Evaluating the Robustness of Process-Pattern Links Using Phylogenies—Insights from a Comparison of Eight Simulation Models
Abstract
Macroevolutionary studies have shown that the shape of phylogenetic trees differs in space, time, and between taxa. It is commonly assumed that these differences in tree shape reflect variability in the underlying ecological and evolutionary processes that produced them, and mechanistic eco-evolutionary models are increasingly used to explore this link. A concern in this context is whether conclusions drawn from such mechanistic models are robust to idiosyncrasies in how eco-evolutionary processes are formalized in the models. Here, we use eight mechanistic macroevolutionary models to study how 52 metrics of phylogenetic tree shape respond to variation in the strength of five fundamental processes: competition, dispersal, environmental filtering, niche conservatism, and speciation. We find that models agree on how some tree metrics respond to changes in these processes, in particular dispersal and speciation. However, no tree metric uniquely correlated with a single process, suggesting that single tree metrics have limited utility as shortcuts for inferring the underlying eco-evolutionary processes. Moreover, although it was possible to infer the underlying processes if the data-generating model was known, inference was not consistent across the different models. We conclude that the relationship between phylogenetic patterns and eco-evolutionary processes in macroevolutionary analysis is likely sensitive to the structural and mechanistic details of how a given process is implemented within models.
What the paper shows and why it matters
(AI-generated)
The shape of phylogenetic trees, the branching family trees of species, differs between regions, periods and organism groups, and this variation is assumed to reflect the ecological and evolutionary processes that produced the trees. Simulation models that build in processes such as competition, dispersal, environmental filtering, niche conservatism and speciation are therefore used to link these processes to tree shape. Turning a verbal process concept into computer code requires many implementation choices (Pontarp et al. 2019), and their consequences are hard to anticipate. Earlier studies typically examined one or two processes and one or two tree metrics at a time. Whether a link between process and tree shape, and any inference built on it, holds regardless of how the model is built was an open question, and it is the one the paper addresses.
The paper finds that links between processes and phylogenetic tree shape differ between simulation models, which suggests that they are sensitive to how each model implements a process, so that inferences drawn with one model carry over poorly to another. It runs eight simulation models across varying strengths of five processes and describes the resulting trees with 52 summary metrics of tree shape. Only for speciation and dispersal do some metrics respond in the same direction in all models that include the process, and no metric responds to one process alone, so no single metric identifies a process. A random forest (a machine-learning method) that uses all metrics recovers process strength for trees from its own model (cross-validated R², the share of variation explained, mostly 0.2 to 0.9). Tested on trees from another model (with a forest restricted to the 32 metrics not strongly correlated with species richness), it transfers less reliably: for speciation it works for several model pairs (not symmetrically), for dispersal mostly not, and for the other three processes generally poorly. For six sister-clade pairs of birds and mammals, the seven models that include speciation disagreed for some pairs on which clade has the higher speciation rate.
Inferring processes from tree shape is therefore possible in principle, but apparently only if the model that generated the data is known. The paper cautions against generalising conclusions from any single simulation framework to real data. The paper recommends checking conclusions against several plausible models and suggests adding other information, such as spatial richness and abundance or trait distributions, as a possible way forward.
References
- Pontarp et al. (2019). The Latitudinal Diversity Gradient: Novel Understanding through Mechanistic Eco-evolutionary Models. Trends in Ecology & Evolution 34, 211–223. doi:10.1016/j.tree.2018.11.009