some proteins have sequence signals that
  determine their translocation to organelles or
  outside the cell
• Claros et al. Curr.Op.Struct.Biol. (1997).
• These patterns are not clear cut, especially for
  the intracellular organelle targeting peptides
• prediction accuracy is limited
• Nielsen et al. Prot.Eng. (1997) v.10, 1
• Combinations of compositional and signal sequence
  analyses have been used in expert systems for the
  prediction of cellular localization
• Nakai Kanehisa Genomics (1992)
• In general not systematic and not rigorously
  tested

http//mendel.imp.univie.ac.at/CELL_LOC/
10
Extracting information from sequence

• Signal peptides short sequences in the protein
  used to target the protein for specific cellular
  compartments.
• Signal patches (clusters of amino acids in close
  proximity in 3D structure, but distant in primary
  sequence) are also found
• Examination of amino acids at structure surface
  can be particularly helpful subtle preferences
  of different amino acids for different
  environments

11
Trans-membrane helix prediction
12
(No Transcript)
13
Helical membrane proteins

• Key components in cell-cell signalling
• Mediate transport of ions and solutes across
  membrane
• Crucial for recognition of self
• Major class of drug targets
• More than 50 of prescription drugs act on GPCRs
  (G-protein coupled receptors)
• Multi-billion dollar industry

14
Many predicted few known

• Solved structures available for very few membrane
  proteins
• Predicted 10K helical membrane proteins in human
  genome (25 of genome!)

Chen and Rost, 2002
15
Helical membrane proteins challenge bioinformatics

• Very little info about 3D structures
• Very hard to crystallize
• Hardly traceable by nuclear magnetic resonance
  (NMR) spectroscopy
• Relatively easy to identify (rough) location of
  helices through low-resolution experiments
• C-terminal fusion with indicator proteins
• Antibody binding

Chen and Rost, 2002
16
Concepts for predicting TM helix location and
topology

• Hydrophobicity scales provide simple criteria for
  prediction
• TM helices are predominantly non-polar
• TM helix length between 12-35 aa
• Globular regions between membrane helices
  typically shorter than 60 aa
• Positive inside rule von Heijne
• Connecting loop regions on inside have more
  positive charge than loop regions on outside

Chen and Rost, 2002
17
Hydrophobicity scales

• Kyte and Doolittle (20 yrs ago)
• Hydropathy scale, moving window approach
• Window of 19 residues discriminated best between
  membrane and globular
• Other work equally successful
• Drawback methods fail to discriminate between
  membrane regions and highly hydrophobic globular
  segments

Chen and Rost, 2002
18
Other clues

• Amino acid preferences for membrane and
  non-membrane proteins
• Training data for methods derived from proteins
  identified as containing TM helices, as well as
  other secondary structure types
• Higher accuracy

Chen and Rost, 2002
19
Including topology helps

• TopPred (von Heijne, 1992)
• Topology prediction, using hydrophobicity
  analysis, possible topologies ranked by
  positive-inside rule
• SOSUI (Hirokawa et al, 1998)
• Combined KD hydropathy, amphiphilicity, relative
  and net charges, protein length

Chen and Rost, 2002
20
Including homology helps

• Alignment of homologs known to help secondary
  structure prediction (Rost and Sander, 1993)
• Note for 20-30 of proteins in any genome, no
  identifiable homologs can be found!
• PHDhtm first method using homology info for
  membrane prediction
• Uses neural networks, DP, multiple alignment
• one of the most accurate prediction methods

Chen and Rost, 2002
21
Including homology helps

• TMAP (Persson and Argos, 1996)
• Derived amino acid propensities from known TMs
• 4-residue caps of membrane helices
• 21 residue TM segments
• Found at outside of membrane N D G F P W Y V
• Found mostly inside A R C K
• Used these propensities to improve prediction

Chen and Rost, 2002
22
Grammatical rules

• TMHMM pioneered building models of predicted
  membrane proteins in one consistent methodology
• Sonnhammer et al 1998, Krogh et al 2001
• Similar concept implemented in HMMTOP
• Tusnady and Simon, 1998
• MEMSAT similar to HMMTOP
• Jones et al, 1994

Chen and Rost, 2002
23
Topology questions

• The topology of a TM protein indicates its
  orientation with respect to the membrane
• which regions are outside (extracellular) and
  which are cytoplasmic
• Predicted topologies turn out to be wrong roughly
  as often as theyre correct

Chen and Rost, 2002
24
Sequence information aiding TM recognition

• Hydrophobic stretches (for lipid bilayer)
• Positive inside rule
• Von Heijne 1986, 1994
• Abundance of positively charged residues
• Improved predictions through use of
• sliding windows
• Multiple alignment
• Neural networks

Chen and Rost, 2002
25
Errors in TM prediction

• Under-prediction (False negative)
• Over-prediction (False positive)
• False merge
• two adjacent helices predicted to be one helix
• False split
• One long helix predicted to be two
• Inexact placement of helices

Chen and Rost, 2002
26
Prediction accuracy (1)

• Performance accuracy overestimated significantly!
• developers have overrated their methods by
  15-50 Chen et al, unpublished
• Why do developers overestimate their method
  accuracy?
• Validation performed on proteins closely related
  to training sequences (and thus not indicative of
  performance on novel sequences)

Chen and Rost, 2002
27
Prediction accuracy (2)

• Membrane helices are not entirely conserved
  across species
• Implies that even related proteins may have
  different topologies ( TM helices, orientation)
  and perform different cellular functions
• N.B. There is no indication that the authors
  meant to imply that proteins that are globally
  alignable have differences in their TM domain
  locations or numbers
• Measures of accuracy of prediction not comparable
  across methods, due to lack of standard benchmark
• Benchmark dataset now available at EBI

Chen and Rost, 2002
28
Chen et al findings

• Most TM methods get the number of helices right
  for most membrane proteins
• 86 of TMH residues predicted by best methods
• 70-75 of proteins get all TM helices predicted
  correctly by top methods
• Topology correct for only half of all proteins

Chen and Rost, 2002
29
Prediction accuracy (4)

• Some papers have claimed that simple
  hydrophobicity scales are as accurate as more
  sophisticated methods
• Chen et al disagree

Chen and Rost, 2002
30
Prediction accuracy (5)

• All methods confuse membrane helices with signal
  peptides
• Best separation provided by ALOM2 (Nakai and
  Kanehisa)
• Optimized to sort proteins into classes of
  sub-cellular localization

Since Rosts paper, the Phobius server was
developed to integrate TM and signal peptide
prediction http//www.ebi.ac.uk/Tools/phobius/inde
x.html
Chen and Rost, 2002
31
Prediction accuracy (6)

• Most methods wrongly predict membrane helices in
  globular proteins
• Most methods overestimate their ability to
  distinguish between globular and membrane
  proteins

Chen and Rost, 2002
32
Emerging and future developments

• Improved prediction by averaging over many
  methods (I.e., consensus approaches)
• Promponas and colleagues CoPreTHi combined 7
  methods, requiring 3 to agree
• Nilsson et al, 2000, used 5 methods
• Accuracy correlated with number of methods
  agreeing

Chen and Rost, 2002
33
Emerging and future developments
Chen and Rost, 2002

• Amphiphilic (aka amphipathic) alpha helix
  identification can improve prediction
• Helical-membrane and signal peptide predictions
  must be combined explicitly
• Best signal peptide prediction tool is SignalP
  (Nielsen et al 1997)
• PSORT, HMMTOP and THHMM integrate these
  predictions
• More thorough combination is still missing

Except, of course, for Phobius, released since
this paper
34
Emerging and future developments

• Databases of TM proteins being produced and
  curated
• Membrane-specific substitution matrices improve
  database search for TM proteins
• Current substitution matrices based on globular
  proteins
• Henikoff and Henikoff have membrane-helix-specific
  substitution matrix PHAT

Chen and Rost, 2002
35
Sequence conservation in TM domains

• Residues on helix-helix interface tend to be more
  conserved than those facing the lipid bilayer
• Conservation in TM helices greater than
  structurally variable regions but not as
  significant as enzyme active sites and other
  functionally critical regions (KS observation)

36
More data from structural studies of TM proteins

• Solved membrane protein structures have also
  shown that helical propensities are different in
  the membrane.
• Glycine and proline, which are thought to be
  helix-breakers in soluble proteins, occur in the
  transmembrane helices of cytochrome c oxidase
• Tsukihara et al, 1995.
• Studying known structures has revealed that
  aromatic residues are often in the bilayer
  interface, possibly anchoring the transmembrane
  helix in the bilayer
• Pawagi et al, 1994.

37
More data from structural studies

• Serine and threonine can satisfy hydrogen bond
  donors and acceptors by hydrogen bonding to
  backbone carbonyls, making membrane localization
  favorable (Engelman et al 1986)
• Analysis of solved membrane proteins show TM
  length ranges from 14-36 aa (varying due to
  variations in lipid bilayer width)
• Canonical alpha helix prediction methods derived
  from soluble proteins are not as effective at
  predicting TM-located helices

38
TMHMM provides a grammar to parse sequences into
subregions
39
TMHMM author findings

• TMHMM correctly predicts 9798 of the
  transmembrane helices.
• TMHMM can discriminate between soluble and
  membrane proteins with both specificity and
  sensitivity better than 99
• although the accuracy drops when signal peptides
  are present
• This high degree of accuracy allowed authors to
  predict reliably integral membrane proteins in a
  large collection of genomes.
• Based on these predictions, authors estimate that
  2030 of all genes in most genomes encode
  membrane proteins
• which is in agreement with previous estimates.
• Proteins with Nin-Cin topologies are strongly
  preferred in all examined organisms
• except Caenorhabditis elegans, where the large
  number of 7TM receptors increases the counts for
  Nout-Cin topologies.

40
Aspects of model

• Specialized modeling of various regions
• Helix caps
• Middle of helix
• Regions close to membrane
• Globular domains (all modeled identically)
• TM amino acid stats derived from known TM domains

41
Training data
42
Signal peptide prediction
43
(No Transcript)
44
(No Transcript)
45
(No Transcript)
46
Chloroplast transit peptides are hard to detect
47
(No Transcript)
48
(No Transcript)
49
A plant GPCR??
Arabidopsis Thaliana GCR2
50
only one Arabidopsis putative GPCR protein
(GCR1) has been characterized in plants (1720),
and no ligand has been defined for any plant GPCR
51
Transmembrane structure prediction suggests that
GCR2 is a membrane protein with seven
transmembrane helices
-Despite bold claim of 7TM GPCR, only two
prediction servers used, no confidence values
indicated, and figure ended up in Supplemental
Material!
DAS
TMPred
52
(No Transcript)
53
Liu et al predicted GCR2 as a seven-transmembrane
protein (7TM), using the TMpred and DAS
programs, but did not report score thresholds to
evaluate the confidence of these predictions.
TMpred and DAS are known to erroneously predict
transmembrane helices within soluble proteins
(55 and 83 false positive rates, respectively)
GCR2 alignment with LanC superfamily (non 7-TM
GPCR)
54
(No Transcript)
55
(No Transcript)
56
We initially predicted that GCR2 was a
seventransmembrane protein using TMpred and
DAS software programs (2). We further used 12
distinct software programs to predict the
topological structure of GCR2 and found that 9 of
them (TMHMM, SOUSI, and DAS TMfilter
excluded) showed that GCR2 is a transmembrane
protein with various numbers of transmembrane
domains. TMHMM has underpredicted
transmembrane domains in many instances (3), and
the only other reported GPCR in Arabidopsis, GCR1
(4), was predicted to be a three-transmembrane
protein by SOSUI. In addition, about 14 of
known transmembrane proteins (established by
crystal structure or biochemical evidence) cannot
be correctly predicted by available software (3).
Thus, computational prediction of membrane
proteins is not yet a mature science and mainly
serves to generate hypotheses for experimental
testing
57
(No Transcript)
58
Discussion
Based on the evidence presented who do you think
is right? TM prediction and validation is
challenging both bioinformaticians
and Experimentalists alike!
59
TM/Signal Peptide/Localization prediction servers

• Phobius http//phobius.sbc.su.se
• -combined topology and signal peptide prediction
• TMHMM http//www.cbs.dtk.dk/services/TMHMM/
• -TM helix prediction

• TargetP http//www.cbs.dtu.dk/services/TargetP/
• -subcellular localization of eukaryotic proteins
• SignalP http//www.cbs.dtu.dk/services/TargetP/
• -predicts the presence and location of signal
  peptide cleavage sites

60
Summary points

• Protein localization is a critical aspect of
  protein function
• Methods for predicting localization can
  over-estimate their expected accuracy
• Datasets used in validation typically differ from
  one method to the next, so results are not
  comparable
• Expect false positive and false negative
  predictions
• Consensus prediction using various types of
  information and predictions are your best bet for
  improving accuracy