Performance study of phylogenetic methods: (Unweighted) quartet methods and neighbor-joining

We present the results of a large-scale experimental study of quartet-based methods (quartet cleaning and puzzling) for phylogeny reconstruction. Our experiments include a broad range of problem sizes and evolutionary rates, and were carefully designed to yield statistically robust results despite the size of the sample space. We measure outcomes in terms of numbers of edges of the true tree correctly inferred by each method (true positives). Our results indicate that these quartet-based methods are much less accurate than the simple and efficient method of neighbor-joining, particularly for data composed of short to medium length sequences. We support our experimental findings by theoretical results that suggest that quartet-cleaning methods are unlikely to yield accurate trees with less than exponentially long sequences. We suggest that a proposed reconstruction method should first be compared to the neighbor-joining method and further studied only if it offers a demonstrable practical advantage.