T. (Tianjian) Qin
Research interests
PhD thesis: Diversification models and neural inference (2026)
This thesis investigates the drivers of species diversification, and to what extent deep learning methods can recover different drivers from extant phylogenies. In particular, I focus on the effects of ecological limits (constraints to the numbers of species that can coexist) and evolutionary relatedness on speciation and extinction rates. First, I introduce a birth--death model, eve, in which speciation rates depend on both species richness and evolutionary relatedness (ER) measured at different phylogenetic scales. I show that tree shape and the distribution of speciation rates across lineages depend strongly on the scale at which ER acts, and that negative species richness dependence can partly mask the influence of ER on standard tree statistics. The model generates a wide range of empirically realistic, often imbalanced, phylogenies.
Second, I develop an ensemble neural-network framework for parameter estimation in diversification models. Combining dense, graph and recurrent neural networks trained on tree topologies, branching times and summary statistics, the method yields estimates faster than maximum-likelihood approaches and is less sensitive to tree size for constant-rate and diversity-dependent models. However, both likelihood-based and neural network estimators struggle under protracted speciation, highlighting limits imposed by the information content of trees.
Third, I use the eve model---which couples ecological limits with ER effects on speciation and extinction---as a testbed to map when neural networks can and cannot infer diversification mechanisms from phylogenies. In many cases the neural networks struggle to tell the three scenarios apart, and when the trees carry little information the estimated parameters tend to drift back toward average values. Strong global richness dependence further erodes recoverability, whereas sufficiently strong ER effects can create narrow regions of practical identifiability. Together, these results delineate the prospects and limits of using flexible diversification models and deep learning to unravel evolutionary dynamics from extant phylogenies.
