TL;DR
Adding new genetic sequences to existing phylogenetic trees can drastically alter their structure, a phenomenon termed "instability." This study reveals that such instability is surprisingly common (nearly 90% of cases) and often occurs in less supported parts of the tree, not near the new sequence's insertion point. Crucially, machine learning models can effectively predict this instability, offering a path towards more robust online phylogenetic inference methods.
What Problem Does This Paper Address?
As new genetic data pour into public databases, there's a growing need for phylogenetic inference methods that can efficiently incorporate this new information without re-analyzing everything from scratch. These "online" methods are essential for keeping evolutionary trees up-to-date. However, a significant concern is that simply adding a new sequence might fundamentally change the inferred relationships among the already existing taxa. This change in tree topology, termed "instability," could lead to incorrect evolutionary conclusions. While some online methods exist, their reliability in maintaining the integrity of the original tree structure when new data is added remains an open question. This paper investigates the frequency, characteristics, and predictability of this phylogenetic instability.
Methods
The core of this study involves analyzing the stability of phylogenetic trees when a single new taxon (sequence) is added. They defined "stability" as the new tree containing the original tree's topology. To measure instability, they used the Taxon Influence Index (TII), which quantifies the topological difference between the tree inferred with the new taxon and the tree inferred without it (after pruning the new taxon from the full tree). They also introduced the "disruption radius" to measure how far from the new taxon's attachment point the topological changes propagate.
To assess the statistical significance of observed changes, they employed the Approximately Unbiased (AU) test.
For prediction, they developed a workflow that involved:
- Inferring a "full tree" for an alignment.
- For each taxon, creating a new alignment without that taxon and inferring a "pruned tree."
- Computing stability measures (TII, disruption radius) for each taxon.
- Calculating a suite of summary statistics from the alignment and inferred trees, designed to capture potential drivers of instability (e.g., branch lengths, bootstrap support, sequence distances, insertion location properties).
- Training random forest models (both classification and regression) using these summary statistics to predict stability (TII = 0 vs. TII ≠ 0) and the degree of instability (normalized TII, disruption radius).
Datasets
The study utilized a collection of 1000 empirical datasets previously used in large-scale studies of Bayesian inference methods. These datasets were curated from three sources: amniotes, mitochondrial DNA sequences from tetrapod species, and alignments from the PhyLoTA database. The alignments varied in size, containing up to 250 sequences and differing in gene and species composition. The authors specifically noted that while their findings are based on these datasets, other data types, such as viral datasets, might exhibit different behaviors.
Key Insights and Ideas
- Instability is the norm, not the exception: The most striking finding is that phylogenetic trees are highly unstable to the addition of new taxa. This challenges the assumption that simple "insertion-only" online methods will reliably produce globally optimal trees.
- Changes are often distant and subtle: Instability doesn't typically manifest as a local rearrangement around the new taxon. Instead, changes are frequently observed in regions of the tree with low bootstrap support, suggesting that existing uncertainties are amplified.
- Predictability is achievable: Despite the widespread instability, the study demonstrates that it's possible to predict it with high accuracy using machine learning. This opens the door for smarter online inference strategies that can identify when a full re-analysis might be necessary.
- Low bootstrap support is a key indicator: Regions of the tree with low bootstrap support appear to be particularly susceptible to topological changes upon taxon addition, acting as "hotspots" for instability.
- Model choice matters: The study highlights that the choice of substitution model can influence TII values, an important consideration for developing robust online methods.
Key Results
- Frequency of Instability: Nearly 88.62% of taxa exhibited non-zero TII and disruption radius, indicating instability.
- Significance of Instability: While instability was frequent, only 48.46% of taxa showed significant instability as determined by the AU test (P < 0.05). This means over half of the unstable cases were not statistically distinguishable from stability.
- Disruption Radius: For about 51.83% of taxa, the normalized disruption radius was greater than 0.8, indicating that instability often propagated far from the attachment point.
- Predictive Accuracy:
- Random forest classification for predicting TII (stable vs. unstable) achieved an Area Under the ROC Curve (AUC) of 0.94.
- Random forest regression for predicting normalized TII achieved an R² score of 0.98.
- Random forest regression for predicting disruption radius achieved an R² score of 0.94.
- Feature Importance: Key predictors for instability included insertion locations in low bootstrap support regions, Pythia difficulty scores (a measure of inference difficulty), and uncertainties in the placement of the added taxon.
Limitations and Future Work
- Single Taxon Addition: The study focused exclusively on the addition of a single taxon. Future work should investigate the effects of adding multiple taxa simultaneously, as this could lead to more complex and potentially different instability patterns.
- Maximum Likelihood Focus: The analysis was conducted using Maximum Likelihood (ML) inference on gene trees. The authors acknowledge that other inference methods (e.g., Bayesian, distance-based) and data types (e.g., viral genomes) might exhibit different stability characteristics.
- Model Selection: While the study used ModelFinder for substitution model selection, the impact of model choice on TII was noted as a factor to consider for online methods.
- Future Directions: The authors suggest that a pipeline involving efficient taxon placement, stability prediction, and targeted tree updates based on these predictions could form the basis of a robust online phylogenetic inference algorithm. They also propose further research into comparing posterior distributions of trees before and after taxon addition for Bayesian inference.
Summary updated