molec evo quiz 7 pt2

Step 1 of a distance method: what is a character in an alignment, and what is lost?

Each alignment column is a character. Distance methods compress all columns into one pairwise number, which is information loss.

What is saturation, and why does it matter for distances?

Multiple substitutions at the same site mean you observe fewer changes than actually happened, so raw p-distance underestimates true distance.

Step 2 of a distance method?

Correct the distances with a substitution model. The larger the uncorrected p-distance, the larger the correction (e.g. A-D 0.46 becomes 0.72; A-B 0.08 stays 0.08).

What does UPGMA stand for and what type of algorithm is it?,Unweighted Pair Group Method with Arithmetic Mean. A simple agglomerative hierarchical clustering algorithm.

List the UPGMA steps.,1) Find the smallest value in the matrix. 2) Join those two taxa. 3) Recompute distances from the new cluster to the rest using arithmetic averages. 4) Repeat until one cluster remains.

How is the distance from a new cluster (AB) to C recomputed in UPGMA?

Arithmetic mean of d(A,C) and d(B,C).

What assumption does UPGMA make, and what tree shape results?,A strict molecular clock (all lineages accumulate equal change since divergence). The tree is ultrametric: every tip is the same total distance from the root. The tree is rooted.

How does rate variation mislead UPGMA? Use the A/B/C/D example.

With d(A,B)=8, d(A,C)=7, d(B,C)=9, D far: B evolved fast and C slightly slower, so UPGMA groups A and C (smallest distance) - grouping by similarity, not ancestry.

How does neighbor-joining (NJ) start?

From a star phylogeny, then it works out how far each taxon is from all the others.

Define net divergence r_i in NJ.,The row sum of the distance matrix for taxon i: its summed distance to all other taxa.

Give the NJ Q matrix formula.

Q(i,j) = (n-2)*d(i,j) - r_i - r_j, where n is the number of taxa and r is net divergence.

Which pair does NJ join each round?,The pair with the smallest Q value; each gets its own branch length. Then recompute the matrix and repeat.

NJ vs UPGMA: clock, branch lengths, rooting?

NJ: no clock assumption, branch lengths vary across lineages, unrooted tree. UPGMA: clock, ultrametric, rooted tree.

NJ is a fast approximation to which criterion?,Minimum evolution.

Pros of distance methods?,Extremely fast (seconds on thousands of sequences); still the method of choice for a quick tree inside a pipeline.

Cons of distance methods?,Collapsing the alignment to a matrix discards information; you cannot trace individual sites; accuracy depends entirely on the distance correction being right.

Why are NJ/distance methods good for guide trees?

They are fast, so they suit initial guide trees such as those used for multiple sequence alignment (MSAs).

Distance methods measure what, versus evolutionary history?

They measure overall similarity, not evolutionary history.

State the parsimony criterion.,The best tree needs the fewest character-state changes. Score each tree by its length (number of changes); the shortest tree wins.

What principle underlies parsimony, and who formalised it?,Occam's razor (fewest assumptions; here an assumption is a state change). Formalised by Willi Hennig (1966); championed by cladists.

What key assumption does parsimony make about homoplasy?,Homoplasy (convergence and reversal) is rare. It happens but is assumed to be the exception.

What is a parsimony-informative site?

A site whose score differs among topologies, so it favours some trees over others (needs at least two states each shared by at least two taxa).

What is a parsimony-uninformative site? Example?

A site that gives the same number of changes on every topology. Site 3 (W:G, X:C, Y:A, Z:A) needs two changes on every tree, so it does not affect the answer.

How is the best tree chosen from informative sites?

Sites vote. In the example, sites 5 and 7 favour grouping (W,X) with (Y,Z) = topology I, which outvotes site 9 favouring another topology.

What is weighted parsimony?,Different change types get different costs. No rate matrix is involved and no probability is calculated with a Markov model.

Parsimony limitations?

Fine for highly similar sequences; branch lengths reflect inferred change, not evolutionary distance; especially susceptible to long-branch attraction; no model of evolution.

What is the cladistics-journal parsimony policy cited in lecture?

Cladistics manuscripts must include a parsimony analysis and not rely solely on model-based methods (ML, Bayesian), or they are considered incomplete and not sent for review.

Why can't we score all trees?

Tree space is too large: 5 taxa give 15 unrooted trees; 10 taxa give 2,027,025.

How does an exhaustive search build trees?

Add taxa one at a time to every branch of every tree built so far, then score all of them.

Describe branch and bound.

Start with a partial tree, add a taxon, compute a lower bound (best possible score any completion could achieve) and compare it to the best complete tree so far. Prune partial trees whose bound is worse. Continue; score complete trees and update the best. Gives the optimal tree(s) under the criterion.

What is a heuristic search?

Modify the current tree to walk through tree space, scoring neighbours (no guarantee of optimality).

Name the three heuristic tree rearrangements, smallest to largest change.,Nearest Neighbor Interchange (NNI): swap neighbouring branches (small). Subtree Prune and Regraft (SPR): cut a subtree and move it (medium). Tree Bisection and Reconnection (TBR): cut the tree in two and reconnect (large).

Why do heuristics get stuck, and how do you reduce it?,They can settle on local optima. Reduce by searching a sufficient amount of tree space (e.g. multiple starts).

Define likelihood in phylogenetics.

L = P(data | hypothesis): the probability of the observed alignment given a tree, branch lengths and a substitution model.

Is likelihood P(tree , data)?|No. Likelihood does not give the probability that a tree is correct.

What is the maximum likelihood tree?

The tree that makes the observed data least surprising (highest P(data | tree, model)).

How does ML compute the likelihood for a single site?,Sum over every possible history (ancestral state assignment) rather than choosing the minimum-change one.

How are site likelihoods combined?

Multiply across sites, or equivalently sum log-likelihoods (log-likelihood avoids numerical underflow).

What is estimated in ML?,Branch lengths (and model parameters) are estimated from the data; the substitution model is chosen and its parameters inferred; topologies are compared by score.

Pros of ML?

Assumptions written as an inspectable, testable model; corrects for multiple hits so it is consistent where parsimony is not; gives branch lengths; allows statistical comparison of trees.

Cons of ML?

Can be slow; only as good as the model, which is always wrong to some degree (a reason cladists object).

What question does Bayesian inference ask versus ML?,ML: how probable is my data given this tree? Bayesian: how probable is this tree given my data?

State Bayes' relation for trees.,posterior is proportional to likelihood x prior. The prior is belief before seeing data; the posterior is belief after.

What is the output of Bayesian phylogenetics?

A distribution over trees, not a single tree.

Label terms in P(A,B).|P(A|B) posterior; P(B|A) likelihood; P(A) prior; P(B) marginal probability.

Spam example: priors P(Spam)=0.4, P(Legit)=0.6; P(FREE

Spam)=0.8, P(FREE|Legit)=0.1. Find P(Spam|FREE).|P(FREE)=0.80.4+0.10.6=0.38; posterior=0.32/0.38=84.2%, up from 40%.

How do data amount affect prior influence?,With plenty of data the likelihood dominates and the prior matters little; with little data the prior contributes more.

What is MCMC and how does it sample the posterior?

A heuristic random walk through tree space that usually accepts uphill moves and sometimes downhill ones, so time spent in each region is proportional to its posterior probability.

Weaknesses of Bayesian inference?,Slow; sensitive to priors; convergence must be checked.

Why not just pick the model with most parameters?

More parameters always fit better (overfitting), so a penalty is needed.

How do AIC and BIC differ?

Both subtract a penalty growing with parameter number. BIC's penalty also grows with data amount, so it picks simpler models.

What is ModelFinder?

A tool inside IQ-TREE that fits dozens of substitution models and reports the best by each criterion (e.g. AIC, BIC).

Which methods need a model choice?,Likelihood and Bayesian methods.

What question does the bootstrap ask?,How consistently does this clade appear if I perturb the dataset (resample sites)?

What can the bootstrap reveal?,Conflicting information between sites; instability from too little data; reliance on a few very informative characters.

What to do with poorly supported branches?,Collapse them (treat as unresolved) when reading the tree.

What is long-branch attraction (LBA)?,Artefactual grouping of fast-evolving (long-branch) lineages because of shared convergent changes.

Which method is most prone to LBA and why?

Parsimony, because it is statistically inconsistent. ML is not immune.

Best fix for LBA?,Add more taxa to break up long branches.

Two categories of failure in tree building?

Data input failures (e.g. paralogs, unbalanced alignments) and tree-building (method/analysis) failures.

Turtle example: effect of 1% of genes?

Of 248 nuclear genes, two were paralogous; removing them changed the turtle placement. Small amounts of bad input can change the result.

Core idea of the topology test paper?

Build an ML tree, build an ML tree under the alternative hypothesis, and compare them (e.g. sponges-first vs ctenophores-first).

What was the procedural error in the high-profile example?

In IQ-TREE, the -z file should contain fully resolved ML trees, but constraint trees were supplied, so the full ML tree was compared with a constrained one. Also unbalanced alignments with critical taxa missing. Moral: check your work.

Tree 1 vs Tree 2 hypotheses in the example?

Tree 1: ((other animals, Sponges), Ctenophores). Tree 2: ((other animals, Ctenophores), Sponges).

A tree-building method is made of what two parts?,An optimality criterion plus a search strategy.

Compare UPGMA, NJ, MP, ML, Bayesian: speed.,UPGMA superfast; NJ superfast; parsimony fast; ML medium to slow; Bayesian slow.

Compare methods: how is the tree chosen?

UPGMA/NJ: clustering (join closest / lowest Q), no alternatives scored. MP: fewest changes. ML: highest P(data|tree,model). Bayesian: sampled from the posterior.

Compare methods: model of evolution and branch lengths.

UPGMA/NJ: only within distance correction; UPGMA lengths forced equal (ultrametric), NJ estimated. MP: no model, lengths = inferred changes. ML: explicit model, lengths estimated. Bayesian: explicit model plus priors, lengths estimated with uncertainty.

Compare methods: main weakness.

UPGMA: clock, similarity not ancestry, discards sites. NJ: discards sites. MP: LBA, no model. ML: only as good as the model. Bayesian: prior sensitivity, convergence checking.

Which methods use pairwise distances vs characters directly?

UPGMA and NJ use pairwise distances; MP, ML and Bayesian use characters directly.

What does bootstrap support tell you overall?,Whether your sites agree on a clade.

Lab 07 goal and platform?,Build and evaluate phylogenetic trees using a Jupyter notebook (Bash kernel) on ICE (GT PACE) plus iTOL in a browser.

What must you do before connecting to ICE?

Connect to the GT VPN (GlobalProtect), then log in at ondemand-ice.pace.gatech.edu with GT credentials and Duo.

Commands to set up the Lab 07 directory?,cd ~; mkdir -p BIOS4225/labs/lab_07; cd BIOS4225/labs/lab_07; cp -r /storage/ice-shared/bios4225/labs/lab_07/* .; ls

What should the lab directory contain?

The notebook plus containers, data and scripts directories.

How do you rename the notebook?,mv Lab_07_building_trees.ipynb Lab07_LastnameInitial.ipynb (e.g. Lab07_LoblawB.ipynb).

How do you start the Jupyter session on ICE?,Interactive Apps > Jupyter; request 2 hours and 1 core; Launch; wait for green card; Connect to Jupyter.

Which kernel should the notebook use?

Bash (top right). If it says Python 3, Kernel > Change Kernel > Bash.

Which three trees get pasted into iTOL?

plant_upgma.nwk, plant_nj.nwk and rotifer_iqtree.nwk (printed in the notebook and on Canvas).

iTOL skills reused from Lab 06?

Upload in the header, pasting a Newick string, the Control panel, and Advanced > Other functions > Midpoint root.

Why do we not need to download from ICE?

Nothing sent to iTOL is generated on the cluster; trees are printed in the notebook or on Canvas, avoiding Globus.

Describe the plant dataset.

rbsml_L5_plant.aln: 17 sequences of ribosomal protein L5 from plants and green algae; small, conserved, easy to align; the expected answer follows the known plant species tree.

Plant taxa roles in the plant dataset?,Chlamydomonas and Ostreococcus: green algae. Physcomitrella: moss. Oryza and Sorghum: grasses.

Why is the plant dataset good for evaluating methods?

The correct answer is known, so you can see what UPGMA, NJ etc. do to it.

Describe the rotifer dataset.,71 glycosyl hydrolase sequences from the Lab 04 MSA: Adineta vaga protein UJR16592.1 plus its 70 BLAST hits.

What was the Lab 04 puzzle and why does a tree help?

An animal protein whose best BLAST matches were overwhelmingly bacterial. A ranked BLAST list is not a phylogeny, so the tree lets you test your hypothesis (e.g. horizontal gene transfer) against an alternative.

What does Quiz 6 cover?,Lecture 7 (tree methods) and the Tuesday lab (constructing and evaluating phylogenetic trees).

What are the next sessions after Lecture 7?

Thursday 15 Oct: molecular clocks and time trees. Tuesday 20 Oct: lab on tree visualisation and hypothesis testing (comparing topologies, consensus trees, testing whether two trees differ).