1. Exact Computation of Coalescent Likelihood under the Infinite Sites Model Yufeng Wu University of Connecticut ISBRA 2009 1

2. Coalescent Likelihood D: a set of binary sequences. Coalescent genealogy: history with coalescent and mutation events. Coalescent likelihood P(D): probability of observing D on coalescent model given mutation rate  Assume no recombination. 00000 0001 0 011 00 1 0000 10001 1 5 2 3 4 Coalescent Mutation

3. Perfect Phylogeny Infinite many sites model of mutations: one mutation per site in history Perfect phylogeny –Site labels tree branches –Each site appears exactly once. –Sequence: list of mutations from root to leaf. –Unique topologically, except root is unknown and order of mutations on the same branch is not fixed. 00000 01100 3

4. Genealogy and Perfect Phylogeny Perfect phylogeny: not enough timing information –Exists many coalescent genealogy for a fixed perfect phylogeny. –Each genealogy: different probability (depending on  ) Coalescent likelihood: sum over all compatible genealogy. 1 5 2 3 4 1 5 2 3 4 1 5 2 3 4

5. Computing Coalescent Likelihood Computation of P(D): classic population genetics problem. Statistical (inexact) approaches: –Importance sampling (IS): Griffiths and Tavare (1994), Stephens and Donnelly (2000), Hobolth, Uyenoyama and Wiuf (2008). –MCMC: Kuhner,Yamato and Felsenstein (1995). Genetree: IS-based, widely used but (sometimes large) variance still exists. How feasible of computing exact P(D)? –Considered to be difficult for even medium-sized data (Song, Lyngso and Hein, 2006). This talk: exact computation of P(D) is feasible for data significantly larger than previously believed. –A simple algorithmic trick: dynamic programming 5

6. Ethier-Griffiths Recursion Build a perfect phylogeny for D. Ancestral configuration (AC): pairs of sequence multiplicity and list of mutations for each sequence type at some time Transition probability between ACs: depends on AC and . Genealogy: path of ACs (from present to root) P(D): sum of probability of all paths. EG: faster summation, backwards in time. (1, 0), (3, 4 0), (1, 3 2 0), (1, 1 0), (1, 5 1 0) (1, 0), (1, 0), (1, 3 2 0), (1, 1 0), (1, 5 1 0) (1, 0), (2, 4 0), (1, 3 2 0), (1, 1 0), (1, 5 1 0) (1, 0), (1, 4 0), (1, 3 2 0), (1, 1 0), (1, 5 1 0) (3, 4 0) 6

7. Computing Exact Likelihood Key idea: forward instead of backwards –Create all possible ACs reachable from the current AC (start from root). Update probability. –Intuition of AC: growing coverage of the phylogeny, starting from root Possible events at root: three branching (b 1, b 2, b 3 ), three mutations (m 1, m 2, m 4 ). Branching: cover new branch Covered branch can mutate Each event: a new AC b1b1 m2m2 b2b2 Start from root AC 7

8. Forward Finding of ACs Finding all ACs by forward looking. Maintain a list of active events in each AC. Update in the new ACs. Rule: at a node in phylogeny, mutated branches  covered branches (unless all branches are covered) Mut branch = covered branch 8

9. Why Forward? Bottleneck: memory Layer of ACs: ACs with k mutation or branching events from root AC, k= 1,2,3… Key: only the current layer needs to be kept. Memory efficient. A single forward pass is enough to compute P(D). 9 Coalescent Mutation

10. Results on Simulated Data Use Hudson’s program ms: 20, 30, 40 and 50 sequences with  = 1, 3 and 5. Each settings: 100 datasets. How many allow exact computation of P(D) within reasonable amount of time? Number of sequences % of feasible data Number of sequences Ave. run time (sec.) for feasible data 10

11. Application: MLE of Mutation Rate Given a set of sequences D, what is the maximum likelihood estimate of the mutation rate  ? Issue: Need to compute for many possible  and root of genealogy is not known. Use exact likelihood –Compute P(D |  ) for  on a grid. –MLE of  : maximize P(D |  ). –Full likelihood: sum P(D |  ) over all possible roots. –Faster computation: compute P(D) for multiple  in one pass. Quicker than computing P(D) for one  each time. 11  P(D) MLE(  )

12. A Mitochondrial Data Mitochondrial data from Ward, et al. (1991). Previously analyzed by Griffiths and Tavare (1994) and others. –55 sequences and 18 polymorphic sites. –Believed to fit the infinite sites model. 12

13. MLE of Mutation Rate for the Mitochondrial Data MLE of  : 4.8 Griffiths and Tavare (1994) IS methods can have variance –Is 4.8 really the MLE? Compute the exact full likelihood for a grid of  between 4.0 to 6.0. 13

14. Conclusion IS seems to work well for the Mitochondrial data –However, IS can still have large variance for some data. –Thus, exact computation may help when data is not very large and/or relatively low mutation rate. –Can also help to evaluate different statistical methods. Research supported by National Science Foundation. 14