Evolutionary shift detection with variable selection method

Abstract

Abrupt environmental changes can lead to evolutionary shifts in trait evolution. Identifying these shifts is an important step in understanding the evolutionary history of phenotypes.

We first introduce ELPASO, an ensemble variable selection method implemented as an R package, designed to detect shifts in optimal trait values. We evaluate its performance against existing methods (R packages l1ou and PhylogeneticEM) across various scenarios. Our results reveal that the choice of selection criterion significantly impacts performance: methods using the Bayesian information criterion (BIC) perform better when signal sizes are small, while the phylogenetic BIC (pBIC) excels when signal sizes are larger. Shifts near the tips of the phylogenetic tree are more difficult to detect, whereas those closer to the root are more easily identified. The performance of all methods is compromised by factors such as shifts in variance, measurement error, and errors in phylogenetic tree reconstruction.

Next, we address shifts in both optimal trait values and variances by introducing ShiVa, a method based on the multi-optima, multi-variance Ornstein-Uhlenbeck (OU) process. ShiVa uses L1 regularization to detect shifts in both variance and optima. Simulation results show that ShiVa has strong predictive ability, particularly when shifts in evolutionary variance are present.

Recognizing that neglecting measurement error can lead to inaccurate shift detection, we developed ShiVa-ME, an extension of ShiVa that explicitly accounts for measurement error. By incorporating L1 regularization for shifts in both optima and variances and utilizing a gradient-based approach to estimate measurement error variance, ShiVa-ME provides more accurate modeling of noisy data. Simulations show that ShiVa-ME outperforms existing methods (l1ou, PhylogeneticEM), particularly in the presence of measurement error, achieving higher predictive log-likelihoods and more accurate estimates of evolutionary and measurement error variances, while reliably identifying true evolutionary shifts.