ML tree inference using gap-coding Derrick J. Zwickl and Mark T. - - PowerPoint PPT Presentation

ML tree inference using gap-coding

Derrick J. Zwickl and Mark T. Holder

• Dept. Ecology and Evolutionary Biology
• Univ. Kansas

Lawrence, Kansas

Acknowledgments Derrick Zwickl Jeet Sukumaran, Jamie Oaks, Tracy Heath, and Jiaye Yu NSF ATOL

Talk outline

• Background
• Simple approaches based on gap-coding:

– Dollo Indel Mixture model. – Mkv with ascertainment bias correction for at least one residue in a column.

• Some very preliminary results from an implementation in

GARLI.

• Future work

• Unaligned sequences, U, are observed.
• The tree, T, which maximize Pr(U|T), is the ML tree (or

we could maximize Pr(T|U) in a Bayesian framework).

• We could use MCMC to marginalize over every possible

alignment, A: Pr(U|T) =

• A

Pr(U, A|T)

• Too computationally intensive for trees with hundreds or

thousands of leaves.

• See Ben Redelings talk tomorrow !

Two-phase methods

• 1. Infer a single alignment, A from the unaligned sequences,

U.

• 2. Infer tree, T, given the alignment, A, and sequences. For

example, the tree that maximizes Pr(U, A|T). Weaknesses:

• 1. Ignores alignment uncertainty by treating {U, A} as if it

were the known data.

• 2. Ignores the fact the progressive alignment strategies use a

guide tree to infer A

Two-phase methods with gaps as missing data

• 1. Infer a single alignment, A. Map {U, A} → Z where Z is

an aligned sequence matrix in which all gaps are treated as missing data.

• 2. Infer tree, T, given Z. For example, the tree that maximizes

Pr(Z|T, θ). Weaknesses:

• 1. Ignores alignment uncertainty by treating {U, A} as if it

were the known data.

• 2. Ignores the fact the progressive alignment strategies use a

guide tree to infer A

• 3. Ignores phylogenetic information from indels.

Addressing alignment uncertainty without marginalizing

• ver all plausible alignments
• Masking or “culling”
• Wheeler’s “elision” method of concatenating alternative

alignments

• SAT´

e – iteratively considering multiple alignments and trees. Results in an estimate of a tree and alignment pair.

• Resampling procedures which are cell-based rather than

column-based: – Jamie Oaks (KU grad. student) is currently testing a cell-based jackknifing procedure that will try let alignment uncertainty affect clade support.

This talk: ML methods that do not ignore information from gaps

• tree inference with the alignment treated as fixed.
• considering single-residue indel events (considering multiple-

residue indel events means that columns an alignment are no longer independent).

• focussing on methods that can be fit well with:

– large-scale tree inference software (GARLI) and – high-performance likelihood calculation library (lib-beagle).

General strategies for fixed alignment, single-residue indel handling Rather than {U, A} → Z where Z is an aligned matrix with gaps converted to missing data:

• Analyze the aligned sequences, {U, A} and treat gaps as a

“5th base” or “21st amino acid”

• “Gap-code” A as binary matrix (1 for presence of a residue,

0 for a gap). Next, analyze Z under a model of substitution and A under a model with a state space of {0, 1}

“5th base” coding Weaknesses:

• The simplest form would allow evolutionary transitions such

as C → G → — → G But, the first and second G residues should not be considered to be homologous!

• Gap states evolve in a “5th-base” coding, but it is odd to

model the evolution of a site that is not there.

• Indels of multiple residues are not handled nicely.

Better “5th base” coding Rivas and Eddy (2008):

• provide a nice discussion of these issues
• develop techniques for computing the transition probability

matrices for “extended” state models

• point out the need to correct for ascertainment bias
• develop

altered versions

• f

the Felsenstein’s pruning algorithm so that at most one branch per tree can have an insertion at a site

• implemented in their software DNAMLE

Stochastic Dollo Alekseyenko, Lee, and Suchard (2008) generalize “stochastic Dollo” models that prohibit re-evolution of a complex character. Their approach could be used to model insertion of a site: Their Figure 4:

Dollo indel mixture model - motivation

Like Rivas and Eddy (2008) and Alekseyenko et

• al. (2008), we would like an indel model that

strictly enforces homology. But we would like most of the likelihood calculations to use the standard pruning algorithm of Felsenstein, so we can use highly optimized code (lib-beagle, GARLI . . .).

Dollo indel mixture model - separating substitution calculations from indel calculations ai refers to the gap state at column i of an alignment. ui is the sequence state at this position of the aligned matrix. zi is the states with gaps mapped to missing. Pr(ui, ai|T) = Pr(zi, ai|T) = Pr(zi|T, ai) Pr(ai|T) We can calculate Pr(zi|T, ai) from the standard pruning algorithm.

Dollo indel mixture model - gap coding For a residue there are 3 states indicating presence/absence: never had it → has it → lost it But if we “gap-code” a site we see two observable states: = never had it OR lost it 1 = has it We can use a model with 3 possible states and the modified pruning algorithm to assure that there can be no more than

• ne insertion per column when we calculate Pr(ai|T)

Dollo indel mixture model - mixture part Rather than infer the insertion rate, we can use a simple mixture approach: Pr(ui, ai|T) = Pr(ui, ai|T, 0 ins.) Pr(0 ins.) + Pr(ui, ai|T, 1 ins.) Pr(1 ins.) If φ is the probability of a site having having an insertion: Pr(ui, ai|T) = Pr(ui, ai|T, 0 ins.)(1 − φ) + Pr(ui, ai|T, 1 ins.)φ The max. likelihood estimate of φ can be found very quickly.

Mk not-fully-absent variant For comparison, we can also analyze the gap-coded data using:

• 2-state model that allows re-insertion,
• standard pruning algorithm.

We should condition our inference on the fact that our data is censored because ai cannot be all gaps. This is slight tweak of the Mkv model (the “morphology model”) of Lewis (2001).

Very preliminary results - simulation study

• Simulation on a tree of estimated for 64 species of Rana.
• Sweeps of simulation of indel rates, indel lengths, and tree

depths.

• Tree inference in GARLI under:

– GTR model ignoring indels, – GTR + Dollo Indel Mixture Model (DIMM), and – GTR + Mkv variant

Results on true alignment

0.02 0.04 0.06 0.08 0.1 0.04 0.08 0.16 0.32

Tree Depth 2.0 Error Deletion + Insertion Rate

dnaMix dna dnaMk 0.02 0.04 0.06 0.08 0.1 0.04 0.08 0.16 0.32

Tree Depth 1.0 Error

dnaMix dna dnaMk 0.02 0.04 0.06 0.08 0.1 0.04 0.08 0.16 0.32

Tree Depth 0.5 Error

dnaMix dna dnaMk 0.02 0.04 0.06 0.08 0.1 0.04 0.08 0.16 0.32

Tree Depth 0.1 Error True Alignment

dnaMix dna dnaMk

Deletion + Insertion Rate

Results on MAFFT alignment

Tree Depth 2.0 Error Tree Depth 1.0 Error Tree Depth 0.5 Error Tree Depth 0.1 Error

0.1 0.2 0.3 0.4 0.5 0.6 0.04 0.08 0.16 0.32

Deletion + Insertion Rate

dnaMix dna dnaMk 0.1 0.2 0.3 0.4 0.5 0.6 0.04 0.08 0.16 0.32 dnaMix dna dnaMk 0.1 0.2 0.3 0.4 0.5 0.6 0.04 0.08 0.16 0.32 dnaMix dna dnaMk

MAFFT Alignment Deletion + Insertion Rate

0.1 0.2 0.3 0.4 0.5 0.6 0.04 0.08 0.16 0.32 dnaMix dna dnaMk

Preliminary Results Summary

• DIMM and Mkv variant are able to use indel information to

improve tree inference when the alignment is reliable

• DIMM

is very sensitive to alignment error (though the MAFFT alignments are quite compressed for these simulations).

• Mkv variant is much less sensitive to alignment error.
• Analyses with Prank alignments are currently underway.
• Analyses using these gap-coding methods are available

in GARLI (demo later today; note: not fully optimized).

Future work

• 1. Multiple tiers of “masking” like approaches
• 2. Comparing Mkv variant and DIMM to identify compression
• 3. Elision-like approaches in ML
• 4. Integration of these approaches within SAT´

e

Future work - richer masking/culling to classify columns

Rather than: We can: included included, with DIMM included, with Mkv variant included, gaps ignored excluded excluded

Future work - identifying compression Mkv allows any number of insertions in a column, while DIMM forces there to be 0 or 1. We can compare partial likelihoods from both models to identify sites that show strong evidence

• f multiple insertion events.

Future work - more appropriate “Elision” method for ML We will be adding the ability to score multiple possible alignments to GARLI. The likelihoods rather that log-likelihood will be added across alternatives (see Redelings and Suchard review chapter 2008).

(e)

X Y Z 1 2 3 4 5 6 7 8 9 1 1 1 0 1 2 Outgroup T A G A G C A C T C A G Taxon A T A G A G C A C T C A G Taxon B T A G T G A A G C C A G Taxon C T A G T G A A G C C A G Taxon D T A G - - - A G C C A G Taxon E T A G - - - A G C C A G X Y Z 1 2 3 4 5 6 7 8 9 1 1 1 0 1 2 T A G A G C A C T C A G T A G A G C A C T C A G T A G T G A A G C C A G T A G T G A A G C C A G T A G A G C - - - C A G T A G A G C - - - C A G

from figure M. S. Y. Lee, TREE, 2001

Future work - integration with SAT´ e

• In the next few weeks we will be adding GARLI as a tree

inference option for SAT´ e

• The subalignment stage of SAT´

e should be an excellent step to incorporate automated masking, compression correction,

• r elision recoding into SAT´

e.

Thanks!

Questions, suggestions, collaborations, and donation

• f test datasets are always welcome!