A Unified Framework for Trees, Multi-Dimensional Scaling and Planar Graphs

TL;DR

This paper introduces a unified framework that applies flexi-Weighted Least Squares to trees, MDS, and planar graphs, comparing their effectiveness in visualizing complex population genetic data.

Contribution

It extends the fWLS method to multi-dimensional scaling and planar graphs, enabling direct comparison of visualization techniques within a common likelihood framework.

Findings

fWLS trees are favored by BIC and are robust under residual resampling.

Neighbor Nets fit data well but are less stable after resampling.

All models exhibit larger errors than sampling variance alone.

Abstract

Least squares trees, multi-dimensional scaling and Neighbor Nets are all different and popular ways of visualizing multi-dimensional data. The method of flexi-Weighted Least Squares (fWLS) is a powerful method of fitting phylogenetic trees, when the exact form of errors is unknown. Here, both polynomial and exponential weights are used to model errors. The exact same models are implemented for multi-dimensional scaling to yield flexi-Weighted MDS, including as special cases methods such as the Sammon Stress function. Here we apply all these methods to population genetic data looking at the relationships of "Abrahams Children" encompassing Arabs and now widely dispersed populations of Jews, in relation to an African outgroup and a variety of European populations. Trees, MDS and Neighbor Nets of this data are compared within a common likelihood framework and the strengths and weaknesses of each method are explored. Because the errors in this type of data can be complex, for example, due to unexpected genetic transfer, we use a residual resampling method to assess the robustness of trees and the Neighbor Net. Despite the Neighbor Net fitting best by all criteria except BIC, its structure is ill defined following residual resampling. In contrast, fWLS trees are favored by BIC and retain considerable strong internal structure following residual resampling. This structure clearly separates various European and Middle Eastern populations, yet it is clear all of the models have errors much larger than expected by sampling variance alone.