Bootstrapping in Phylogenetics – Tree Reliability, Resampling & Bootstrap Values

Bootstrapping is a statistical method used to assess the reliability of a phylogenetic tree, introduced by Felsenstein in 1985. Since its introduction, it has become one of the standard tools routinely applied in phylogenetic studies. The method addresses an important problem: since substitutions in sequences occur as random events, the number of substitutions observed on any given branch of a real tree can deviate considerably from the mean value predicted by the evolutionary model being used. As a result, the distances measured between sequences are subject to chance fluctuations, and it becomes necessary to determine whether the tree topology obtained is a genuine reflection of the underlying phylogenetic signal or whether it has been influenced by such random fluctuations. Bootstrapping provides a way of answering this question.

1. Basic Concept

• Substitutions in sequences are random events; even under a correct evolutionary model, the number of substitutions on a branch can deviate substantially from the expected mean.
• This means distances measured between sequences are subject to chance fluctuations, and we need to know whether these fluctuations are influencing the tree obtained.
• Bootstrapping generates multiple new data sets that differ slightly from the real data through controlled random resampling.
• Tree-construction is repeated on each resampled data set to see whether the same tree topology is obtained.
• If phylogenetic signal is strong: information about relatedness is spread throughout the whole sequence length → resampling makes very little difference → same topology recovered repeatedly → well-supported clade.
• If phylogenetic signal is weak: the noise introduced by resampling may be sufficient to change the result → tree-construction method gives a different topology on different replicates → weak/unreliable support for that grouping.

2. Method of Resampling

• Randomized data sets are generated by resampling the columns of the original alignment, not the sequences themselves.
• Each bootstrap alignment is of the same length as the original.
• Every column in the new alignment is created by copying one column chosen at random from the original.
• Sampling is done with replacement:
  • Some columns may be selected more than once.
  • Some columns may not be selected at all.
• Because of this, each bootstrap alignment carries slightly different information from the original.

Difference from Reshuffling

• Phylogenetic methods treat each column of an alignment independently of the others.
• Reshuffling = same set of columns, only rearranged into a different order. Since column order does not matter to the method, a reshuffled alignment gives exactly the same answer as the original sequences, and therefore tells us nothing new.
• Bootstrapping (resampling with replacement) actually changes which columns are present, and how many times → produces genuinely different, informative data sets.

3. Procedure for Bootstrap Analysis

1. Generate many randomized data sets by resampling columns (usually 100 or 1,000 replicates).
2. Apply the same tree-construction method (e.g., Neighbor-Joining) to each randomized data set.
3. Some resulting trees match the original tree; others differ.
4. For each clade in the original tree, calculate the percentage of bootstrap trees containing that same clade.
5. This percentage = bootstrap value, a measure of confidence in that clade.

4. Interpretation of Bootstrap Values

Example from primate NJ tree (1,000 replicates):

• 100% support: Hominidae, Cercopithecidae, Catarrhini, Platyrrhini, tarsiers, tree shrews.
• 90% support: Prosimian group (lemur + bushbaby).
• New-world monkeys as a whole: 100% support, but low values within the group → branching order not well resolved.
• Weakest point (24%): node linking prosimians, tarsiers, and tree shrews → very low confidence; likely incorrect grouping (confirmed by other evidence).

General Guidelines

• No fixed rule for a “safe” cutoff.
• Values above ~70% are generally considered reasonably strong evidence.
• High bootstrap value ≠ guaranteed correctness — e.g., tarsier + tree shrew grouping had 75% support but is likely wrong; this incorrect relationship disappeared when a more realistic evolutionary model was used.
• Conclusion: Bootstrap values indicate reliability but do not “prove” a clade is true.

5. Example: Human, Chimpanzee, and Gorilla Relationship

• Node showing (human + chimpanzee + pygmy chimpanzee) clade had 71% support.
• Sandwiched between two very high-support nodes:
  • Two chimpanzee species together: 96%
  • Gorilla + human + both chimpanzees together: very high/near 100%
• Uncertainty concerns only the branching order among gorilla, human, and chimpanzees.

Three possible topologies:

1. (gorilla, (human, chimpanzees)) – topology actually obtained; gorilla most distant.
2. (human, (gorilla, chimpanzees))
3. (chimpanzees, (gorilla, human))

• Topologies 2 and 3 appeared occasionally in bootstrap replicates but far less frequently than topology 1.
• Current scientific consensus: humans and chimpanzees are the closest pair of the three — consistent with the obtained topology, though this was debated in the past.

6. Consensus Trees

• Bootstrap results are often summarized as a consensus tree.

Construction Steps

1. Determine the frequency of occurrence of every possible clade across all bootstrap trees.
2. Rank clades in descending order of frequency.
3. Build the consensus tree by adding clades one at a time from the top of the ranking, adding only clades that are consistent with clades already added.

Key Points

• Consensus topology may differ slightly from the tree built using the original full data set.
• Choice of presentation:
  • Original tree labeled with bootstrap percentages, OR
  • Consensus tree, which may show slightly higher support for some clades not in the original tree.
• Well-supported clades (high bootstrap values) appear in both the original and consensus trees.
• Difference in presentation mainly affects only the least well-determined, weakly supported parts of the tree.

(Quick Recap)

• Bootstrapping = statistical technique to test reliability of a phylogenetic tree (Felsenstein, 1985).
• Works by resampling alignment columns with replacement and rebuilding the tree repeatedly.
• Bootstrap value = % of replicate trees containing a given clade.
• >70% generally = strong support; low values = unreliable branching.
• High values do not guarantee biological correctness (caution needed).
• Results often summarized as a consensus tree built from clade frequencies.