Inferring reticulate evolution networks from consensus gene trees

1
Inferring reticulate evolution networksfrom
consensus gene trees

• Samantha Riesenfeld
• Richard Karp
• Dept. of Computer Science
• U.C. Berkeley

2
Reticulate evolution

• Phylogenetic tree model
• requires independence of evolutionary lineages
• What about when lineages mix?
• very common in plants, fish, bacteria
• phylogenetic networks
• Two main types of reticulate evolution
• hybrid speciation
• lateral gene transfer

3
Types of reticulate evolution

• Hybrid speciation

• Lateral gene transfer

4
Network reconstruction

• Goal given data from a set of species
• e.g. DNA sequences, gene-order data, etc.
• reconstruct the phylogenetic network
• Approach 1 combined analysis
• combine data sets reconstruct network
• e.g. NeighborNet (Bryant Moulton 02)
• Approach 2 separate analysis
• reconstruct separate gene trees from data
• infer network from gene trees

5
Our approach separate analysis

• Works with any phylogenetic tree model,
  reconstruction method
• Offers good understanding of the history
• evolution of different parts of genetic data
• Promising experimental results
• SpNet (Nakhleh et al. 04) shows significant
  improvement over existing combined analysis
  program (NeighborNet)

6
gt-networks

• Reconstructing general networks is difficult
• start with a restricted class of networks
• Use galled-trees, or gt-networks, as model
• (Wang et al. 01, Gusfield et al. 03)
• Reticulation events are independent
• The network is a Directed Acyclic Graph in which
  no node appears in more than one reticulation
  cycle, or gall

7
gt-network with two induced trees
?
l?
w2
w3
x3
x3
v2
u2
u3
v3
w1
y
1
7
8
9
10
3
5
u1
v1
x1
4
2
6
8
the binary gene trees from the network
9
Reconstruction from binary trees

• Reconstruction from binary gene trees
• Efficient algorithm by Nakhleh et al. for
  reconstructing gt-networks from binary trees
• Problem the trees need to be accurate
• rarely is one tree unambiguously reconstructed
• Idea use consensus trees
• for each genetic region, take a consensus tree
  of the top-ranked gene trees

10
Getting a consensus tree (1)
11
Getting a consensus tree (2)
12
New input consensus trees
13
Reconstruction from consensus trees

• Nakhleh et al give algorithm to reconstruct a
  network with just one reticulation event
• Implemented in SpNet
• Need for broader applicability
• What about networks with multiple reticulation
  cycles?
• Problem is more complicated
• structure is lost in the consensus trees

14
Results our algorithm

• Input q consensus trees (q 2)
• the leaves are labeled with the same n taxa
• Output a gt-network, if one exists, that
  contains refinements of all the input trees
• network has m cycles (m is minimal, m 1)
• Runs in polynomial time (i.e. O(qmn2))
• no algorithm known previously that runs in less
  than exponential time (i.e. qnO(m))
• a significant improvement for large n

15
Whats next?

• Experimental testing
• gt-network reconstruction on simulated data
• How often are real networks galled?
• Open questions
• Reconstructing a more general class of networks?
• Dealing with the confounding factors of gene
  duplication and loss?
• Considering the case when not all gene trees
  include all taxa?

16
Thank you