ayesian Phylogenetic Inference using Sequential Monte arlo lgorithms lexandre ouchard-ôté *, Sriram Sankararaman *, and Michael I. Jordan *, * omputer Science ivision, University of alifornia erkeley epartment of Statistics, University of alifornia erkeley
Phylogenetic tree inference Topic of this talk: integration over the space of trees using Sequential Monte arlo (SM) Motivation: ayesian approach to phylogenetic inference Put a prior on trees, use the posterior for reconstruction Heavy use of integrals over the space of trees: e.g. for handling nuisance parameters, computing minimum risk estimators, ayes factors, etc. Prelude: a parallel with the simpler problem of maximization over the space of trees
Maximization over phylogenies Two strategies: Local and sequential search Key difference: representation Local Sequential State t Trees over the observed species
Maximization: local strategy Meta-algorithm: 1.Start at arbitrary state
Maximization: local strategy Meta-algorithm: 1.Start at arbitrary state 2.Iterate: i. Evaluate neighbors ii. Move to a nearby tree
Maximization: local strategy Meta-algorithm: 1.Start at arbitrary state 2.Iterate: i. Evaluate neighbors ii. Move to a nearby tree 3.Return best state visited Example: stochastic annealing
Maximization over phylogenies Two strategies: Local and sequential search Key difference: representation Local Sequential State t Trees over the observed species Partial state p Forests over the observed species
Maximization over phylogenies Two strategies: Local and sequential search Key difference: representation Local Sequential State t Partial state p Trees over the observed species Forests over the observed species
Maximization: sequential strategy Meta-algorithm: 1.Start at the initial, unconstrained partial state =
Maximization: sequential strategy Meta-algorithm: 1.Start at the initial, unconstrained partial state 2.Iterate: i. Extend partial state ii. Estimate best successor =
Maximization: sequential strategy Meta-algorithm: 1.Start at the initial, unconstrained partial state 2.Iterate: i. Extend partial state ii. Estimate best successor 3.Return best final state = Example: neighbor joining
Parallel lassification of phylogenetic algorithms Local strategy Sequential strategy Maximization Stochastic annealing, Neighbor-joining, Integration
Parallel lassification of phylogenetic algorithms Local strategy Sequential strategy Maximization Stochastic annealing, Neighbor-joining, Integration MM algorithms???
Parallel lassification of phylogenetic algorithms Local strategy Sequential strategy Maximization Stochastic annealing, Neighbor-joining, Integration MM algorithms??? Sequential Monte arlo (SM)
Outline ackground: Importance sampling and Sequential Monte arlo SM for phylogenetic inference Framework for designing proposals Experiments: comparisons with MM
Preview: omparative advantages SM MM + Trivial to parallelize + Easier to get data likelihood estimate + No burn-in + Easier to resample hyper-parameters + Easier to design proposal distribution
Preview: omparative advantages SM MM + Trivial to parallelize + Easier to get data likelihood estimate + No burn-in + Easier to resample hyper-parameters + Easier to design proposal distribution Not exclusive: the two approaches can be combined
Phylogenetic setup: ultrametric trees root G.. G.. } Tree T Hidden sequences X E G..} G.. G.. GGT.. TTT.. GT.. GT.. Observations Y = y
Phylogenetic setup: ultrametric trees root G.. G.. height } Tree T Hidden sequences X E G..} G.. G.. GGT.. TTT.. GT.. GT.. Observations Y = y
Phylogenetic setup: ultrametric trees root G.. G.. height } Tree T Hidden sequences X E G..} G.. G.. GGT.. TTT.. GT.. GT.. Observations Y = y Target distribution: with density: T Y d = π γ(t) Z
Phylogenetic setup: ultrametric trees root G.. G.. height } Tree T Hidden sequences X E G..} Joint density evaluated G.. at G.. (t, y) GGT.. TTT.. GT.. GT.., summing over Observations Y = y hidden x Target distribution: with density: T Y d = π γ(t) Z
Phylogenetic setup: ultrametric trees root G.. G.. height } Tree T Hidden sequences X E G..} Joint density evaluated G.. at G.. (t, y) GGT.. TTT.. GT.. GT.., summing over Observations Y = y x P(Y = y) T Y = d π hidden Target distribution: ata likelihood (intractable) with density: γ(t) Z
Sequential Monte arlo (SM) ackground: Importance Sampling (IS) t1 t2 t3 IS : pproximation for π q 1.Sample trees from a proposal q: ti ~ q
Sequential Monte arlo (SM) ackground: Importance Sampling (IS) IS : pproximation for π w1 w2 w3 1.Sample trees from a proposal q: ti ~ q 2. ompute weights w i = γ(t i )/q(t i ) 3. Normalize weights
Sequential Monte arlo (SM) ackground: Importance Sampling (IS) IS : pproximation for π w1 w2 w3 1.Sample trees from a proposal q: ti ~ q 2. ompute weights w i = γ(t i )/q(t i ) "Particle" 3. Normalize weights
Sequential Monte arlo (SM) ackground: Problem with importance sampling: π is high-dimensional Most particles will have tiny weights
Sequential Monte arlo (SM) ackground: Importance Sampling q π γ Z
Sequential Monte arlo (SM) ackground: Importance Sampling q π γ Z SM: a sequence of proposals q q q q π 1 π 2 π R = π γ Z
Sequential Monte arlo (SM) ackground: Importance Sampling q π γ Z SM: a sequence of proposals q q q q π 1 π 2 π R = π γ Z SM for phylogenies: states (forest) π r are distributions over partial
Sequential Monte arlo (SM) 1. Initialize [] π 1
Sequential Monte arlo (SM) SM : pproximation for π 1. Initialize [] π 2 2. Iterate : i. Sample partial states p i π 1 q π 1
Sequential Monte arlo (SM) SM : pproximation for π 1. Initialize [] π 2 2. Iterate : i. Sample partial states p i π 1 q π 1 p1 p3 p2
Sequential Monte arlo (SM) SM : pproximation for π 1. Initialize [] π 2 p 1 p 2 p 3 2. Iterate : i. Sample partial states p i π 1 q π 1 p1 p3 p2
Sequential Monte arlo (SM) SM : pproximation for π 1. Initialize [] π 2 2. Iterate : i. Sample partial states w1 w2 w3 p i π 1 q ii. ompute weights w i = γ(p i ) γ(p i ) 1 q(p i p i ) iii. Normalize weights
Sequential Monte arlo (SM) π SM : pproximation for π 1. Initialize [] π 2 2. Iterate : i. Sample partial states p i π 1 q π 1 ii. ompute weights w i = γ(p i ) γ(p i ) 1 q(p i p i ) iii. Normalize weights
Intuition: why it works asic result: SM is asymptotically consistent p w = γ(p ) γ(p ) 1 q(p p ) p p w = γ(p ) γ(p) 1 q(p p ) w = γ(p) 1 1 q( p)
Intuition: why it works asic result: SM is asymptotically consistent p w = γ(p ) γ(p ) 1 q(p p ) p p w = γ(p ) γ(p) 1 q(p p ) w = γ(p) 1 1 q( p) w w w = γ(p ) q( p )
Intuition: why it works asic result: SM is asymptotically consistent p w = γ(p ) γ(p ) 1 q(p p ) p p w = γ(p ) γ(p) 1 q(p p ) w = γ(p) 1 1 q( p) w w w = γ(p ) q( p )
Intuition: why it works ompare: Weights along a SM path Importance sampling w w w = γ(p ) q( p ) w = γ(t) q(t)
esigning a proposal q Issue: Over-counting Two ways of t1 t2 One way of building t2 building t1
esigning a proposal q Useful abstraction: q induce a partial order (poset) P
esigning a proposal q Useful abstraction: q induce a partial order (poset) P p1 p2 if q can propose a path from p1 to p2
esigning a proposal q Useful abstraction: q induce a partial order (poset) P Poset s Hesse diagram:
esigning a proposal q Useful abstraction: q induce a partial order (poset) P Poset s Hesse diagram: Use proposal that have tree-shaped Hesse diagrams
esigning a proposal q Example: a proposal that has a tree-shaped Hesse diagram. 1. Pick a pair of trees to merge uniformly at random 2. Pick a height for the new tree such that = height( ) < height( )
esigning a proposal q t1 t2 = < height( ) height( )
Experiments: setup Synthetic-small Synthetic-med Real data Source Generated from the model Subset of HGP Likelihood model rownian motion on frequencies Number of sites 100 11,511 Number of nodes 25 51 25 Number of leaves 13 26 13
Synthetic experiments Goal: comparison against MM ompetitor: standard MM sampler, 4 tempering chains, shared sum-product implementation Metric: symmetric clade difference of the Minimum ayes Risk reconstructed tree to the generating tree atapoints computed by increasing the number of particles (for SM) and the number of sampling steps (for MM)
omparison with MM Synthetic-small Symmetric clade difference 0.4 0.3 0.2 0.1 SM MM 10 100 1000 10000 100000 Wall clock time (ms) in logscale
omparison with MM Synthetic-medium Symmetric clade difference 0.5 0.4 0.3 0.2 0.1 MM SM 10 100 1000 10000 100000 1x10 6 1x10 7 Wall clock time (ms) in logscale
Experiments on real data Goal: show that the method scales to large number of sites Number of particle (10,000) determined using synthetic experiments, timing experiments with different numbers of cores: Wall clock time (ms) 65000 55000 45000 1 2 3 4 Number of cores
onclusion SM can be applied to a wide range of phylogenetic models; previous work limited to oalescent priors [Teh et al. 07] Order theoretic framework for designing proposals Experiments: There are regimes where SM outperforms MM Promising applications of SM in phylogenetic inference: 1. Quickly analyze large datasets 2. Initialization and large step proposal for MM chains