Syst. Biol. 50(4):465–467, 2001

The Troubled Growth of Statistical

JOSEPH FELSENSTEIN Department of Genetics, , Seattle, Washington 98919 USA; E-mail: [email protected]

Statistical inference of phylogenies almost of a small band of pioneers exploring new didn’t happen. The story of the origin, territory. growth, and spread of “statistical phyloge- The stage would now seem set for a grad- netics” needs to be told, because it is so ual spread of statistical methods, but reality strange. It is not the straightforward story of was not to be this simple. In 1969 I began gradual spread that one might imagine. to attend the annual Numerical Taxonomy It starts with the development of numeri- conferences convened by Bob Sokal. In cal methods in systematics, whose modern 1971, at that meeting in Ann Arbor, Gareth proponents were Sokal and Sneath. Their Nelson advocated Willi Hennig’s strictly work, embodied in their book Principles of monophyletic classiŽcation. It became clear Numerical Taxonomy, set off an explosion of that some systematists wanted to take a well- work by mathematical clusterers, but did not deŽned, almost algorithmic approach. Hen- win many converts in systematics. In the nig set forth well-deŽned methods for infer- early 1960s two groups started work on nu- ring phylogenies (provided there was no in- merical inference of phylogenies. Edwards ternal conict in the data), an approach with and Cavalli-Sforza, working on trees of enormous appeal to a new generation of mor- human populations and using gene frequen- phological systematists. cies, invented parsimony and distance ma- The difŽculty was that although well- trix methods. In an attempt to choose be- deŽned philosophical principles had great tween these two approaches, they turned to appeal, numerical methods and statistics did maximum likelihood, which proved to be not. Nonetheless, some method was needed different from both of the methods. At about to resolve conict among characters. For the the same time, Camin and Sokal began using new generation of morphological systema- parsimony on discrete characters, partly to tists, parsimony was to play this role. They prove how inaccurate it would be. Their pa- also raised the temperature of debate. Phylo- per in in 1965 was widely noticed. It genetic systematists found themselves trying contained detailed discussions of their meth- to dislodge a dominant school of evolution- ods of searching in tree space for the most- ary systematists who would not make their parsimonious tree. methods explicit and algorithmic. The phy- In 1966 Eck and Dayhoff used parsi- logenetic systematists responded with an in- mony on molecular sequences for the Žrst creasing stridency. time, and in 1967 Walter Fitch and Emanuel Pheneticists were algorithmic but not phy- Margoliash gave the Žrst detailed descrip- logenetic. The rest of the numerical taxon- tion of distance methods and showed their omy crowd were interested in phylogenies application to protein sequences. When I but not committed to Hennig’s message. That started writing phylogeny programs and there would be a schism over this message fretting about the statistical properties of among the numerical types was inevitable, these methods, I had read Sokal and Sneath and a few allied themselves with the phylo- and also Camin and Sokal and had met genetic systematists and moved away from a Sokal. The International Congress of Hu- statistical framework. The polarization was man Genetics in 1966 brought Edwards, increasingly felt within the Numerical Tax- Cavalli-Sforza, and Fitch to Chicago, where onomy Conference meetings, which were I was studying. George Estabrook passed among the only forums for debate on new through our department in Chicago soon systematic methods. More and more mor- after, and I met Steve Farris in 1967 at phological systematists began to show up, a population symposium in Syra- taking the phylogenetic systematic position, cuse, New York. I rapidly got the sense wearing with pride Ernst Mayr’s epithet for

465 466 SYSTEMATIC BIOLOGY VOL. 50 them, “cladists”. The atmosphere grew tense Use Hennig’s terminology—autapomor- and warlike. ² phy, symplesiomorphy, and so forth— At the same time numerical criteria were rather than terms like ancestral or gaining ground in the new Želd of molecular derived. evolution. Molecular evolutionists were not Classify cladistically; use only mono- interested in philosophical frameworks or is- ² phyletic groups. sues of how to classify. They were more prag- Do biogeography by vicariance (pace matic and eclectic. In the early 1970s major ² Hennig). advances were made in the numerical treat- Use only computer programs written by ment of molecular sequences. Fitch discov- ² leaders in the Hennig Society, all others ered his algorithm for counting the minimum are fundamentally awed. number of steps necessary to evolve a DNA Use only parsimony methods. Compati- site on a given phylogeny. David Sankoff ² bility methods are evil. generalized it, and then went on to propose Do not weight characters. a parsimony framework that integrated se- ² Be hostile to molecular data. quence alignment and inference of phyloge- ² Consider your methods to be hypo- nies. In this he was about 20 years ahead of ² thetico-deductive. his time. Fossils are tobe treated the same as living The climax of tensions in the Numeri- ² species. cal Taxonomy meetings came in 1979 at Parasites always have exactly the same Harvard. I recall having to steel myself for ² phylogenies as their hosts. about 6 months in advance just to be able It is important to go around saying that to cope with the meeting. The two camps ² one cannot infer ancestor–descendant sat on on opposite sides of the auditorium relationships. (cladists on the right, others on the left), and It is important to go around saying that the atmosphere was totally partisan. When ² species are individuals, not classes. a speaker from one side made a point, the Be skeptical of the reality of the species other groaned, while adherents murmured ² as nonoperational. approval. It was not a relaxing meeting. I had History: William of Ockham told Popper a premonition of an organizational split, hav- ² to tell Hennig to use parsimony. ing been through similar events 10 years ear- lier in radical student organizations. For those trying to survive outside the fold, The analogy proved exact. About 6 months the early 1980s were scary. We felt that the later the cladists founded the Willi Hennig systematics journals were becoming closed Society as a direct reaction to a breakdown to us. In 1981 I sent a paper on a statistical ap- of negotiations over the format of the next proach to weighting overseas, to the Biolog- Numerical Taxonomy meeting. Many of the ical Journal of the Linnean Society, rather than most important young systematists quickly see it handled by Systematic Zoology. It was joined, in a spirit of crusading zeal. Their in- touch and go as to whether people outside uence in such journals as Systematic Zoology the Hennig Society could continue to work rose rapidly. on phylogenies. There was the recurrent fear Historians like David Hull have tended to that the National Science Foundation (NSF) see this as a struggle over classiŽcation, phy- Systematic Biology program would come to logenetic versus phenetic or evolutionary. It be controlled by reviewers who would auto- was that in part, but in addition many of the matically reject our grants. In fact, this never people who had been developing numerical happened (though I have heard rumors that and statistical methods for inferring phylo- NSF Systematic Biology took formal steps to genies found themselves outside the Hennig ensure that the strife would not affect their Society. And outside versus inside was im- review of proposals). portant. The cladists of that era had accepted The Žghting in systematics grew more in- a number of points as an intellectual pack- tense than in any Želd I have known. I used to age. At one point in the mid-1980s I tried to think that we fought a lot in when I worked summarize the package and came up with in , but in that Želd we these points, in order of importance: used to sit side by side at meetings without 2001 FELSENSTEIN—HISTORICAL ESSAY 467 growing red-faced, hissing at each other, or tic parsimony. Many younger systematists spreading scurrilous rumors. In systematics, think of their inferences of phylogeny as be- however, the controversy attracted extreme ing basically statistical, and their choices of personalities, mostly to the other side of the methods are driven by pragmatism rather issue (probably both sides felt this way; I still than by prior philosophical commitment. do). Within the Hennig Society, an operationalist Systematics and molecular evolution strain of thought emerged in the 1980s as passed through this Dark Ages in the 1980s. “pattern cladistics”; it shared with phenet- By the early 1990s the mood was beginning ics the view that the patterns in the data to lighten. Conicts within the Hennig should be used whether or not they reected Society drove a large fraction of young sys- history. This view was heavily criticized and tematists to desert that organization, though effectively silenced within the society, but its without abandoning their commitment to very critics then adopted a quite similar view phylogenetic systematics. Ed Wiley, Joel when they said that parsimony should be Cracraft, and Vicki Funk took the important used even when it is statistically inconsistent. step of taking over the Society of Systematic Although the cladistic parsimony school in Zoology, playing a role in moving it out the Hennig Society has renewed its criticism of the American Society of Zoologists and of statistical approaches, these have become into meeting with the Society for the Study established as an important part of the stan- of Evolution and the American Society of dard repertory. In the era of Markov Chain Naturalists. These “evolution meetings” Monte Carlo methods, Hidden Markov Mod- grew rapidly, providing a forum where open els, and mathematical genomics, little sign is and friendly exchange of views could take left of the harrowing conicts of the 1980s. place. We old survivors of the systematics wars At the same time, statistical tools for phy- sit around nursing our post-traumatic stress logenies were becoming available, including disorder and telling our war stories to the maximum likelihood, the bootstrap, invari- few who will listen. The late-1990s genera- ants, the KHT test, and parametric bootstrap- tion wonders what these old folks are grum- ping. This was reected in available software bling about. such as PAUP, MacClade, and my own pack- Historians and philosophers of science age, PHYLIP. Computers continued to get have scarcely noticed these developements. faster and could more easily carry out sta- While they looked elsewhere, debate shifted tistical methods. The prestige of parsimony from cladistics-versus-phenetics to whether as the only philosophically sound method inference of phylogenies should be statis- of inferring phylogenies was tarnished by tical. The focus of systematics has shifted the realization that it sometimes would be massively away from classiŽcation: It is the statistically inconsistent. Computer simula- phylogenies that are central, and it is nearly tion tests by John Huelsenbeck and others irrelevant how they are then used in taxon- showed that distance and likelihood meth- omy. Once historians awake to these devel- ods work at least as well as parsimony. opments they will Žnd a fascinating story, There have thus emerged two schools of and one that happens to have a reasonably thought, statistical phylogenetics and cladis- happy ending.