1. FROM PROTEIN SEQUENCES TO PHYLOGENETIC TREES
Simon Harris
Wellcome Trust Sanger Institute, UK

2. Agenda Remind you that molecular phylogenetics is complex
the more you know about the compared proteins and the method used, the better
Try to avoid the black box approach as much as possible!
Give an overview of some phylogenetic methods and software used with protein alignments - some practical issues…

3. * * * 1 2 3 4 5 From DNA/protein sequences to trees Distance methods
Sequence data 2
Align Sequences 3
Phylogenetic signal?
Patterns—>evolutionary processes?
Character based methods
Distance methods * 4
Choose a method
Distance calculation
(which model?) MB ML MP
Weighting?
(sites, changes)?
Model?
Model?
Optimality criterion
Single tree LS ME NJ *
Calculate or estimate best fit tree * 5
Test phylogenetic reliability
Modified from Hillis et al., (1993). Methods in Enzymology 224,

4. Phylogenies from proteins
Parsimony
Distance matrices
Maximum likelihood
Bayesian methods * * *

5. Characters based methods Explicit model of sequence evolution
A G G G C G
B A G G C T
C A T T C T
D A T T T G
Characters based methods
Distance methods
Distance matrix
A B C D
A - d1 d2 d3
B d4 d5
C d6 D MP ML BM
Explicit model of sequence evolution B C D A

6. Phylogenetic trees from protein alignments
Distance methods - model for distance estimation
Simple formula (e.g. Kimura, use of Dij, LogDet)
Complex models
Probability of amino acid changes - Mutational Data Matrices
Site rate heterogeneity
Maximum likelihood and Bayesian methods- MDM based models are used for lnL calculations of sites -> lnL of trees
Homogenous versus heterogeneous models
Estimations of data specific rate matrices (amino acid groupings - GTR like)

7. Software: an overview CLUSTALXv2 - kimura distances
PHYLIPv distance, MP, and ML methods (and more)
Some complex protein models
PAM, JTT ± site rate heterogeneity
Bootstrapping - bootstrap support values
TREE-PUZZLEv5.2 - distance and a ML method
ML - quartet method
Complex protein models
JTT, WAG…matrices ± site rate heterogeneity
From quartets to n-taxa tree - PUZZLE support values
Some sequence statistics - aa frequency and heterogeneity between sequences
Tree comparisons - KH test
PHYML - ML methods (protein and DNA)
MRBAYES - Bayesian
Data partitioning
Posteriors as support values
PHYLOBAYESv2.3 P4
All the things you can dream off… almost… ask Peter Foster
Heterogeneous models among taxa or sites
Estimation of amino acid rate matrices for grouped categories (6x6 rate matrices can be calculated)

8. PHYLIP3.62 Protpars: parsimony
Protdist: models for distance calculations:
PAM1, JTT, Kimura formula (PAM like), others...
Correction for rate heterogeneity between sites! Removal of invariant sites? (not estimated, see TREE-PUZZLE5.2!)
NJ and LS distance trees (± molecular clock)
Proml: protein ML analysis (no estimation of site rate heterogeneity - see TREE-PUZZLE5.2)
Coefficient of variation (CV) versus alpha shape parameter CV=1/alpha1/2
Bootstrapping

9. Distance methods A two step approach - two choices!
1) Estimate all pairwise distances
Choose a method (100s) - has an explicit model for sequence evolution
Simple formula
Complex models - PAM, JTT, site rate variation
2) Estimate a tree from the distance matrix
Choose a method: with (ME, LS) or without an optimality criterion (NJ)?

10. Simple and complex models
dij = -Ln (1 - Dij - (Dij2/5)) (Kimura)
Simple and fast but can be unreliable - underestimates changes, hence distances, which can lead to misleading trees - PHYLIP, CLUSTALX
Dij is the fraction of residues that differs between sequence i and j (Dij = 1 - Sij)
dij = ML [P(n), (G, pinv), Xij] (bad annotation!)
ML is used to estimate the dij based on the sequence alignment and a given model - MDM, gamma shape parameter and pinv - PHYLIP, PUZZLE. Each site is used for the calculation of dij, not just the Dij value.
More realistic complexity in relation to protein evolution and the subtle patterns of amino acid exchange rates…
Note: the values of the different parameters (alpha+pinv) have to be either estimated, or simply chosen (MDM), prior the dij calculations

11. 1) Choosing/estimating the parameter of a model
1) Mutation Data Matrices: PAM, JTT, WAG…
What are the properties of the protein alignment (% identity, amino acid frequencies, globular, membrane)?
Can be corrected for the specific dataset amino acid frequencies (-F) - in some software only
Compare ML of different models for a given data and tree
ModelGenerator and ProtTest are designed for this
2) Alpha and pInv values have to be estimated on a tree
TREE-PUZZLE can do that. Reasonable trees give similar values…

12. 2) Inferring phylogenetic trees from the estimated dij
a) Without an optimality criterion
Neighbor-joining (NJ) (NEIGHBOR)
Different algorithms exist - improvement of the computing
If the dij are additive, or close to it, NJ will find the ME tree
BIONJ, WEIGBOR, FastME
b) With an optimality criterion
Least squares (FITCH)
Minimum evolution (in PAUP - now also PHYLIP)

13. Fitch Margoliash Method 1968
Seeks to minimise the weighted squared deviation of the tree path length distances from the distance estimates -uses an objective function
E = S S wij |dij - pij|a
i=1
j=i+1
T-1 T
E = the error of fitting dij to pij
T = number of taxa
if a = 2 weighted least squares
wij = the weighting scheme
dij = F(Xij) pairwise distances estimate - from the data using a specific model (or simply Dij)
pij = length of path between i and j implied on a given tree
dij = pij for additive datasets (all methods will find the right tree)

14. With Vk being the length of the branch k on a tree
Minimum Evolution Method
For each possible alternative tree one can estimate the length of each branch from the estimated pairwise distances between taxa (using the LS method) and then compute the sum (S) of all branch length estimates. The minimum evolution criterion is to choose the tree with the smallest S value
S = S Vk
k=1
2T-3
With Vk being the length of the branch k on a tree

15. Distance methods Advantages: Disadvantages: Can be fast (NJ)
Some distance methods (LogDet) can be superior to more complex approached (ML) in some conditions (shown for DNA alignments)
Distance trees can be used to estimate parameter values for more complex models and then used in a ML method
Provides trees with branch lengths
Disadvantages:
Can lose information by reducing the sequence alignment into pairwise distances
Can produce misleading (like any method) trees in particular if distance estimates are not realistic (bad models), deviates from additivity

16. Character based methods
Maximum likelihood based methods
Quartet puzzling method - TREE-PUZZLE
Standard ML - PHYML, PROML (PHYLIP)
Bayesian based methods
MrBayes v3.1
Phylobayes v2.3
P4 (Peter Foster)

17. Characters based methods Explicit model of sequence evolution
A G G G C G
B A G G C T
C A T T C T
D A T T T G
Characters based methods
Distances methods
Distance matrix
A B C D
A - d1 d2 d3
B d4 d5
C d6 D MP ML BM
Explicit model of sequence evolution B C D A

18. TREE-PUZZLE5.2
Protein maximum likelihood method using “quartet puzzling”
With various protein rate matrices (JTT, WAG…)
Can include correction for rate heterogeneity between sites - pinv + gamma shape (can estimates the values)
Can estimate amino acid frequencies from the data
List site rates categories for each site (2-16)
Composition statistics
Molecular clock test
Can deal with large datasets
Can be used for ML pairwise distance estimates with complex models - used with puzzleboot to perform bootstrapping with PHYLIP

19. A gamma distribution can be used to model site rate heterogeneity
Yang 1996
TREE, 11,

20. TREE-PUZZLE5.2 The quartet ML tree search method has four steps:
1) Parameters (pInv-gamma) are estimated on a NJ n-taxa tree
2) Calculate the ML tree for all possible quartets (4-taxa)
3) Combine quartets in a n-taxa tree (puzzling step)
4) Repeat the puzzling step numerous times (with randomised order of quartet input)
5) Compute a majority rule consensus tree
from all n-trees - has the puzzle support value
Puzzle support values are not bootstrap values!

22. TREE-PUZZLE5.2 Models for amino acid changes:
PAM, JTT, BLOSUM64, mtREV24, WAG (with correction for amino acid frequencies)
Correction for specific dataset amino acid frequencies
Discrete gamma model for rate heterogeneity between sites 4-16 categories.
-> output gives the rate category for each site. Can be used to partition your data and analyse them separately…
Taxa composition heterogeneity test
Molecular clock test

23. Combination of categories that contributes the most to the likelihood
(computation done without clock assumption assuming quartet-puzzling tree):

24. TREE-PUZZLE5.2
Can be used to calculate pairwise distances with a broad diversity of models - puzzleboot (Holder & Roger)
Can be used in combination with PHYLIP programs for bootstrapping:
SEQBOOT
NJ or LS…
CONSENSE
But PHYML can do ML bootstrapping in a fair amount of time…

25. TREE-PUZZLE5.2 Advantages: Disadvantages:
Can handle larger numbers of taxa for maximum likelihood analyses
Implements various models (BLOSUM, JTT, WAG…) and can incorporate a correction for rate heterogeneity (pinv+gamma)
Can estimate for a given tree the gamma shape parameter and the fraction of constant sites and attribute to each site a rate category
Disadvantages:
Quartet based tree search - amplification of the long branch attraction artefact within each quartet analysis?

26. MrBayes 3.1 Bayesian approach Complex models for amino acid changes:
Iterative process leading to improvement of trees and model parameters and that will provide the most probable trees (and parameter values)
Complex models for amino acid changes:
PAM and JTT, WAG (with correction for amino acid frequencies, but you have to type it!?!?!)
Correction for rate heterogeneity between sites (pinv, discrete gamma, site specific rates)
Powerful parameter space search
Tree space (tree topologies)
Shape parameter (alpha shape parameter, pinv)
Can work with large dataset
Provides probabilities of support for clades (posterior probabilities)

27. MrBayes 3.1
MrBayes will produce a population of trees and parameter values - obtained by a Markov chain (mcmcmc). If the chain is working well these will have converged to “probable” values
In practice we plot the results of an mcmcmc to determine the region of the chain that converged to probable values. The “burn in” is the region of the mcmcmc that is ignored for calculation of the consensus tree
Trees and parameter values from the region of equilibrium are used to estimate a consensus tree
The number of trees recovering a given clade corresponds to the posterior for that clade, the probability that this clade exists
The mcmcmc uses the lnL function to compare trees between generations

28. MrBayes 3.1 Most methods provide a single tree and parameters value
Bootstrapping provide a distribution of tree topologies
Puzzling steps also provides a distribution tree topologies
Bootstrap values - Puzzle support values - Posteriors values ???
But not to sure how to interpret these different support values - in each case the support values are for a given dataset and method used
Posteriors are typically higher then bootstrap and puzzle support values?!

29. MrBayes 3.1: some options

30. MrBayes 3.1: an example Etc… Block I Block II #NEXUS begin data;
dimensions ntax=8 nChar=500;
format datatype=protein gap=- missing=?;
matrix
Etc…
Begin mrbayes;
log start filename=d.res.nex.log replace;
prset aamodelpr=fixed(wag);
lset rates=invgamma Ngammacat=4;
set autoclose=yes;
mcmc ngen=5000 printfreq=500 samplefreq=10 nchains=4 savebrlens=yes
startingtree=random filename=d.res.nex.out;
quit;
end; [
log start filename=d.res.nex.con.log replace;
sumt filename=d.res.nex.out.t burnin=150 contype=allcompat; ]
Block I
Block II

32. #generations (mcmcmc)
A Bayesian analysis
Propose a starting tree topology and parameters values (branch length, alpha, pinv), calculate lnL
Change one of these and compare the lnL with previous proposal
If the lnL is improved accept it
If not, accept it only sometimes
Do many of these…
Plot the change of lnL in relationship to the number of generations run
Determine the region where the chain converged and calculate the consensus tree for that region
-> consensus tree with posteriors for clade support
Tree lnL
Zooming in
Tree lnL
#generations (mcmcmc)

33. #generations (mcmcmc)
alpha
“Burn in” determines the trees to be ignored for consensus tree calculation
Was the chain run long enough?
Do we get the same result from an independent chain?
pinv
#generations (mcmcmc)

34. Consensus tree with a burn in of 1500 (150)
Showing posterior values for the different clades - probability for a given clade to be correct (for the given data and method used!!!) A | B
| C
| |(0.98)
| | | D
| | |(0.99)
|(0.49) E
| F
|(0.96)
| G
|(0.81) H

35. Model choice in protein analyses
Rate matrix choice (20x20 matrices)
WAG, BLOSSUM62, etc…
Recoding protein datasets
20x20 --> 6x6 rate matrix (or else)
Implemented in P4 and PhyloBayes v2.3

36. Effect of using different rate matrices on phylogenetics
PHYML
MtRev matrix
PHYML
WAG matrix
Keane et al. 2006

37. Numerous eukaryotes do not possess mitochondria
They possess instead hydrogenosomes or mitosomes
What is the evolutionary origin of these organelles and what is their function?

38. Trichomonas NuoF localises in the hydrogenosomes
Complex I
News and views by Gray (2005). Nature 434,

39. Amino acids categories - recoding in p4
Sulfhydryl: C (1)
Smallhydrophilic: S, T, A, P, G (2)
Acid,amide: D, E, N, Q (3)
Basic: H, R, K (4)
Smallhydrophobic: M, I, L, V (5)
Aromatic: F, Y, W (6)
1- x1 x2 x3 x4 x5
x6 x7 x8 x9
x10 x11 x12
x13 x14
x15 6-
Recoding into 6 states (1-6) allows the estimation of a GTR like matrix with 14 free parameters

40. Why recode amino acids? Potential advantages:
Allows generation of a rate matrix specific for the investigated alignment
Contributes to mitigating amino acid composition heterogeneity and homoplasy due to frequent changes within categories - equivalent to DNA transversion analyses
Potential disadvantage:
Loss of potential useful signal by reducing the alphabet from 20 to 6 letters (or else)

41. Recoding effect on NuoF phylogeny
GTRrecodedDayhoff classes (6x6)
+pInv+G
WAG (20x20)
+pInv+G
1.0
0.92
1.0
1.0
1.0
a-proteobacteria
0.96
0.8
1.0
1.0

42. Summary
No single program allows thorough phylogenetic analyses of protein alignments
Combination of PHYLIPv3.6, TREE-PUZZLEv5.2, PHYML, MrBAYESv3.1, PHYLOBAYESv2.3 and P4 allow detailed protein phylogenetics
Experimenting with your data and available methods/models can lead to interesting and biologically relevant results (data <-> method)
Incorporate site rate heterogeneity correction in the model or reduce heterogeneity by data editing (with and without invariant sites?)
Partitioning of the alignment (variant - various rates, invariant sites, secondary structure, protein domains…)
Amino acid groupings (6 categories - GTR like)
LogDet for proteins - rare/absent changes? For long alignments?
DNA based LogDet or the protein alignment…?
Do not take support values as absolute. Any support value is for a given method and data, only!

43. outside PM
inside
TM domains
TM domain have very specific structural requirements: AA composition from TM domain is very distinct from non TM domains!
Extracellular and intracellular domains may also have important functional differences --> different functional constrains can lead to different AA composition.
More generally in any protein, surface exposed AA composition is typically distinct from “internal” AA.

44. B C A D B C A D outside PM inside Global alignment (AA)
Sub-alignment 1 B C
Model 1 A D
Sub-alignment 1 B C
Model 2 A D

45. outside PM
inside
Partition 1
Partition 2
Model 1
Model 2 B C A D

46. B C A D outside PM inside Global alignment (AA)
AA recoding, can mitigate compositional difference between domains
Sulfhydryl: C (1)
Smallhydrophilic: S, T, A, P, G (2)
Acid,amide: D, E, N, Q (3)
Basic: H, R, K (4)
Smallhydrophobic: M, I, L, V (5)
Aromatic: F, Y, W (6) B C A D

47. B C A D outside PM inside Global alignment (AA) Global alignment (DNA)
DNA based models
LogDet? Codon 1,2 or 3? B C A D