Phylogenetic supertrees: seeing the data for the trees

1
Phylogenetic supertreesseeing the data for the
trees

• Olaf R. P. Bininda-Emonds
• Technische Universität München

2
Outline

• the fundamental issue characters versus trees
• open questions are trees data?
• loss of contact with primary character data
• loss of information
• novel solutions
• data duplication
• the nature of supertrees
• analytical issues
• conclusions
• are supertrees a valid phylogenetic technique?

3
The fundamental issues
4
The basic distinction

• Supertrees
• source data phylogenies
• basic unit membership criterion / statement of
  relationship
• at best, can be viewed as a proxy for a shared
  derived character

• Conventional studies
• source data measurable attribute of an organism
• basic unit character
• can be viewed as a putative statement of
  relationship

5
The fundamental issue

• supertrees combine trees, not real data
• has led to many criticisms of supertree
  construction
• but also lends advantages to the approach

6
Supertree construction
7
Supertree methods

• Direct
• strict consensus supertrees
• MinCutSupertree (and variants)
• semi-strict supertrees
• Lanyon (1993)
• Goloboff and Pol (2002)

• Indirect
• most matrix representation (MR) supertrees
• parsimony (MRP and variants)
• compatibility (MRC)
• minimum flip supertrees (MRF)
• average consensus (MRD)
• gene tree parsimony

8
Are trees data?
9
Open questions

• loss of contact with raw (character) data
• loss of information
• novel solutions
• data duplication
• the nature of supertrees consensus or
  phylogenetic hypothesis?
• analytical issues

10
Loss of information

• a tree is a graphical representation of the
  primary signal in a character-based data set
• strength of primary signal can be measured (e.g.,
  bootstrap frequencies)
• but information regarding nature of any
  conflicting subsignals lost

11
Potential problems

• all trees and clades on them have equal support a
  priori
• prevents signal enhancement (sensu de Queiroz
  et al., 1995) in combined data sets
• coherent subsignals in different data partitions,
  when combined, outweigh conflicting primary
  signals
• throwing away of information should cause a
  supertree analysis to be less accurate than a
  total evidence one, where primary data are
  combined

12
No loss of accuracy

• simulation studies indicate loss of information
  is not detrimental
• MRP (and variants) (Bininda-Emonds and Sanderson,
  2001)
• average consensus (Lapointe and Levasseur, 2001)
• both methods perform about on a par with total
  evidence analyses of primary character data
• and show similar behaviour to total evidence
  analyses

13
Maximizing contact

• weighting according to evidential support in
  source trees
• possible for all MR methods, average consensus,
  and MinCutSupertree (and gene tree parsimony?)
• causes MRP to outperform total evidence analyses
  of primary character data in simulation
  (Bininda-Emonds and Sanderson, 2001)
• bootstrapping of primary character data
• both non-parametric (Moore et al., in prep) and
  parametric versions (Huelsenbeck et al., in prep)

14
Non-parametric bootstrapped supertrees
)
(
consensus of supertrees
bootstrapped source trees
bootstrapped supertree
original data
15
Open questions

• loss of contact with raw (character) data
• loss of information
• novel solutions
• data duplication
• the nature of supertrees consensus or
  phylogenetic hypothesis?
• analytical issues

16
Novel clades

• all supertree methods have the potential to yield
  novel statements
• relationships between taxa that do not co-exist
  on any single source tree (sensu Sanderson et
  al., 1998)
• defining characteristic of method

17
Unsupported clades

• some supertree methods have the potential to make
  statements that are not only novel, but also
  contradicted (unsupported) by every source tree
• violation of a weaker form of co-Pareto property
• co-Pareto relationship of a given kind in the
  consensus is present in at least one input tree

18

• from Goloboff and Pol (2002)

19
Comparing supertree methods

• indirect, optimization-based methods seem more
  prone to producing unsupported clades

• ?
• MRP (and variants)
• MRF?
• average consensus?
• MinCutSupertree (and variants)?
• gene tree parsimony?

• ?
• strict consensus supertrees
• semi-strict supertrees
• MRC

20
Questions unsupported clades

• how should they be treated?
• how common are they?

21
Appropriateness

• Conventional studies
• unsupported clades (at level of resulting trees)
  arise via signal enhancement
• have direct character support in the combined
  matrix

• Supertrees
• subsignals are invisible
• unsupported clades lack any support among source
  trees ? should be regarded as spurious (Pisani
  and Wilkinson, 2002)
• not equivalent to signal enhancement

22
(No Transcript)
23
Incidence of unsupported clades

• circumstantial evidence hints that they are rare
• only a few reported in the literature
• theoretical Goloboff and Pol (2002) Wilkinson
  et al. (2001)
• empirical Bininda-Emonds and Bryant (1998)
  Wilkinson et al. (2001)
• estimated that 8 of the 198 clades in the
  carnivore MRP supertree ( 4) had no support
  among the source trees (Bininda-Emonds et al.,
  1999)
• dinosaur MRP supertree (Pisani et al., 2002) has
  no unsupported clades

24
Unsupported clades are very rare

• simulation results (MRP only)
• occur most often with source trees that are
• few in number (n 5)
• large in size (up to 50 taxa)
• possess identical taxon sets (consensus
  setting)
• most often means lt 0.21 of all simulated
  clades
• overall incidence was 131 of 282 137 clades (lt
  0.05)
• empirical results
• both the carnivore and lagomorph MRP supertrees
  have no unsupported clades whatsoever

25
Open questions

• loss of contact with raw (character) data
• loss of information
• novel solutions
• data duplication
• the nature of supertrees consensus or
  phylogenetic hypothesis?
• analytical issues

26
Data duplication

• character data are often recycled between
  phylogenetic analyses
• e.g., total evidence analyses, molecular studies
  of the same gene
• the same character data may contribute to more
  than one source tree
• overrepresented in a supertree analysis ? data
  duplication
• also violates assumption of data non-independence

27

• data duplication among cetartio-dactyl source
  trees in the Liu et al. (2001) mammal order MRP
  supertree
• from Gatesy et al. (2002)

28
Minimizing duplication

• data duplication a potential problem for all
  supertree methods
• use of trees does not reveal directly source of
  underlying data set
• but can be minimized / avoided with careful data
  collection protocols
• e.g., supertrees of Daubin et al. (2001) and
  Kennedy and Page (2002) lack data duplication

29
Is data duplication unavoidable?

• no phylogenies are independent given a single
  Tree of Life
• all characters and data sources have been subject
  to the same evolutionary processes and history
• want to combine phylogenetic hypotheses that can
  reasonably be viewed as being independent

30
Is the problem overrated?

• supertrees combine phylogenetic hypotheses
• emergent property ? composed of more than their
  raw character data
• manipulation of data (weighting, alignment,
  recoding)
• method and assumptions of analysis
• for example
• strongly conflicting molecular phylogenies for
  whales can be explained largely by the choice of
  outgroup (Messenger and McGuire, 1998)
• alignment and weighting of primary data also
  important

31
Is data duplication overrated?

• data duplication is often only partial
• most combined data sets represent unique
  combinations of individual data sets
• easy to deal with data sets that are supersets of
  others
• signal enhancement means that each unique
  combination could justifiably be viewed as an
  independent hypothesis
• also independent from constituent data sets

32
Are supertrees unfairly singled out?

• data duplication also exists in conventional
  studies (but less obviously so and to a lesser
  known extent)
• morphological ? single features often described
  by multiple characters
• molecular ? secondary structure (e.g., stems in
  tRNA, protein folding) and codon structure mean
  primary mutations may require secondary
  compensatory ones
• total evidence ? mixing of phenotypic and
  genotypic data must represent data duplication at
  some level

33
Open questions

• loss of contact with raw (character) data
• loss of information
• novel solutions
• data duplication
• the nature of supertrees consensus or
  phylogenetic hypothesis?
• analytical issues

34
The nature of supertrees

• is the supertree itself a legitimate phylogenetic
  hypothesis?
• many would say no, arguing instead that they
  are a
• form of consensus
• historical summary of systematic effort
• therefore, supertrees should not be used to
  answer biological questions

35
Supertrees as consensus

• association derives from
• similar methodology (combining trees rather than
  data)
• both containing polytomies
• resulting topologies may be suboptimal given
  underlying data
• why are consensus trees not valid phylogenetic
  hypotheses?
• especially if polytomies viewed as soft rather
  than hard

36
Dealing with incongruence

• all supertree methods must somehow deal with
  incongruence among source trees
• ignore it strict consensus, semi-strict,
  MinCutSupertree, MRC
• fix it MRF
• explain it biologically gene tree parsimony
• optimize it average consensus and MRP

37
Incongruence as homoplasy

• a repeated criticism of MRP is that inferred
  homoplasy on supertree has no biological meaning
• convergence and reversals meaningless with
  respect to a membership criterion
• but why is MRP singled out?
• similar arguments should apply at least to
  average consensus

38
Parsimony and parsimony

• Principle of parsimony
• a criterion for deciding among scientific
  theories or explanations
• Plurality should not be posited without
  necessity ? choose the simplest explanation of a
  phenomenon

• Cladistic parsimony
• specific application of principle of parsimony
• prefer the tree with the fewest number of
  evolutionary steps (i.e., character state
  changes)
• additional changes over minimum number represent
  homoplasy

39
Homoplasy and supertrees

• notions of homoplasy, convergence, and reversals
  have nothing to do with parsimony per se
• or really even with cladistic parsimony
• post hoc biological interpretation of
  incongruence
• incongruence on an MRP supertree is simply
  incongruence
• idea of homoplasy in this context is
  epistemologically, not biologically meaningless

40
Open questions

• loss of contact with raw (character) data
• loss of information
• novel solutions
• data duplication
• the nature of supertrees consensus or
  phylogenetic hypothesis?
• analytical issues

41
Limitations of total evidence

• analytical limitations of combined primary data
  sets also result in a loss of information
• data must be compatible
• use of a single optimization criterion ? usually
  MP, but ML now also possible
• some data still not analyzable under either
  framework (e.g., DNA-DNA hybridization,
  morphometric data)
• use of simplistic models of evolution
• MP differential weighting (including titv
  ratio)
• ML same model for every partition
• alignment problems

42
Advantages to supertrees

• no loss of information all phylogenetic
  hypotheses can be combined
• even those that arent based on any data
• process amounts to partitioned analyses
• each partition can be analyzed according to most
  appropriate model of evolution, and optimization
  criterion
• can be done in parallel
• results then combined with little loss of
  accuracy
• or hopefully less than loss of information for a
  total evidence analysis entails

43
A phylogeny of mammals

• The superteam
• have complete supertrees for
• Carnivora
• Chiroptera
• Insectivora
• Lagomorpha
• Marsupialia
• Primates
• total of 1923 species (41.5)

• Molecular data
• Murphy et al. (2001a)
• 9779 bp from 18 genes for 64 species
• Madsen et al. (2001)
• 8655 bp from 4 genes for 82 species
• Murphy et al. (2001b)
• 16 397 bp from 22 genes for 44 species (lt 1)

44
Summary
45
Whither supertrees?

• criticisms of supertree construction have been
  launched at two levels
• at the supertree approach as a whole
• at individual supertree methods

46
Of approaches

• supertree problem inherently difficult because of
  missing data
• results in the lack of a single right answer

47
Of approaches

• trees are data
• potential loss of information not detrimental
• key is to think in terms of phylogenetic
  hypotheses
• still awaiting a response from the cladistic
  community

48
and methods

• all methods will go astray if its assumptions are
  violated
• e.g., parsimony and long-branch attraction,
  likelihood and wrong model, regression and data
  non-independence
• for supertrees, key is to try and establish
• what each methods boundary conditions are
• how robust each method is to violations of its
  assumptions
• what the properties of each method are (in
  relation to our desired objective)