The troubled growth of statistical phylogenetics.

Statistical inference of phylogenies almost didn’t happen. The story of the origin, growth, and spread of “statistical phylogenetics” needs to be told, because it is so strange. It is not the straightforward story of gradual spread that one might imagine. It starts with the development of numerical methods in systematics, whose modern proponents were Sokal and Sneath. Their work, embodied in their book Principles of Numerical Taxonomy, set off an explosion of work bymathematical clusterers, but did not win many converts in systematics. In the early 1960s two groups started work on numerical inference of phylogenies. Edwards and Cavalli-Sforza, working on trees of human populations and using gene frequencies, invented parsimony and distance matrix methods. In an attempt to choose between these two approaches, they turned to maximum likelihood, which proved to be different from both of the methods. At about the same time, Camin and Sokal began using parsimony on discrete characters, partly to prove how inaccurate it would be. Their paper inEvolution in 1965waswidely noticed. It contained detailed discussions of theirmethods of searching in tree space for the mostparsimonious tree. In 1966 Eck and Dayhoff used parsimony on molecular sequences for the Žrst time, and in 1967 Walter Fitch and Emanuel Margoliash gave the Žrst detailed description of distance methods and showed their application to protein sequences. When I started writing phylogeny programs and fretting about the statistical properties of these methods, I had read Sokal and Sneath and also Camin and Sokal and had met Sokal. The International Congress of Human Genetics in 1966 brought Edwards, Cavalli-Sforza, and Fitch to Chicago, where I was studying. George Estabrook passed through our department in Chicago soon after, and I met Steve Farris in 1967 at a population biology symposium in Syracuse, New York. I rapidly got the sense of a small band of pioneers exploring new territory. The stage would now seem set for a gradual spread of statistical methods, but reality was not to be this simple. In 1969 I began to attend the annual Numerical Taxonomy conferences convened by Bob Sokal. In 1971, at that meeting in Ann Arbor, Gareth Nelson advocated Willi Hennig’s strictly monophyletic classiŽcation. It became clear that some systematists wanted to take awelldeŽned, almost algorithmic approach. Hennig set forth well-deŽned methods for inferring phylogenies (provided there was no internal conict in the data), an approach with enormous appeal to anewgenerationofmorphological systematists. The difŽculty was that although welldeŽned philosophical principles had great appeal, numerical methods and statistics did not. Nonetheless, some method was needed to resolve conict among characters. For the new generation of morphological systematists, parsimony was to play this role. They also raised the temperature of debate. Phylogenetic systematists found themselves trying to dislodge a dominant school of evolutionary systematists who would not make their methods explicit and algorithmic. The phylogenetic systematists responded with an increasing stridency. Pheneticists were algorithmic but not phylogenetic. The rest of the numerical taxonomy crowd were interested in phylogenies butnot committed toHennig’smessage. That there would be a schism over this message among the numerical types was inevitable, and a few allied themselves with the phylogenetic systematists and moved away from a statistical framework. The polarization was increasingly felt within the Numerical Taxonomy Conference meetings, which were among the only forums for debate on new systematic methods. More and more morphological systematists began to show up, taking the phylogenetic systematic position, wearing with pride Ernst Mayr’s epithet for