Prediction of protein localization and membrane protein topology Gunnar von Heijne Department of Biochemistry and Biophysics Stockholm Bioinformatics Center Stockholm University
Stockholm Bioinformatics Center sorting
Protein localization
Protein sorting in a eukaryotic cell SP
The ’canonical’ signal peptide n-region: positively charged h-region: hydrophobic c-region: more polar, small residues in -1, -3 mTP
mTPs are rich in R & K and can form amphiphilic helices (Abe et al., Cell 100:551) cTP mTP bound to Tom20
Typical chloroplast transit peptide ANN
A simple artificial neural network (ANN) output layer input layer Inside ANN
Artificial neural networks: a summary - a high-quality dataset (positive and negative examples) - an ANN architecture (can be optimized) - all internal parameters in the ANN are systematically optimized during a training session - evaluate the predictive performance using cross- validation ChloroP
ChloroP (Prot.Sci. 8:978) TargetP
TargetP - a four-state SP/mTP/cTP/other predictor (JMB 300:1105) performance
TargetP sensitivity/specificity sensspec SP mTP cTP other sens = tp/(tp+fn)spec = tp/(tp+fp) Other predictors
Other ways to predict localization - amino acid composition - sequence homology - domain structure - phylogenetic profiles - expression profiles Membrane proteins
Popular prediction programs SignalP (NN, HMM) ChloroP TargetP LipoP MitoProt PSORT Membrane proteins
Membrane protein topology
A simulated lipid bilayer (Grubmüller et al.)
Only two basic structures ( Quart.Rev.Biophys. 32:285) Helix bundle ß-barrel Lipid/prot interactions
Most MPs are synthesized at the ER SP
The basic model (courtesy Bill Skach) prediction
Topology prediction
TM helix lengths are typically residues (Bowie, JMB 272:780) Trp, Tyr
Trp & Tyr are enriched in the region near the lipid headgroups (Prot.Sci. 6:808; 7:2026) Loop lengths
Loops tend to be short (Tusnady & Simon, JMB 283:489) PI rule
The ’positive inside’ rule (EMBO J. 5:3021; EJB 174:671, 205:1207; FEBS Lett. 282:41) Bacterial IM in: 16% KR out: 4% KR Eukaryotic PM in: 17% KR out: 7% KR Thylakoid membrane in: 13% KR out: 5% KR Mitochondrial IM In: 10% KR out: 3% KR in out prediction
The positive-inside rule applies to all organisms (Nilsson, Persson & von Heijne, submitted) number of genomes amino acid
Topology can be manipulated (Nature 341:456) Lep constructs expressed in E. coli PK
Topology prediction - a classical problem in bioinformatics 4 characteristics
Three important characteristics ~20 hydrophobic residues predictors ’Positive inside’ rule Trp, Tyr
Popular topology predictors TMHMM (HMM) HMMTOP (HMM) TopPred (h-plot + PI-rule) MEMSAT (dynamic programming) TMAP (h-plot, mult. alignment) PHD (NN, mult. alignment) toppred
TopPred (JMB 225:487) seqanal/interfaces/ toppred.html - construct all possible topologies - rank based on + E. coli LacY TMHMM
TMHMM (Sonnhammer et al., ISMB 6:175, Krogh et al., JMB 305:567) h & l models A hidden Markov model-based method
HMMTOP (Tusnady & Simon, JMB 283:489) performance
Helix & loop models in TMHMM HMMTOP
TMHMM performance (Krogh et al., JMB 305:567; Melén et al. JMB 327:735) Discrimination globular/membrane: sens & spec > 98% Correct topology: 55-60% Single TM identification: sensitivity: 96% specificity: 98% Training set: 160 membrane proteins 650 globular proteins # of TM proteins
Can performance be improved? Consensus predictions Multiple alignments Experimental constraints # of TM proteins
’Consensus’ predictions indicate reliability (FEBS Lett. 486:267) 60 E. coli proteins majority level fraction correct/coverage 5 prediction methods used 46% of 764 predicted E. coli IM proteins are in the 5/0 or 4/1 classes Partial consensus
TMHMM reliability scores (Melén et al. JMB 327:735) TMHMM output: 1. Mean probability p mean 2. Minimum probability p min (label) 3. P bestPath /P allPaths Sequence: M C Y G K C I p(i): p(h): p(o): Label: i i i i i h h S3 results
TMHMM (score 3) Prediction accuracy vs. coverage Test set bias percent correct coverage ~70%~45% 92 bacterial proteins
”Experimentally known topologies” is a biased sample percent score interval Estimate true performance
Correlation between accuracy and TMHMM S3 score mean score percent correct genomes
Expected TMHMM performance on proteomes E. coli S. cerevisiae test set C. elegans coverage percent correct Add C-term.
Original TMHMM prediction, one TM helix missing TMHMM prediction with C-terminus fixed to inside Experimental information helps (JMB 327:735) improvement
When the location of the C-terminus is known, the correct topology is predicted for an estimated ~70% of all membrane proteins (~ 55% when not known) Reporter fusions Experimental information helps (JMB 327:735)