Areas under LBA
To detect the areas most effected by LBA, we ran an analysis of the six taxa data sets (see section ) over a range of branch lengths and with two scenarios for the position of the long branches (see Figure 1). Figure 2 shows where the location of LBA, as the black region when the β branches are long and the α branches are shorter under the Felsenstein-like topology. As a control, the Farris-like topology shows how the parsimony bias can be perceived as increased accuracy under the same permutations of branch length. In these figures, the darker the color means the less amount of time the MP analysis and the correct topology were in accordance. In other words, the yellow areas are regions where MP always returned the correct topology (i.e. 100 out of 100 trials) and the black areas are where MP never returned the correct topology. The gradient obviously then covers the percentage of time at intermediate levels of accuracy. What is also interesting to note is the extreme cut off between the areas of correct prediction and those that are incorrect, especially when examining the Felsenstein-like topology of Figure 2. This very black region essentially shows the Felsenstein zone or the conditions under which parsimony suffers from LBA. Figure 3 shows the results for Parsimony under the Farris topology.
One problem with most phylogenetic algorithms is the loss of detectable signal with extremely long trees. The length of the tree is the sum of all the branch lengths it has and those with an extreme length or long trees are difficult to decipher. This problem is clearly visible when examining the upper right of the figures under both topologies. We hypothesis that as the branch lengths get longer the percentage correct will converge to 0.95% as this is a random guess out of the 105 possible topologies.
This analysis served as a search space basis for where LBA should be detected. By comparing the differences between the Felsenstein-like and Farris-like topologies it is clearly visible which areas should be detected. When analyzed with ML these regions do not appear but the loss of signal is still present (see Figure 4). The comparison of both the topologies and the ML method adds further descriptive details and confidence to the search space we are examining.
LBE is not functioning as theory predicts
For a method to accurately detect LBA, it needs to discern between these two types of topologies and find the area of LBA. The region found by searching the branch length space should be the same predicted by LBE. Surprisingly this was not the case.
As is seen in Figures 5 and 6, LBE seems to completely miss the area it is intended to detect (the upper left corner). For a more in depth investigation we analyzed the data by examining specific scenarios that should show extreme LBA. In the majority of cases examined, the parsimony trees outputted in the LBA zone really did suffer from LBA as predicted, but the method failed to recognize it and the short sister taxa of the removed taxa was incorrectly grouped with the other long branch.
Further, LBE predicted LBA under the Farris-like topology, where we know a priori that the data set does not suffer from LBA. A few inconsistent categorizations would be understandable because no method is perfect. But this situation, where similar branch lengths give similar conservative predictions under both topologies, calls into question what the method is actually predicting.
It is consistently classifying the wrong area of the Felsenstein-like topology as LBA and the same area of the Farris-like topology. In reality, this is an area suffering from loss of signal. But even in other areas of loss of signal, i.e. the lower right corner of Figure 2, it is classifying it as not having any LBA. Even though this is technically correct the loss of signal should produce a random-like result in the prediction of LBA, not an extremely confident vote that it is not suffering from LBA. Keeping in mind the method is detecting LBA as the least refuted hypothesis, it seems odd that the only area detected as having LBA is not actually suffering from it and those areas that are suffering from LBA have inconsistent results.
What is more bothersome is that the LBE does not seem to consistently categorize based on specific examples of branch length. Under the full method of LBE with the branch length step included (see section ), the method only categorizes a maximum of 25% of any permutation of α and β as suffering from LBA. When the steps that use branch length estimation (i.e. ML) are removed, the LBE method categorizes more areas with a greater percentage of LBA, (45% in Figure 7) but looses its conservative nature with respect to areas that have lost signal. In this case, it inaccurately predicts a large area that had previously been defined as having lose of signal as having LBA be the least refuted hypothesis.
Why it may not work
Siddall and Whiting make the claim that, “... if each of the two branches individually group in precisely the same place as the other when they are allowed to stand alone in an analysis, one can hardly argue that they are attracted to this placement by the absent branch. [1]” While this seems logical, one needs to remember that a common way to avoid LBA in the first place is to add additional taxa to break up long branches [17, 18]. One possible reason that extracting taxa doesn’t work to detect LBA is that parsimony is sensitive to the removal of taxa, creating artificial long branches in the reran analysis. In the case of our analysis, removing a taxa would still be classified as not LBA because it created an artificially long branch consisting of a full α branch along with a half α branch. This then would attract either the original long branch taxa and it would look the same as the original LBA tree and then be rejected as LBA. In other words the extraction creates a problem with sampling, not splitting up longer branches by adding taxa, a typical pitfall when dealing with LBA. The long branch is not being attracted by the excluded long branch but it is being attracted to the extended branch caused by not breaking it up. This creates a double error and deceives the procedure into thinking it is not a case of LBA
We can thus split the branch length search space into three major areas: the area masked by the ML step (I), the area misled by the artificially long branch (II), and the area that is correct until it reaches a point of loss of signal (III), as seen in Figure 8. Area I can be seen by comparing Figure 5 and Figure 7. The deciding factor when the branch lengths are α ≥ β is the final step that estimates the percentile of the outgroup and questionable taxa are among the top 25%. But we know, based on the design of our experiment, that this will not be the case in this area and so the detection or confusion that it is LBA is masked artificially. This mask is removed when we remove this final step from the analysis, as is seen in Figure 7 and the area looks like a continuation of a loss of signal area.
Area II is much more hypothetical but seems to fit the data reasonably well. When examining Figures 5 and 6 there is a noticeable but rough line at about y = – 2 * x + 2. We hypothesis that the shape of this line is a function of the branch lengths. This area is obviously crucial as seen in Figure 5 because it is the area suffering from LBA. In other words, the predictive power of LBE is being masked by this artificial long branch in the exact area needed for accurate prediction of LBA. This triangle directly corresponds to the areas under LBA, thus making the technique inadvisable.
Finally, area III is where the LBE method is actually mostly correct or the area not suffering from some other artifact. Unfortunately, this area is not suffering from LBA but eventually it losses phylogenetic signal. It is the most clearly seen in Figure 4 where ML can determine to a greater extent the phylogenetic signal. At approximately the same point LBE makes incorrect predictions because of the loss of signal. This area is not under a LBA bias for MP and so is correctly labeled as not having LBA but this is not informative. This really does not add a lot of strength to the procedure because it is already unambiguous.
Long branch shortening
Due to the problems associated with Long Branch Extraction, an alternate approach could be used. Rather than removing the suspected long branch that would cause changes in the overall phylogeny, a series of iterative steps are taken to shorten the branch to diminish the phylogenetic signal being sent from the questionable branch and then see if that changes the phylogeny. If the phylogeny changes, long branch attraction is suspected.
Assuming the questionable taxon (qtaxa) falls basal in the MP analysis and is suspect (this is similar to step 1 of LBE), LBS performs the following three step test:
1. Rather than sampling from all the other taxa, construct the ancestral sequence to all taxa excluding the outgroup and qtaxa. With this sequence, you have the combined signal of all the other taxa, or a summary of that clade.
2. Using the constructed sequence and the questionable taxa, hybridize the two in a random fashion. We are not implying crossing over, albeit that should be tested as well, but using a binomial distribution, characters are exchanged between the sampled ancestor and the suspected long branch ataxa. This causes the branch to be shortened by reducing the differences between the taxon and its hypothetical ancestor. However, since this ancestor is unknown, the characters for the questionable taxon are modified by sampling from the hypothetical ancestor. In Figures 9, 10, 11, 12, different sampling frequencies were used. A sampling frequency of 30% means that a random 30% of the target taxon characters are modified.
3. Re-run the analysis with the hybridized sequence included in place of the qtaxa. If the taxa moves after reducing its own signal and adding some signal from the monophyletic clade you have some evidence of LBA The parameter or probability of switching in the binomial distribution is increased and steps 2 and 3 are repeated until either the probability reaches 1 or consistently (i.e. multiple runs) shows the hybridized qtaxa clading with the hypothetical clade.
One of the weaknesses of such an approach is the lack of an absolute answer. You don’t get a final answer of yes or no (as to whether LBA is occurring) but added evidence that there is a problem. This evidence comes in the form of a probability or percentage of the branch that needs to be shortened to form the monophyletic clade. If the probability comes out high, 0.9 to 1.0, you can be fairly sure that LBA is not occurring that that there is strong phylogenetic signal supporting the current position in the phylogeny. If it is very low, then long branch attraction has occurred and is causing an incorrect tree to be inferred. This evidence can help the researcher to understand if the questionable taxa (qtaxa) is sending a strong signal to be in the current location or a weak one. A weak signal implies that the location is inferred only because of analogous evolution and not homology. This implication can then be interpreted as the determination or detection of LBA.
In Figures 9, 10, 11, 12, dark black indicates regions where Long Branch Shortening(LBS) fails to predict whether long branch attraction is occurring. Gray areas indicate regions where LBS successfully determined whether or not long branch attraction was present in the resulting phylogeny. With a 0% sampling frequency (Figure 9), the target taxon is not modified at all and thus the phylogeny does not change. LBS then reports that no LBA exists anywhere. In this case, the Felsenstein Zone (the black region in the upper left portion of the graph) is clear and LBS is unable to detect long branch attraction. As the sampling frequency increases, the target taxon becomes more like the clade and LBS is more able to detect long branch attraction in the Felsenstein Zone. However, this comes at a price. The region where there is no long branch attraction is now reported incorrectly (the lower left portion of the graph). This is due to the fact that the target taxon has become so much more like the other taxa that at 90% sampling (Figure 12), long branch attraction is always reported because the target taxon always moves; resulting in a different phylogeny.
Maximum parsimony and maximum likelihood
Lastly, as this paper addresses algorithms to detect regions where Maximum Parsimony would report the incorrect tree, it is important to compare Maximum Parsimony and Maximum Likelihood in terms of the best scoring tree versus the true phylogeny. Figures 13 and 14 address this question. In the figures, dark areas indicate topologies where Maximum Likelihood and Maximum Parsimony generated the same phylogeny. In the gray regions, Maximum Likelihood generated the correct phylogeny while Maximum Parsimony failed and in the white regions, Maximum Parsimony generated the correct phylogeny and Maximum Likelihood failed. This shows that in the region where there is not much phylogenetic signal (the upper right region) both methods are equally likely to generate the correct tree. In the region near the Felzenstein Zone, Maximum Likelihood is able to generate the correct phylogeny but only for a small additional part of the region. Much of the Felsenstein Zone is still unable to be determined in both the 4 and 6 taxa cases by either Parsimony or Likelihood.