copyright 2003 limsoon wong recognition of protein features limsoon wong institute for infocomm...

Copyright 2003 limsoon wong

Recognition of Protein Features

Limsoon Wong

Institute for Infocomm ResearchBI6103 guest lecture on ?? March 2004


Lecture Plan

• Membrane proteins

• Subcellular localization


Recognition of Transmembrane Helices


Eukaryotic Cells

• Eukaryotic cells have membrane-bound compartments with specialized functions


Lipids & Membrane

• Membrane is a double layer of lipids and associated proteins which define subcellular compartments or enclose the cell

• Lipids consist of a “polar head group” and long-chain fatty acids• This dual nature promotes formation of lipid bilayers

• “Hydrophobic tails” are shielded from aqueous environment

• Water-soluble (i.e., charged or polar) molecules cant pass through this impermeable barrier

• Permeability across the bilayer is regulated by membrane proteins that span the bilayer and function like channels or pores


all- -barrel

Membrane Proteins

• Two types of membrane proteins: Integral vs peripheral

• Two types of integral membrane proteins: all- vs -barrel


Topography & Topology

• topography: predict location of transmembrane segment

• topology: predict location of N- and C-termini wrt lipid bilayer

• We focus on topography prediction for all- membrane proteins

Lipid molecules


Datasets

• Jayasinghe et al. Protein Sci, 10:455-458, 2001– 59 high resolution membrane proteins– www.biocomp.unibo.it/gigi/ENSEMBLE

• Moller et al. Bioinformatics, 16:1159--1160, 2000– 151 low resolution membrane proteins

• Jones et al., Biochem., 33(10):3038--3049, 1994– 38 multi-spanning and 45 single-spanning membrane proteins– topologies experimentally determined

• Sonnhammer et al., ISMB, 6:175-182, 1998– 108 multi-spanning and 52 single-spanning membrane proteins

– most of experimentally determined topologies, but less reliably determined than Jones et al.


Monne et al., JMB, 288:141--145, 1999:

Turn Propensity Scale for TM Helices

• E. coli Lep protein contains two TM domains (H1, H2) and C-terminal doman P2

• Translocation of P2 to lumenal side is easy to test by glycoslation

• Replace H2 by 40 residue poly-L segment LIK4L21XL7VL10Q3P

• The poly-L segment can form either one long TM or 2 closely-spaced TM helices, depending on what is substituted for X

ER


Monne et al., JMB, 288:141--145, 1999:


• Using the poly-L segment, measure “turn” propensity of the 20 amino acids by substituting them for the X in the poly-L segment

• Hydrophobic residues (I, V, L, F, C, M, A) do not induce turn

• Charged and polar residues (except S & T) induce turn

• Exercise:– What are the charged/polar

residues?

– What could be reason of S & T not inducing turn?

glycoslated

non-glycoslated


Monne et al., JMB, 288:141--145, 1999

• In all- membrane proteins, – hydrophobic residues

prefer membrane env and have low turn propensity

– charged & polar residues induce turn formation to avoid membrane interior

prediction of TM helix distinction of 1 long TM

helix vs 2 closely spaced TM helices

Monne et al., JMB, 288:141--145, 1999:



Monne et al., JMB, 288:141--145, 1999

• Inside of cellular membrane is hydrophobic

• Segment of protein that spans membrane is expected to contain many hydrophobic amino acids

Locate segments that have high average “hydrophobicity” score

Wiess et al, ISMB, 1:420--421, 1993 Hydrophobicity Approach


Wiess et al, ISMB, 1:420--421, 1993 Hydrophobicity Approach

• find a segment of 10 to 70aa with hp > 0.71

• expand to longer segment with hp > 0.35

• mark this segment as TM

• repeat above starting from position after previous segment

• Caveats:– may be unable to

distinguish hydrophobic core of nonmembrane proteins vs. transmembrane regions

– what are the right thresholds?

Adjustable thresholds


An Example: Bacteriorhodopsin

http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=protein&list_uids=461610&dopt=GenPept&term=bacteriorhodopsin&qty=1

1 gigtllmlig tfyfiargwg vtdkkareyy aitilvpgia saaylsmffg iglttvevag 61 maepleiyya ryadwlfttp lllldlalla nadrttigtl igvdalmivt gligalshtp

121 larytwwlfs tiaflfvlyy lltvlrsaaa elsedvqttf ntltalvavl wtaypilwii

181 gtegagvvgl gvetlafmvl dvta

7 transmembrane helices






• After applying hydrophobicity scale...



• Compute hydrophobicity score, hp > 7




TM identified: 6/7, TM FP: 0TM residue identified: 62/117, TM residue FP: 4



• Expand segment, maintain hp > 5, avoid low hydrophobicity




TM identified: 6/7, TM FP: 0TM residue identified: 100/117, TM residue FP:15


Sonnhammer et al., ISMB, 6:175-182, 1998:

TMHMM, A HMM Approach

• There are 3 main locations of a residue:– TM helix core (viz., in hydrophobic tail of membrane– TM helix cap (viz., in head of membrane)

• cytoplasmic vs • non-cytoplasmic side of the helix core

– loops• cytoplasimc vs • non-cytoplasmic (short) vs • non-cytoplasmic (long)

So needs HMM with 7 states• Exercise: What is the 7th state for?

cyto

non-cyto



TMHMM, Architecture

cyto

non-cyto

Each state has an associated probabilitydistribution over the 20 amino acids characterizing the variability of amino acids in the region it models



TMHMM, Architecture

• The first 3 and last 2 core states have to be traversed. But all other core states can be bypassed.

• This models core regions of 5--25 residues



TMHMM, Architecture

• The states of globular, loop, & cap regions. • The caps are 5 residues each. Since core is 5--25

residues, this allows for helices 15--35 residues long

To model bias in amino acid usage near cap

To model neutral aminoacid distribution



TMHMM, Training the HMM• Stage 1: Baum-Welch is used for maximum likelihood estimation from

“diluted” labeled training data. As precise end of TM is only approximately known, we “dilute” by unlabeling 3 residues on each side of a helix boundary to accommodate this

• Stage 2: Baum-Welch is used for maximum likelihood estimation from

“relabeled” training data. The original training data are diluted as by unlabeling 5 residues on each side of a helix boundary. Model from Stage 1 is used to produce “relabeled training data” by relabeling this part under constraints of remaining labels

• Stage 3: Model from Stage 2 is further tuned by a method for “discriminative” training, to maximize probability of correct prediction (Krogh, ISMB, 5:179--186, 1997)


Krogh, ISMB, 5:179--186, 1997:

Discriminative HMM Training



TMHMM, Example

Non-cytoplasmic Cytoplasmic TM segment

Datasets• Jones et al., Biochem., 33(10):3038--3049, 1994• Sonnhammer et al., ISMB, 6:175-182, 1998



TMHMM, Accuracy (10-CV)

All TM segments& their orientationcorrectly predicted

All TM segmentscorrectly predicted,ignoring orientation

precision

Jone

s et a

l

Sonnhammer

et al


NN HMM1 HMM2

ENSEMBLE

Martelli et al. Bioinformatics, 19:i205--i211, 2003

ENSEMBLE


ENSEMBLE:

The Neural Network Part

• The NN part is a cascade shown above, a la Rost et al., Protein Science, 1995

h1

h2

h5

HMM

LOOP

Inputlayer17*2inputs

1

17

15 hiddenunits

17 * 20input units

Feed-forwardback-propagationneural network


ENSEMBLE:

The HMM1 Part

• HMM1 models the hydrophobic nature of most TM helices, a la Krogh et al. JMB 2001 & Sonnhammer et al., ISMB 1998


ENSEMBLE:

The HMM2 Part

• HMM2 models TM helices that are mix of hydrophobic and hydrophilic residues, ala Martelli et al., Bioinformatics 2002.


NN HMM1 HMM2

ENSEMBLE

ENSEMBLE:

Predicting if a residue is in TM

NN(p,i) = NN(H,p,i) NN(L,p,i) HMM1(p,i) = AP1(H,p,i) AP1(I,p,i) AP1(O,p,i)

HMM2(p,i) = AP2(H,p,i) AP2(I,p,i) AP2(O,p,i)

• E(p,i) = (NN(p,i) + HMM1(p,i) + HMM2(p,i)) / 3

position

helix

loop (inner I, outer O)

E(p,i) > 0 means residue i of protein p is in TM helix


Ensemble: Topography PredictionFariselli et al., Bioinformatics, 2003

NN HMM1 HMM2

ENSEMBLE MaxSubSeq

TM helix found by MaxSubSeq butwould be missed w/o it

This path istaken means positions m to j form a helix


Ensemble:

Topography Prediction Results

60%

65%

70%

75%

80%

85%

90%

Jayasinghe(CV)

Moller

NN

HMM1

HMM2

ENSEMBLE

TMHMM2.0

MEMSAT

PHD

HMMTOP

A prediction is considered correct if (a) the number of TM segments is correct and(b) the overlap between a predicted and a real TM segment > 8aa


Topology Prediction: Postive-Inside RuleGavel et al., FEBS, 282:41--46, 1991

• Positively-charged residues (Lys and Arg) are enriched more than 2 fold in stromal vs luminal loops


Topology Prediction:

Ensemble

“positive-inside” rule


Ensemble:

Topology Prediction Results

40%

45%

50%

55%

60%

65%

70%

75%

80%

Jayasinghe(CV)

Moller

ENSEMBLE(rule 4)

TMHMM2.0

MEMSAT

PHD

HMMTOP

ENSEMBLE(rule 1)


Short Break


Subcellular Localization


Compartments and Sorting

• Eukaryotic cells requires proteins be targeted to their subcellular destinations

• Protein sorting is determined by specific amino acid sequences, or “signals”, within the protein

• Secretory pathway targets proteins to plasma membrane, some membrane-bound organelles such as lysosomes, or to export proteins from the cell


Secretory Pathway

• The secretory pathway consists of the endoplasmic reticulum (ER), Golgi apparatus and transport vesicles

• The transport vesicles carry proteins from one compartment to the other

• Exocytosis is mediated by fusion of secretory vesicles with the plasma membrane.

• Endocytosis is the opposite of exocytosis and involves the uptake of extracellular material by pinching off vesicles from the plasma membrane

• The contents of the endocytic vesicles are delivered to the lysosomes by membrane fusion

• Lysosomes contain hydrolytic enzymes that breakdown macromolecules into the smaller subunits which can be utilized by the cell for its own biosynthesis


Datasets

• Reinhartdt & Hubbard, NAR, 26:2230--2236, 1998– 2427 eukaryotic proteins for 4 locations (cytoplasmic, extracellular, nuclear,&

mitochondrial)

– 997 prokaryotic proteins for 3 locations (cytoplasmic, extracellular, & periplasmic)

• Park & Kanehisa, Bioinformatics, 19:1656--1663, 2003– 7589 eukaryotic proteins from 709 organisms for 12 locations

(chloroplast, cytoplasmic, cytoskeleton, ER, extracellular, golgi, lysosomal, mitochondrial, nuclear, peroxisomal, plasma membrane, vacuolar)

• Chou & Cai, JBC., 277:45765--45769, 2002– 2191 proteins for 12 locations

• Emanuelsson et al., JMB, 300:1005--1016, 2000

• Gardy et al., NAR, 31:3613--3617, 2003


Common Eukaryotic Protein Sorting Signals

For a comprehensive list of cellular localization sites, see

http://mendel.imp.univie.ac.at/CELL_LOC/index.html


Schematic View of SortingSignals

cleavage site

~25aa


Sequence Logos ofSP, mTP, & cTP

SPsignal peptide

mTPmitochondrial

transfer peptide

cTPchloroplast

transit peptide


Neural Network Approach: TargetPEmanuelsson et al., JMB, 300:1005--1016, 2000

• cTP, mTP, SP– 4 hidden units– feedforward NNs– input windows:

• 55aa (cTP), 35aa (mTP), 27aa (SP)

• sparsely encoded

• Integrating Network– 0 hidden unit– feedforward NN– input is taken from the

outputs of cTP, mTP, SP networks over 100aa at N-terminal

cTP: chloroplast transit peptide, mTP: mitochondria transfer peptide, SP: signal peptide


TargetP:

Performance

Dataset: Emanuelsson et al., JMB, 2000


Expert System Approach: PSORT Horton & Nakai, ISMB, 1997

A simplified version of the decision tree thatPSORT uses tocheck and reasonover various sorting signals


A Refinement: PSORT-BGardy et al., NAR, 31:3613--3617, 2003

SCL-BLAST

Motifs HMMTOPOuter

MembraneProtein

SubLocCSignal

Peptides

BayesianNetwork

Localization sitesor “unknown”

• Sites considered– cytoplasm– inner membrane– periplasm– outer membrane– extracellular space


PSORT-B:

SCL-BLAST

• Homology to a protein of known localization is good indicator of a protein’s actual localization site

BLAST target protein against a database of proteins whose localization sites are known

Return localization sites of hits at E-value of 10e-10

over 80% of length


PSORT-B:

Motifs

• Some motifs in PROSITE may be able to identify subcellular localization with 100% precision

Scan target protein against a database of such motifs (28 such 100%-precision motifs are known)

Return localization sites corresponding to the motif hits


PSORT-B:

HMMTOP

-helical transmembrane region is reliable indicator of localization to inner membrane

Scan target protein for transmembrane helices using HMMTOP

Return localization site as “inner membrane” if >2 helices found


PSORT-B:

Outer Membrane Proteins

• Outer-membrane proteins have characteristics -barrel structure

Identify freq seq occurring only in -barrel proteins (279 such freq seq known)

Scan target protein for these freq seq

Return localization site as “outer membrane” if >2 such freq seq found


PSORT-B:

SubLocC

• Overall amino acid composition is useful for recognizing cytoplasmic proteins

Trained SVM on overall amino acid composition to predict cytoplasmic vs non-cytoplasmic, as in SubLoc

Analyze target protein’s amino acid composition using this SVM


PSORT-B:

Signal Peptides• Presence of signal peptide at N-

terminal means protein not cytoplasmic

Train HMM and SVM to recognize signal peptides and their cleavage sites

If high-confidence cleavage site found by HMM in first 70aa of target protein, then “non-cytoplasmic”

If low-confidence cleavage site found, pass candidate signal peptide to SVM to confirm

If confirmed, then “non-cytoplasmic” Otherwise, “unknown”


PSORT-B:

Bayesian Network

• Bayesian Network integrates results from the 6 modules

• Produces a score for each of the 5 possible localization sites

• If a site scores >7.5, then predicts as a localization site of the target protein

• If no site scores >7.5, then makes no prediction


PSORT-B:

Performance of Individual Modules

Dataset: Gardy et al., NAR, 2003


PSORT-B:

Performance wrt Localization Sites

PSORT-B is a considerable improvement over original PSORT

Dataset: Gardy et al., NAR, 2003


PSORT vs PSORT-B:

Some Remarks

• PSORT considers various signal/features in a top-down way driven by its reasoning tree

• PSORT-B generates all signal/features in a bottom-up way, then integrate them for decision making using Bayesian Network

• Machine learning “beats” human expert? Probably the number of features/rules needed is too much/complicated


Amino acid composition of proteins residing in different sites are different


Amino Acid Composition Differences

• each cellular location has own characteristic physio-chemical environment

• proteins in each location have adapted thru evolution to that environment

• thus reflected in the protein structure and amino acid composition

• If the above is true, the amino acid composition differences wrt cellular location sites should be more pronounced on protein surfaces than protein interior

• Exercise: Why?


Adaptation of Protein SurfacesAndrade et al., JMB, 1998

Proportion ofjth amino acid type in ith protein

• To test the theory of adaptation of protein surfaces to subcellular localization, we do a plot of 3 types of composition vectors along their first two principal components


Adaptation of Protein Surfaces Andrade et al., JMB, 1998

Total amino acidcomposition vector

Surface amino acidcomposition vector

Interior amino acidcomposition vector

• Clearly total & surface composition vectors show better separation than interior composition vectors


Amino Acid Composition

• This means can use amino acid composition vectors, especially those from protein surfaces, to predict subcellular localization!

• Let’s see how this turn out….


Neural Networks: NNPSLReinhardt & Hubbard, NAR, 26:2230--2236, 1998

Input1

Input20

cytoplasmic

extracellular

mitochodrial

nuclear

fraction of each aminoacid in the input protein


NNPSL:

Performance

• Outputs NNPSL have values 0 to 1. The difference () between the highest and the next highest nodes can be used as a reliability index

0 < < 0.2

0.2 < < 0.4

0.4 < < 0.6

0.6 < < 0.8

0.8 < < 1

Dataset: Reinhardt & Hubbard,NAR, 1998


Performance Emanuelsson, BIB, 3:361--376, 2002

(940 proteins)

(2738 proteins)

Dataset: Emanuelsson et al., JMB, 2000


Markov ChainYuan, FEBS Letters, 451:23--26, 1999

Why?


Markov Chain:

Performance

NNPSL 4th Order Markov(Eukaryotic)

Dataset: Reinhardt & Hubbard,NAR, 1998


Support Vector Machines: SubLocHua & Sun, Bioinformatics, 17:721--728, 2001

extracellularvs rest

nuclearvs rest

cytoplasmicvs rest

mitochondrialvs rest

ArgmaxX X-vs-rest

SVM

SVM

SVM

SVMThe SVMs use • polynomial kernel with d = 9 (prokaryotic),

K(Xi,Xj) = (Xi ·Xj + 1)d

• RBF kernel with =16 (eukaryotic),K(Xi, Xj) = exp(- |Xi - Xj|2

20-dimensional vector giving amino

acid composition of the input protein


SubLoc:

Performance

NNPSL SubLoc

(Eukaryotic)

Dataset: Reinhardt & Hubbard, NAR, 1998


SubLoc: Robustness of Amino Acid Composition Approach

• Amazingly, accuracy of SubLoc is virtually unaffected when the first 10, 20, 30, & 40 amino acids in a protein are deleted

• Amino acid composition is a robust indicator of subcellular localization, and is insensitive to errors in N-terminal sequences


Amino Acid Composition:Taking it Further

• How about pairs of consecutive amino acids? (a.k.a 2-grams) How about 3-grams, …, k-grams?

• How about pseudo amino acid composition?

• How about presence of entire functional domains? (I.e. think of the presence/absence of a functional domain as a summary of amino acid sequence info...)


Functional Domain CompositionChou & Cai, JBC, 277:45765--45769, 2002

Training seqs of various localizationsites

BLAST againstdb of known functional domains(SBASE-A)

aminoacid

composition+

Train SVM using these vectors

xi = 1 means ith domain is present


Functional Domain Composition:

Performance

• Not so good• Why? Number of known domains in SBASE-A too small Need to handle situation where a protein has no

hit in known domains



Functional Domain CompositionCai & Chou, BBRC, 305:407--411, 2003

Training seqs of various localizationsites

BLAST againstdb of known functional domains(Interpro)

NN-5875D:Train k-NN (k=1) using these vectors

or, if nohit found

Pseudo aminoacid composition

Aminoacidcomposition

NN-40D:Train k-NN (k=1) using these vectors

If a protein got a hit in Interpro,use NN-5875D; else use NN-40D


Functional Domain Composition:

Performance



Notes


References (Transmembrane)

• Wiess et al. “Transmembrane segment prediction from protein sequence data”, ISMB, 420--421, 1993

• Gavel et al. “The positive-inside rule applies to thylakoid membrane proteins”, FEBS 282:41--46, 1991

• Monne et al. “A turn propensity scale for transmembrane helices”, JMB, 288:141--145, 1999

• Sonnhammer et al. “A hidden Markov model for predicting transmembrane helices in protein sequences”, ISMB, 6:175--182, 1998

• Martelli et al. “An ENSEMBLE machine learning approach for the prediction of all-alpha membrane proteins”, Bioinformatics, 19(suppl):i205--i211, 2003



• Von Heijne. “Membrane protein structure prediction”, JMB, 225: 487--494, 1992

• Jacoboni et al. “Prediction of the transmembrane regions of beta-barrel membrane proteins with a neural network-based predictor”, Protein Sci., 10:779--787, 2001

• Martelli et al. “a sequence-profile-based HMM for predicting and discriminating beta barrel membrane proteins”, Bioinformatics, 18:S46--S53, 2002

• Moller et al. “Evaluation of methods for the prediction of membrane spanning regions”, Bioinformatics, 17:646--653, 2001

• Fariselli et al. “MaxSubSeq: an algorithm for segment-length optimization. The case study of the transmembrane spanning segments”, Bioinformatics, 19:500--505, 2003



• Rost et al. “Transmembrane helices predicted at 95% accuracy”, Protein Sci., 4:521--533, 1995

• Krogh et al. “Predicting transmembrane protein topology with a hidden Markov model: Application to complete genomes”, JMB, 305:567--580, 2001

• Andersson et al. “Different positively charged amino acids have similar effectson the topology of a polytopic transmembrane protein in E. coli”, JBC, 267:1491--1495, 1992


References (Subcellular Localization)

• Horton & Nakai, “Better prediction of protein cellular localization sites with the k-nearest neighbours classifier”, ISMB, 5:147--152, 1997

• Gardy et al., “PSORT-B: Improving protein subcellular localization for Gram-negative bacteria”, NAR, 31:3613--3617, 2003

• Emanuelsson, “Predicting protein subcellular localization from amino acid sequence information”, BIB, 3:361--376, 2002

• Andrade et al., “Adaptation of protein surfaces to subcellular location”, JMB, 276:517--525, 1998

• Yuan, “Prediction of protein subcellular locations using Markov chain models”, FEBS Letters, 451:23--26, 1999



• Emanuelsson et al., “ChloroP, a neural network-based method for predicting chloroplast transit peptides and their cleavage sites”, Protein Sci., 8:978--984, 1999

• Emanuelsson et al., "Predicting subcellular localization of proteins based on their N-terminal amino acid sequence", JMB, 300:1005-1016, 2000

• Hua & Sun, “Support vector machine approach for protein subcellular localization prediction”, Bioinformatics, 17:721--728, 2001

• Reinhardt & Hubbard, “Using neural networks for prediction of the subcellular location of proteins”, NAR, 26:2230--2236, 1998



• Cai & Chou, “Nearest neighbour algorithm for predicting protein subcellular location by combining functional domain composition and pseudo-amino acid composition”, BBRC, 305:407--411, 2003

• Chou & Cai, “Using functional domain composition and support vector machines for prediction of protein subcellular location”, JBC, 277:45765--45769, 2002

• Park & Kanehisa, “Prediction of protein subcellular locations by support vector machines using compositions of amino acids and amino acid pairs”, Bioinformatics, 19:1656--1663, 2003

copyright 2003 limsoon wong recognition of protein features limsoon wong institute for infocomm...

Documents