What use are Exponential Weights for flexi-Weighted Least Squares Phylogenetic Trees?

TL;DR

This paper explores exponential weights in flexi-Weighted Least Squares for phylogenetic trees, showing they can fit evolutionary data better than polynomial weights and are computationally efficient.

Contribution

It introduces exponential weight families for phylogenetic tree estimation, compares their performance with polynomial weights, and demonstrates their effectiveness and fast convergence.

Findings

Exponential weights can outperform polynomial weights in data fitting.

Both weight models approximate model-based variances well.

Iterated least squares converges quickly with minimal fit change.

Abstract

The method of flexi-Weighted Least Squares on evolutionary trees uses simple polynomial or exponential functions of the evolutionary distance in place of model-based variances. This has the advantage that unexpected deviations from additivity can be modeled in a more flexible way. At present, only polynomial weights have been used. However, a general family of exponential weights is desirable to compare with polynomial weights and to potentially exploit recent insights into fast least squares edge length estimation on trees. Here describe families of weights that are multiplicative on trees, along with measures of fit of data to tree. It is shown that polynomial, but also multiplicative weights can approximate model-based variance of evolutionary distances well. Both models are fitted to evolutionary data from yeast genomes and while the polynomial weights model fits better, the exponential weights model can fit a lot better than ordinary least squares. Iterated least squares is evaluated and is seen to converge quickly and with minimal change in the fit statistics when the data are in the range expected for the useful evolutionary distances and simple Markov models of character change. In summary, both polynomial and exponential weighted least squares work well and justify further investment into developing the fastest possible algorithms for evaluating evolutionary trees.