History of speciation

From Wikipedia, the free encyclopedia

Charles Darwin in 1868

 

The scientific study of speciation — how species evolve to become new species — began around the time of Charles Darwin in the middle of the 19th century. Many naturalists at the time recognized the relationship between biogeography (the way species are distributed) and the evolution of species. The 20th century saw the growth of the field of speciation, with major contributors such as Ernst Mayr researching and documenting species' geographic patterns and relationships. The field grew in prominence with the modern evolutionary synthesis in the early part of that century. Since then, research on speciation has expanded immensely. 

The language of speciation has grown more complex. Debate over classification schemes on the mechanisms of speciation and reproductive isolation continue. The 21st century has seen a resurgence in the study of speciation, with new techniques such as molecular phylogenetics and systematics. Speciation has largely been divided into discrete modes that correspond to rates of gene flow between two incipient populations. Today however, research has driven the development of alternative schemes and the discovery of new processes of speciation.

Early history

The only figure in Darwin's 1859 On the Origin of Species, a tree of lineages splitting to form new species

 

Charles Darwin introduced the idea that species could evolve and split into separate lineages, referring to it as specification in his 1859 book On the Origin of Species. It was not until 1906 that the modern term speciation was coined by the biologist Orator F. Cook. Darwin, in his 1859 publication, focused primarily on the changes that can occur within a species, and less on how species may divide into two. It is almost universally accepted that Darwin's book did not directly address its title. Darwin instead saw speciation as occurring by species entering new ecological niches.

Darwin's views

Controversy exists as to whether Charles Darwin recognized a true geographical-based model of speciation in his publication On the Origin of Species. In chapter 11, "Geographical Distribution", Darwin discusses geographic barriers to migration, stating for example that "barriers of any kind, or obstacles to free migration, are related in a close and important manner to the differences between the productions of various regions [of the world]". F. J. Sulloway contends that Darwin's position on speciation was "misleading" at the least and may have later misinformed Wagner and David Starr Jordan into believing that Darwin viewed sympatric speciation as the most important mode of speciation. Nevertheless, Darwin never fully accepted Wagner's concept of geographical speciation.

The evolutionary biologist James Mallet maintains that the mantra repeated concerning Darwin's Origin of Species book having never actually discussed speciation is specious. The claim began with Thomas Henry Huxley and George Romanes (contemporaries of Darwin's), who declared that Darwin failed to explain the origins of inviability and sterility in hybrids. Similar claims were promulgated by the mutationist school of thought during the late 20th century, and even after the modern evolutionary synthesis by Richard Goldschmidt. Another strong proponent of this view about Darwin came from Mayr. Mayr maintained that Darwin was unable to address the problem of speciation, as he did not define species using the biological species concept. However, Mayr's view has not been entirely accepted, as Darwin's transmutation notebooks contained writings concerning the role of isolation in the splitting of species. Furthermore, Many of Darwin's ideas on speciation largely match the modern theories of both adaptive radiation and ecological speciation.

Biogeographic influence

The German traveller Moritz Wagner (1813–1887)

Recognition of geographic factors involved in species populations was present even before Darwin, with many naturalists aware of the role of isolation in species relationships. In 1833, C. L. Gloger published The Variation of Birds Under the Influence of Climate in which he described geographic variations, but did not recognize that geographic isolation was an indicator of past speciation events. Another naturalist in 1856, Wollaston, studied island beetles in comparison to mainland species. He saw isolation as key to their differentiation. However, he did not recognize that the pattern was due to speciation. One naturalist, Leopold von Buch (1825) did recognize the geographic patterns and explicitly stated that geographic isolation may lead to species separating into new species. Mayr suggests that Von Buch was likely the first naturalist to truly suggest geographic speciation. Other naturalists, such as Henry Walter Bates (1863), recognized and accepted the patterns as evidence of speciation, but in Bate's case, did not propose a coherent model.

In 1868, Moritz Wagner was the first to propose the concept of geographic speciation in which he used the term Separationstheorie. Edward Bagnall Poulton, the evolutionary biologist and a strong proponent of the importance of natural selection, highlighted the role of geographic isolation in promoting speciation, in the process coining the term "sympatric speciation" in 1904.

Wagner and other naturalists who studied the geographic distributions of animals, such as Karl Jordan and David Starr Jordan, noticed that closely related species were often geographically isolated from one another (allopatrically distributed) which lead to the advocation of the importance of geographic isolation in the origin of species. Karl Jordan is thought to have recognized the unification of mutation and isolation in the origin of new species — in stark contrast to the prevailing views at the time. David Starr Jordan reiterated Wagner's proposal in 1905, providing a wealth of evidence from nature to support the theory, and asserting that geographic isolation is obvious but had been unfortunately ignored by most geneticists and experimental evolutionary biologists at the time. Joel Asaph Allen suggested the observed pattern of geographic separation of closely related species be called "Jordan's Law" (or Wagner's Law). Despite the contentions, most taxonomists did accept the geographic model of speciation.

Many of the early terms used to describe speciation were outlined by Ernst Mayr. He was the first to encapsulate the then contemporary literature in his 1942 publication Systematics and the Origin of Species, from the Viewpoint of a Zoologist and in his subsequent 1963 publication Animal Species and Evolution. Like Jordan's works, they relied on direct observations of nature, documenting the occurrence of geographic speciation. He described the three modes: geographic, semi-geographic, and non-geographic; which today, are referred to as allopatric, parapatric, and sympatric respectively. Mayr's 1942 publication, influenced heavily by the ideas of Karl Jordan and Poulton, was regarded as the authoritative review of speciation for over 20 years—and is still valuable today.

A major focus of Mayr's works was on the importance of geography in facilitating speciation; with islands often acting as a central theme to many of the speciation concepts put forth. One of which was the concept of peripatric speciation, a variant of allopatric speciation (he has since distinguished the two modes by referring to them as peripatric and dichopatric). This concept arose by an interpretation of Wagner's Separationstheorie as a form of founder effect speciation that focused on small geographically isolated species. This model was later expanded and modified to incorporate sexual selection by Kenneth Y. Kaneshiro in 1976 and 1980.

Modern evolutionary synthesis

Many geneticists at the time did little to bridge the gap between the genetics of natural selection and the origin of reproductive barriers between species. Ronald Fisher proposed a model of speciation in his 1930 publication The Genetical Theory of Natural Selection, where he described disruptive selection acting on sympatric or parapatric populations — with reproductive isolation completed by reinforcement. Other geneticists such as J. B. S. Haldane did not even recognize that species were real, while Sewall Wright ignored the topic, despite accepting allopatric speciation.

The primary contributors to the incorporation of speciation into modern evolutionary synthesis were Ernst Mayr and Theodosius Dobzhansky. Dobzhansky, a geneticist, published Genetics and the Origin of Species in 1937, in which he formulated the genetic framework for how speciation could occur. He recognized that speciation was an unsolved problem in biology at the time, rejecting Darwin’s position that new species arose by occupation of new niches — contending that reproductive isolation was instead based on barriers to gene flow. Subsequently, Mayr conducted extensive work on the geography of species, emphasizing the importance of geographic separation and isolation, in which he filled Dobzhansky‘s gaps concerning the origin of biodiversity (in his 1942 book). Both of their works gave rise, not without controversy, to the modern understanding of speciation; stimulating a wealth of research on the topic. Furthermore, this extended to plants as well as animals with G. Ledyard Stebbins’s book, Variation and Evolution in Plants and the much later, 1981 book, Plant Speciation by Verne Grant. 

Ernst Mayr, an influential evolutionary biologist

 

In 1947, "a consensus had been achieved among geneticists, paleontologists and systematists and that evolutionary biology as an independent biological discipline had been established" during a Princeton University conference. This 20th century synthesis incorporated speciation. Since then, the ideas have been consistently and repeatedly confirmed.

Contemporary work

After the synthesis, speciation research continued largely within natural history and biogeography — with much less emphasis on genetics. The study of speciation has seen its largest increase since the 1980s with an influx of publications and a host of new terms, methods, concepts, and theories. This "third phase" of work — as Jerry A. Coyne and H. Allen Orr put it — has led to a growing complexity of the language used to describe the many processes of speciation. The research and literature on speciation have become, "enormous, scattered, and increasingly technical".

From the 1980s, new research tools increased the robustness of research, assisted by new methods, theoretical frameworks, models, and approaches. Coyne and Orr discuss the modern, post-1980s developments centered around five major themes:

1. genetics (also a primary factor in the Modern Synthesis),
2. molecular biology and analysis (namely, phylogenetics and systematics);
3. comparative analysis;
4. mathematical modeling and computer simulations; and
5. the role of ecology.

Ecologists became aware that the ecological factors behind speciation were under-represented. This saw the growth in research concerning ecology's role in facilitating speciation — rightly designated ecological speciation. This focus on ecology generated a host of new terms relating to the barriers to reproduction (e.g. allochronic speciation, in which gene flow is reduced or removed by timing of breeding periods; or habitat isolation, in which species occupy different habitats within the same area). Sympatric speciation, regarded by Mayr as unlikely, has become widely accepted. Research on the influence of natural selection on speciation, including the process of reinforcement, has grown.

Researchers have long debated the roles of sexual selection, natural selection, and genetic drift in speciation. Darwin extensively discussed sexual selection, with his work greatly expanded on by Ronald Fisher; however, it was not until 1983 that the biologist Mary Jane West-Eberhard recognized the importance of sexual selection in speciation. Natural selection plays a role in that any selection towards reproductive isolation can result in speciation — whether indirectly or directly. Genetic drift has been widely researched from the 1950s onwards, especially with peak-shift models of speciation by genetic drift. Mayr championed founder effects, in which isolated individuals, like those found on islands near a mainland, experience a strong population bottleneck, as they contain only a small sample of the genetic variation in the main population. Later, other biologists such as Hampton L. Carson, Alan Templeton, Sergey Gavrilets, and Alan Hastings developed related models of speciation by genetic drift, noting that islands were inhabited mostly by endemic species. Selection's role in speciation is widely supported, whereas founder effect speciation is not, having been subject to a number of criticisms.

Classification debate

Speciation represented as a continuum of gene flow where ${\displaystyle m}$ equals the rate of gene exchange. The three primary geographic modes of speciation (allopatric, parapatric, and sympatric) can exist within this continuum, as well as other non-geographic modes.

 

Throughout the history of research concerning speciation, classification and delineation of modes and processes have been debated. Julian Huxley divided speciation into three separate modes: geographical speciation, genetic speciation, and ecological speciation. Sewall Wright proposed ten different, varying modes. Ernst Mayr championed the importance of physical, geographic separation of species populations, maintaining it to be of major importance to speciation. He originally proposed the three primary modes known today: geographic, semi-geographic, non-geographic; corresponding to allopatric, parapatric, and sympatric respectively. 

The phrase "modes of speciation" is imprecisely defined, most often indicating speciation occurring as a result of a species geographic distribution. More succinctly, the modern classification of speciation is often described as occurring on a gene flow continuum (i.e., allopatry at ${\displaystyle m=0}$ and sympatry at ${\displaystyle m=0.5}$) This gene flow concept views speciation as based on the exchange of genes between populations instead of seeing a purely geographic setting as necessarily relevant. Despite this, concepts of biogeographic modes can be translated into models of gene flow (such as that in the image at left); however, this translation has led to some confusion of language in the scientific literature.

Comparisons of the three classic geographic modes of speciation: allopatric, parapatric and sympatric; with peripatric speciation included as a special case of allopatric speciation.

 

As research has expanded over the decades, the geographic scheme has been challenged. The traditional classification is considered by some researchers to be obsolete, while others argue for its merits. Proponents of non-geographic schemes often justify non-geographic classifications, not by rejection of the importance of reproductive isolation (or even the processes themselves), but instead by the fact that it simplifies the complexity of speciation. One major critique of the geographic framework is that it arbitrarily separates a biological continuum into discontinuous groups. Another criticism rests with the fact that, when speciation is viewed as a continuum of gene flow, parapatric speciation becomes unreasonably represented by the entire continuum—with allopatric and sympatric existing in the extremes. Coyne and Orr argue that the geographic classification scheme is valuable in that biogeography controls the strength of the evolutionary forces at play, as gene flow and geography are clearly linked. James Mallet and colleagues contend that the sympatric vs. allopatric dichotomy is valuable to determine the degree in which natural selection acts on speciation. Kirkpatrick and Ravigné categorize speciation in terms of its genetic basis or by the forces driving reproductive isolation. Here, the geographic modes of speciation are classified as types of assortive mating. Fitzpatrick and colleagues believe that the biogeographic scheme "is a distraction that could be positively misleading if the real goal is to understand the influence of natural selection on divergence." They maintain that, to fully understand speciation, "the spatial, ecological, and genetic factors" involved in divergence must be explored. Sara Via recognizes the importance of geography in speciation but suggests that classification under this scheme be abandoned.

History of modes and mechanisms

Sympatric speciation

Sympatric speciation, from its beginnings with Darwin (who did not coin the term), has been a contentious issue. Mayr, along with many other evolutionary biologists, interpreted Darwins's view of speciation and the origin of biodiversity as arising by species entering new ecological niches—a form of sympatric speciation. Before Mayr, sympatric speciation was regarded as the primary mode of speciation. In 1963, Mayr provided a strong criticism, citing various flaws in the theory. After that, sympatric speciation fell out of favor with biologists and has only recently seen a resurgence in interest. Some biologists, such as James Mallet, believe that Darwin's view on speciation was misunderstood and misconstrued by Mayr. Today, sympatric speciation is supported by evidence from laboratory experiments and observations from nature.

Hybrid speciation

For most of the history of speciation, hybridization (polyploidy) has been a contentious issue, as botanists and zoologists have traditionally viewed hybridization's role in speciation differently. Carl Linnaeus was the earliest to suggest hybridization in 1760, Øjvind Winge was the first to confirm allopolyploidy in 1917, and a later experiment conducted by Clausen and Goodspeed in 1925 confirmed the findings. Today it is widely recognized as a common mechanism of speciation.

Historically, zoologists considered hybridization to be a rare phenomenon, while botanists found it to be commonplace in plant species. The botanists G. Ledyard Stebbins and Verne Grant were two of the well known botanists who championed the idea of hybrid speciation during the 1950s to the 1980s. Hybrid speciation, also called polyploid speciation (or polyploidy) is speciation that results by an increase in the number of sets of chromosomes. It is effectively a form of sympatric speciation that happens instantly. Grant coined the term recombinational speciation in 1981; a special form of hybrid speciation where a new species results from hybridization and is itself, reproductively isolated from both its parents. Recently, biologists have increasingly recognized that hybrid speciation can occur in animals as well.

Reinforcement

The young naturalist Alfred Russel Wallace in 1862

The concept of speciation by reinforcement has a complex history, with its popularity among scholars changing significantly over time. The theory of reinforcement experienced three phases of historical development:

1. plausibility based on unfit hybrids;
2. implausibility based on the finding that hybrids may have some fitness;
3. plausibility based on empirical studies and biologically complex and realistic models.

It was originally proposed by Alfred Russel Wallace in 1889, termed the Wallace effect—a term rarely used by scientists today. Wallace's hypothesis differed from the modern conception in that it focused on post-zygotic isolation, strengthened by group selection. Dobzhansky was the first to provide a thorough, modern description of the process in 1937, though the actual term itself was not coined until 1955 by W. Frank Blair.

In 1930, Ronald Fisher laid out the first genetic description of the process of reinforcement in The Genetical Theory of Natural Selection, and in 1965 and 1970 the first computer simulations were run to test for its plausibility. Later, population genetic and quantitative genetic studies were conducted showing that completely unfit hybrids lead to an increase in pre-zygotic isolation. After Dobzhansky's idea rose to the forefront of speciation research, it garnered significant support—with Dobzhansky suggesting that it illustrated the final step in speciation (e.g. after an allopatric population comes into secondary contact). In the 1980s, many evolutionary biologists began to doubt the plausibility of the idea, based not on empirical evidence, but largely on the growth of theory that deemed it an unlikely mechanism of reproductive isolation. A number of theoretical objections arose at the time. Since the early 1990s, reinforcement has seen a revival in popularity, with perceptions by evolutionary biologists accepting its plausibility—due primarily from a sudden increase in data, empirical evidence from laboratory studies and nature, complex computer simulations, and theoretical work.

The scientific language concerning reinforcement has also differed over time, with different researchers applying various definitions to the term. First used to describe the observed mating call differences in Gastrophryne frogs within a secondary contact hybrid zone, reinforcement has also been used to describe geographically separated populations that experience secondary contact. Roger Butlin demarcated incomplete post-zygotic isolation from complete isolation, referring to incomplete isolation as reinforcement and completely isolated populations as experiencing reproductive character displacement. Daniel J. Howard considered reproductive character displacement to represent either assortive mating or the divergence of traits for mate recognition (specifically between sympatric populations). Under this definition, it includes pre-zygotic divergence and complete post-zygotic isolation. Maria R. Servedio and Mohamed Noor consider any detected increase in pre-zygotic isolation as reinforcement, as long as it is a response to selection against mating between two different species. Coyne and Orr contend that, "true reinforcement is restricted to cases in which isolation is enhanced between taxa that can still exchange genes".

History of molecular evolution

From Wikipedia, the free encyclopedia

The history of molecular evolution starts in the early 20th century with "comparative biochemistry", but the field of molecular evolution came into its own in the 1960s and 1970s, following the rise of molecular biology. The advent of protein sequencing allowed molecular biologists to create phylogenies based on sequence comparison, and to use the differences between homologous sequences as a molecular clock to estimate the time since the last common ancestor. In the late 1960s, the neutral theory of molecular evolution provided a theoretical basis for the molecular clock, though both the clock and the neutral theory were controversial, since most evolutionary biologists held strongly to panselectionism, with natural selection as the only important cause of evolutionary change. After the 1970s, nucleic acid sequencing allowed molecular evolution to reach beyond proteins to highly conserved ribosomal RNA sequences, the foundation of a new conception of the early history of life.

Early history

Before the rise of molecular biology in the 1950s and 1960s, a small number of biologists had explored the possibilities of using biochemical differences between species to study evolution. Alfred Sturtevant predicted the existence of chromosomal inversions in 1921 and with Dobzhansky constructed one of the first molecular phylogenies on 17 Drosophila Pseudo-obscura strains from the accumulation of chromosomal inversions observed from the hybridization of polyten chromosomes. Ernest Baldwin worked extensively on comparative biochemistry beginning in the 1930s, and Marcel Florkin pioneered techniques for constructing phylogenies based on molecular and biochemical characters in the 1940s. However, it was not until the 1950s that biologists developed techniques for producing biochemical data for the quantitative study of molecular evolution.

The first molecular systematics research was based on immunological assays and protein "fingerprinting" methods. Alan Boyden—building on immunological methods of George Nuttall—developed new techniques beginning in 1954, and in the early 1960s Curtis Williams and Morris Goodman used immunological comparisons to study primate phylogeny. Others, such as Linus Pauling and his students, applied newly developed combinations of electrophoresis and paper chromatography to proteins subject to partial digestion by digestive enzymes to create unique two-dimensional patterns, allowing fine-grained comparisons of homologous proteins.

Beginning in the 1950s, a few naturalists also experimented with molecular approaches—notably Ernst Mayr and Charles Sibley. While Mayr quickly soured on paper chromatography, Sibley successfully applied electrophoresis to egg-white proteins to sort out problems in bird taxonomy, soon supplemented that with DNA hybridization techniques—the beginning of a long career built on molecular systematics.

While such early biochemical techniques found grudging acceptance in the biology community, for the most part they did not impact the main theoretical problems of evolution and population genetics. This would change as molecular biology shed more light on the physical and chemical nature of genes.

Genetic load, the classical/balance controversy, and the measurement of heterozygosity

At the time that molecular biology was coming into its own in the 1950s, there was a long-running debate—the classical/balance controversy—over the causes of heterosis, the increase in fitness observed when inbred lines are outcrossed. In 1950, James F. Crow offered two different explanations (later dubbed the classical and balance positions) based on the paradox first articulated by J. B. S. Haldane in 1937: the effect of deleterious mutations on the average fitness of a population depends only on the rate of mutations (not the degree of harm caused by each mutation) because more-harmful mutations are eliminated more quickly by natural selection, while less-harmful mutations remain in the population longer. H. J. Muller dubbed this "genetic load".

Muller, motivated by his concern about the effects of radiation on human populations, argued that heterosis is primarily the result of deleterious homozygous recessive alleles, the effects of which are masked when separate lines are crossed—this was the dominance hypothesis, part of what Dobzhansky labeled the classical position. Thus, ionizing radiation and the resulting mutations produce considerable genetic load even if death or disease does not occur in the exposed generation, and in the absence of mutation natural selection will gradually increase the level of homozygosity. Bruce Wallace, working with J. C. King, used the overdominance hypothesis to develop the balance position, which left a larger place for overdominance (where the heterozygous state of a gene is more fit than the homozygous states). In that case, heterosis is simply the result of the increased expression of heterozygote advantage. If overdominant loci are common, then a high level of heterozygosity would result from natural selection, and mutation-inducing radiation may in fact facilitate an increase in fitness due to overdominance. (This was also the view of Dobzhansky.)

Debate continued through 1950s, gradually becoming a central focus of population genetics. A 1958 study of Drosophila by Wallace suggested that radiation-induced mutations increased the viability of previously homozygous flies, providing evidence for heterozygote advantage and the balance position; Wallace estimated that 50% of loci in natural Drosophila populations were heterozygous. Motoo Kimura's subsequent mathematical analyses reinforced what Crow had suggested in 1950: that even if overdominant loci are rare, they could be responsible for a disproportionate amount of genetic variability. Accordingly, Kimura and his mentor Crow came down on the side of the classical position. Further collaboration between Crow and Kimura led to the infinite alleles model, which could be used to calculate the number of different alleles expected in a population, based on population size, mutation rate, and whether the mutant alleles were neutral, overdominant, or deleterious. Thus, the infinite alleles model offered a potential way to decide between the classical and balance positions, if accurate values for the level of heterozygosity could be found.

By the mid-1960s, the techniques of biochemistry and molecular biology—in particular protein electrophoresis—provided a way to measure the level of heterozygosity in natural populations: a possible means to resolve the classical/balance controversy. In 1963, Jack L. Hubby published an electrophoresis study of protein variation in Drosophila; soon after, Hubby began collaborating with Richard Lewontin to apply Hubby's method to the classical/balance controversy by measuring the proportion of heterozygous loci in natural populations. Their two landmark papers, published in 1966, established a significant level of heterozygosity for Drosophila (12%, on average). However, these findings proved difficult to interpret. Most population geneticists (including Hubby and Lewontin) rejected the possibility of widespread neutral mutations; explanations that did not involve selection were anathema to mainstream evolutionary biology. Hubby and Lewontin also ruled out heterozygote advantage as the main cause because of the segregation load it would entail, though critics argued that the findings actually fit well with over dominance hypothesis.

Protein sequences and the molecular clock

While evolutionary biologists were tentatively branching out into molecular biology, molecular biologists were rapidly turning their attention toward evolution. 

After developing the fundamentals of protein sequencing with insulin between 1951 and 1955, Frederick Sanger and his colleagues had published a limited interspecies comparison of the insulin sequence in 1956. Francis Crick, Charles Sibley and others recognized the potential for using biological sequences to construct phylogenies, though few such sequences were yet available. By the early 1960s, techniques for protein sequencing had advanced to the point that direct comparison of homologous amino acid sequences was feasible. In 1961, Emanuel Margoliash and his collaborators completed the sequence for horse cytochrome c (a longer and more widely distributed protein than insulin), followed in short order by a number of other species.

In 1962, Linus Pauling and Emile Zuckerkandl proposed using the number of differences between homologous protein sequences to estimate the time since divergence, an idea Zuckerkandl had conceived around 1960 or 1961. This began with Pauling's long-time research focus, hemoglobin, which was being sequenced by Walter Schroeder; the sequences not only supported the accepted vertebrate phylogeny, but also the hypothesis (first proposed in 1957) that the different globin chains within a single organism could also be traced to a common ancestral protein. Between 1962 and 1965, Pauling and Zuckerkandl refined and elaborated this idea, which they dubbed the molecular clock, and Emil L. Smith and Emanuel Margoliash expanded the analysis to cytochrome c. Early molecular clock calculations agreed fairly well with established divergence times based on paleontological evidence. However, the essential idea of the molecular clock—that individual proteins evolve at a regular rate independent of a species' morphological evolution—was extremely provocative (as Pauling and Zuckerkandl intended it to be).

The "molecular wars"

From the early 1960s, molecular biology was increasingly seen as a threat to the traditional core of evolutionary biology. Established evolutionary biologists—particularly Ernst Mayr, Theodosius Dobzhansky and G. G. Simpson, three of the founders of the modern evolutionary synthesis of the 1930s and 1940s—were extremely skeptical of molecular approaches, especially when it came to the connection (or lack thereof) to natural selection. Molecular evolution in general—and the molecular clock in particular—offered little basis for exploring evolutionary causation. According to the molecular clock hypothesis, proteins evolved essentially independently of the environmentally determined forces of selection; this was sharply at odds with the panselectionism prevalent at the time. Moreover, Pauling, Zuckerkandl, and other molecular biologists were increasingly bold in asserting the significance of "informational macromolecules" (DNA, RNA and proteins) for all biological processes, including evolution. The struggle between evolutionary biologists and molecular biologists—with each group holding up their discipline as the center of biology as a whole—was later dubbed the "molecular wars" by Edward O. Wilson, who experienced firsthand the domination of his biology department by young molecular biologists in the late 1950s and the 1960s.

In 1961, Mayr began arguing for a clear distinction between functional biology (which considered proximate causes and asked "how" questions) and evolutionary biology (which considered ultimate causes and asked "why" questions) He argued that both disciplines and individual scientists could be classified on either the functional or evolutionary side, and that the two approaches to biology were complementary. Mayr, Dobzhansky, Simpson and others used this distinction to argue for the continued relevance of organism biology, which was rapidly losing ground to molecular biology and related disciplines in the competition for funding and university support. It was in that context that Dobzhansky first published his famous statement, "nothing in biology makes sense except in the light of evolution", in a 1964 paper affirming the importance of organismal biology in the face of the molecular threat; Dobzhansky characterized the molecular disciplines as "Cartesian" (reductionist) and organismal disciplines as "Darwinian".

Mayr and Simpson attended many of the early conferences where molecular evolution was discussed, critiquing what they saw as the overly simplistic approaches of the molecular clock. The molecular clock, based on uniform rates of genetic change driven by random mutations and drift, seemed incompatible with the varying rates of evolution and environmentally-driven adaptive processes (such as adaptive radiation) that were among the key developments of the evolutionary synthesis. At the 1962 Wenner-Gren conference, the 1964 Colloquium on the Evolution of Blood Proteins in Bruges, Belgium, and the 1964 Conference on Evolving Genes and Proteins at Rutgers University, they engaged directly with the molecular biologists and biochemists, hoping to maintain the central place of Darwinian explanations in evolution as its study spread to new fields.

Gene-centered view of evolution

Though not directly related to molecular evolution, the mid-1960s also saw the rise of the gene-centered view of evolution, spurred by George C. Williams's Adaptation and Natural Selection (1966). Debate over units of selection, particularly the controversy over group selection, led to increased focus on individual genes (rather than whole organisms or populations) as the theoretical basis for evolution. However, the increased focus on genes did not mean a focus on molecular evolution; in fact, the adaptationism promoted by Williams and other evolutionary theories further marginalized the apparently non-adaptive changes studied by molecular evolutionists.

The neutral theory of molecular evolution

The intellectual threat of molecular evolution became more explicit in 1968, when Motoo Kimura introduced the neutral theory of molecular evolution. Based on the available molecular clock studies (of hemoglobin from a wide variety of mammals, cytochrome c from mammals and birds, and triosephosphate dehydrogenase from rabbits and cows), Kimura (assisted by Tomoko Ohta) calculated an average rate of DNA substitution of one base pair change per 300 base pairs (encoding 100 amino acids) per 28 million years. For mammal genomes, this indicated a substitution rate of one every 1.8 years, which would produce an unsustainably high substitution load unless the preponderance of substitutions was selectively neutral. Kimura argued that neutral mutations occur very frequently, a conclusion compatible with the results of the electrophoretic studies of protein heterozygosity. Kimura also applied his earlier mathematical work on genetic drift to explain how neutral mutations could come to fixation, even in the absence of natural selection; he soon convinced James F. Crow of the potential power of neutral alleles and genetic drift as well.

Kimura's theory—described only briefly in a letter to Nature—was followed shortly after with a more substantial analysis by Jack L. King and Thomas H. Jukes—who titled their first paper on the subject "non-Darwinian evolution". Though King and Jukes produced much lower estimates of substitution rates and the resulting genetic load in the case of non-neutral changes, they agreed that neutral mutations driven by genetic drift were both real and significant. The fairly constant rates of evolution observed for individual proteins was not easily explained without invoking neutral substitutions (though G. G. Simpson and Emil Smith had tried). Jukes and King also found a strong correlation between the frequency of amino acids and the number of different codons encoding each amino acid. This pointed to substitutions in protein sequences as being largely the product of random genetic drift.

King and Jukes' paper, especially with the provocative title, was seen as a direct challenge to mainstream neo-Darwinism, and it brought molecular evolution and the neutral theory to the center of evolutionary biology. It provided a mechanism for the molecular clock and a theoretical basis for exploring deeper issues of molecular evolution, such as the relationship between rate of evolution and functional importance. The rise of the neutral theory marked synthesis of evolutionary biology and molecular biology—though an incomplete one.

With their work on firmer theoretical footing, in 1971 Emile Zuckerkandl and other molecular evolutionists founded the Journal of Molecular Evolution.

The neutralist-selectionist debate and near-neutrality

The critical responses to the neutral theory that soon appeared marked the beginning of the neutralist-selectionist debate. In short, selectionists viewed natural selection as the primary or only cause of evolution, even at the molecular level, while neutralists held that neutral mutations were widespread and that genetic drift was a crucial factor in the evolution of proteins. Kimura became the most prominent defender of the neutral theory—which would be his main focus for the rest of his career. With Ohta, he refocused his arguments on the rate at which drift could fix new mutations in finite populations, the significance of constant protein evolution rates, and the functional constraints on protein evolution that biochemists and molecular biologists had described. Though Kimura had initially developed the neutral theory partly as an outgrowth of the classical position within the classical/balance controversy (predicting high genetic load as a consequence of non-neutral mutations), he gradually de-emphasized his original argument that segregational load would be impossibly high without neutral mutations (which many selectionists, and even fellow neutralists King and Jukes, rejected).

From the 1970s through the early 1980s, both selectionists and neutralists could explain the observed high levels of heterozygosity in natural populations, by assuming different values for unknown parameters. Early in the debate, Kimura's student Tomoko Ohta focused on the interaction between natural selection and genetic drift, which was significant for mutations that were not strictly neutral, but nearly so. In such cases, selection would compete with drift: most slightly deleterious mutations would be eliminated by natural selection or chance; some would move to fixation through drift. The behavior of this type of mutation, described by an equation that combined the mathematics of the neutral theory with classical models, became the basis of Ohta's nearly neutral theory of molecular evolution.

In 1973, Ohta published a short letter in Nature suggesting that a wide variety of molecular evidence supported the theory that most mutation events at the molecular level are slightly deleterious rather than strictly neutral. Molecular evolutionists were finding that while rates of protein evolution (consistent with the molecular clock) were fairly independent of generation time, rates of noncoding DNA divergence were inversely proportional to generation time. Noting that population size is generally inversely proportional to generation time, Tomoko Ohta proposed that most amino acid substitutions are slightly deleterious while noncoding DNA substitutions are more neutral. In this case, the faster rate of neutral evolution in proteins expected in small populations (due to genetic drift) is offset by longer generation times (and vice versa), but in large populations with short generation times, noncoding DNA evolves faster while protein evolution is retarded by selection (which is more significant than drift for large populations).

Between then and the early 1990s, many studies of molecular evolution used a "shift model" in which the negative effect on the fitness of a population due to deleterious mutations shifts back to an original value when a mutation reaches fixation. In the early 1990s, Ohta developed a "fixed model" that included both beneficial and deleterious mutations, so that no artificial "shift" of overall population fitness was necessary. According to Ohta, however, the nearly neutral theory largely fell out of favor in the late 1980s, because the mathematically simpler neutral theory for the widespread molecular systematics research that flourished after the advent of rapid DNA sequencing. As more detailed systematics studies started to compare the evolution of genome regions subject to strong selection versus weaker selection in the 1990s, the nearly neutral theory and the interaction between selection and drift have once again become an important focus of research.

Microbial phylogeny

While early work in molecular evolution focused on readily sequenced proteins and relatively recent evolutionary history, by the late 1960s some molecular biologists were pushing further toward the base of the tree of life by studying highly conserved nucleic acid sequences. Carl Woese, a molecular biologist whose earlier work was on the genetic code and its origin, began using small subunit ribosomal RNA to reclassify bacteria by genetic (rather than morphological) similarity. Work proceeded slowly at first, but accelerated as new sequencing methods were developed in the 1970s and 1980s. By 1977, Woese and George Fox announced that some bacteria, such as methanogens, lacked the rRNA units that Woese's phylogenetic studies were based on; they argued that these organisms were actually distinct enough from conventional bacteria and the so-called higher organisms to form their own kingdom, which they called archaebacteria. Though controversial at first (and challenged again in the late 1990s), Woese's work became the basis of the modern three-domain system of Archaea, Bacteria, and Eukarya (replacing the five-domain system that had emerged in the 1960s).

Work on microbial phylogeny also brought molecular evolution closer to cell biology and origin of life research. The differences between archaea pointed to the importance of RNA in the early history of life. In his work with the genetic code, Woese had suggested RNA-based life had preceded the current forms of DNA-based life, as had several others before him—an idea that Walter Gilbert would later call the "RNA world". In many cases, genomics research in the 1990s produced phylogenies contradicting the rRNA-based results, leading to the recognition of widespread lateral gene transfer across distinct taxa. Combined with the probable endosymbiotic origin of organelle-filled eukarya, this pointed to a far more complex picture of the origin and early history of life, one which might not be describable in the traditional terms of common ancestry.

A New Approach to Understanding How Machines Think

Neural networks are famously incomprehensible — a computer can come up with a good answer, but not be able to explain what led to the conclusion. Been Kim is developing a “translator for humans” so that we can understand when artificial intelligence breaks down.

 

Been Kim, a research scientist at Google Brain, is developing a way to ask a machine learning system how much a specific, high-level concept went into its decision-making process.

Rachel Bujalski for Quanta Magazine

John Pavlus

Contributing Writer

---

January 10, 2019

Original link:  https://www.quantamagazine.org/been-kim-is-building-a-translator-for-artificial-intelligence-20190110/

---

 

If a doctor told that you needed surgery, you would want to know why — and you’d expect the explanation to make sense to you, even if you’d never gone to medical school. Been Kim, a research scientist at Google Brain, believes that we should expect nothing less from artificial intelligence. As a specialist in “interpretable” machine learning, she wants to build AI software that can explain itself to anyone.

Since its ascendance roughly a decade ago, the neural-network technology behind artificial intelligence has transformed everything from email to drug discovery with its increasingly powerful ability to learn from and identify patterns in data. But that power has come with an uncanny caveat: The very complexity that lets modern deep-learning networks successfully teach themselves how to drive cars and spot insurance fraud also makes their inner workings nearly impossible to make sense of, even by AI experts. If a neural network is trained to identify patients at risk for conditions like liver cancer and schizophrenia — as a system called “Deep Patient” was in 2015, at Mount Sinai Hospital in New York — there’s no way to discern exactly which features in the data the network is paying attention to. That “knowledge” is smeared across many layers of artificial neurons, each with hundreds or thousands of connections.

As ever more industries attempt to automate or enhance their decision-making with AI, this so-called black box problem seems less like a technological quirk than a fundamental flaw. DARPA’s “XAI” project (for “explainable AI”) is actively researching the problem, and interpretability has moved from the fringes of machine-learning research to its center. “AI is in this critical moment where humankind is trying to decide whether this technology is good for us or not,” Kim says. “If we don’t solve this problem of interpretability, I don’t think we’re going to move forward with this technology. We might just drop it.”

Kim and her colleagues at Google Brain recently developed a system called “Testing with Concept Activation Vectors” (TCAV), which she describes as a “translator for humans” that allows a user to ask a black box AI how much a specific, high-level concept has played into its reasoning. For example, if a machine-learning system has been trained to identify zebras in images, a person could use TCAV to determine how much weight the system gives to the concept of “stripes” when making a decision.

TCAV was originally tested on machine-learning models trained to recognize images, but it also works with models trained on text and certain kinds of data visualizations, like EEG waveforms. “It’s generic and simple — you can plug it into many different models,” Kim says.

Quanta Magazine spoke with Kim about what interpretability means, who it’s for, and why it matters. An edited and condensed version of the interview follows.

You’ve focused your career on “interpretability” for machine learning. But what does that term mean, exactly?

Rachel Bujalski for Quanta Magazine

There are two branches of interpretability. One branch is interpretability for science: If you consider a neural network as an object of study, then you can conduct scientific experiments to really understand the gory details about the model, how it reacts, and that sort of thing.

The second branch of interpretability, which I’ve been mostly focused on, is interpretability for responsible AI. You don’t have to understand every single thing about the model. But as long as you can understand just enough to safely use the tool, then that’s our goal.

But how can you have confidence in a system that you don’t fully understand the workings of?

I’ll give you an analogy. Let’s say I have a tree in my backyard that I want to cut down. I might have a chain saw to do the job. Now, I don’t fully understand how the chain saw works. But the manual says, “These are the things you need to be careful of, so as to not cut your finger.” So, given this manual, I’d much rather use the chainsaw than a handsaw, which is easier to understand, but would make me spend five hours cutting down the tree.

You understand what “cutting” is, even if you don’t exactly know everything about how the mechanism accomplishes that.

Yes. The goal of the second branch of interpretability is: Can we understand a tool enough so that we can safely use it? And we can create that understanding by confirming that useful human knowledge is reflected in the tool.

How does “reflecting human knowledge” make something like a black box AI more understandable?

Here’s another example. If a doctor is using a machine-learning model to make a cancer diagnosis, the doctor will want to know that the model isn’t picking up on some random correlation in the data that we don’t want to pick up. One way to make sure of that is to confirm that the machine-learning model is doing something that the doctor would have done. In other words, to show that the doctor’s own diagnostic knowledge is reflected in the model.

So if doctors were looking at a cell specimen to diagnose cancer, they might look for something called “fused glands” in the specimen. They might also consider the age of the patient, as well as whether the patient has had chemotherapy in the past. These are factors or concepts that the doctors trying to diagnose cancer would care about. If we can show that the machine-learning model is also paying attention to these factors, the model is more understandable, because it reflects the human knowledge of the doctors.

Google Brain’s Been Kim is building ways to let us interrogate the decisions made by machine learning systems.

Rachel Bujalski for Quanta Magazine

Is this what TCAV does — reveal which high-level concepts a machine-learning model is using to make its decisions?

Yes. Prior to this, interpretability methods only explained what neural networks were doing in terms of “input features.” What do I mean by that? If you have an image, every single pixel is an input feature. In fact, Yann LeCun [an early pioneer in deep learning and currently the director of AI research at Facebook] has said that he believes these models are already superinterpretable because you can look at every single node in the neural network and see numerical values for each of these input features. That’s fine for computers, but humans don’t think that way. I don’t tell you, “Oh, look at pixels 100 to 200, the RGB values are 0.2 and 0.3.” I say, “There’s a picture of a dog with really puffy hair.” That’s how humans communicate — with concepts.

How does TCAV perform this translation between input features and concepts?

Let’s return to the example of a doctor using a machine-learning model that has already been trained to classify images of cell specimens as potentially cancerous. You, as the doctor, may want to know how much the concept of “fused glands” mattered to the model in making positive predictions of cancer. First you collect some images — say, 20 — that have examples of fused glands. Now you plug those labeled examples into the model.

Then what TCAV does internally is called “sensitivity testing.” When we add in these labeled pictures of fused glands, how much does the probability of a positive prediction for cancer increase? You can output that as a number between zero and one. And that’s it. That’s your TCAV score. If the probability increased, it was an important concept to the model. If it didn’t, it’s not an important concept.

“Concept” is a fuzzy term. Are there any that won’t work with TCAV?

If you can’t express your concept using some subset of your [dataset’s] medium, then it won’t work. If your machine-learning model is trained on images, then the concept has to be visually expressible. Let’s say I want to visually express the concept of “love.” That’s really hard.

We also carefully validate the concept. We have a statistical testing procedure that rejects the concept vector if it has the same effect on the model as a random vector. If your concept doesn’t pass this test, then the TCAV will say, “I don’t know. This concept doesn’t look like something that was important to the model.”

Rachel Bujalski for Quanta Magazine

Is TCAV essentially about creating trust in AI, rather than a genuine understanding of it?

It is not — and I’ll explain why, because it’s a fine distinction to make.

We know from repeated studies in cognitive science and psychology that humans are very gullible. What that means is that it’s actually pretty easy to fool a person into trusting something. The goal of interpretability for machine learning is the opposite of this. It is to tell you if a system is not safe to use. It’s about revealing the truth. So “trust” isn’t the right word.

So the point of interpretability is to reveal potential flaws in an AI’s reasoning?

Yes, exactly.

How can it expose flaws?

You can use TCAV to ask a trained model about irrelevant concepts. To return to the example of doctors using AI to make cancer predictions, the doctors might suddenly think, “It looks like the machine is giving positive predictions of cancer for a lot of images that have a kind of bluish color artifact. We don’t think that factor should be taken into account.” So if they get a high TCAV score for “blue,” they’ve just identified a problem in their machine-learning model.

TCAV is designed to bolt on to existing AI systems that aren’t interpretable. Why not make the systems interpretable from the beginning, rather than black boxes?

There is a branch of interpretability research that focuses on building inherently interpretable models that reflect how humans reason. But my take is this: Right now you have AI models everywhere that are already built, and are already being used for important purposes, without having considered interpretability from the beginning. It’s just the truth. We have a lot of them at Google! You could say, “Interpretability is so useful, let me build you another model to replace the one you already have.” Well, good luck with that.

So then what do you do? We still need to get through this critical moment of deciding whether this technology is good for us or not. That’s why I work “post-training” interpretability methods. If you have a model that someone gave to you and that you can’t change, how do you go about generating explanations for its behavior so that you can use it safely? That’s what the TCAV work is about.

Rachel Bujalski for Quanta Magazine

TCAV lets humans ask an AI if certain concepts matter to it. But what if we don’t know what to ask — what if we want the AI system to explain itself?

We have work that we’re writing up right now that can automatically discover concepts for you. We call it DTCAV — discovery TCAV. But I actually think that having humans in the loop, and enabling the conversation between machines and humans, is the crux of interpretability.

A lot of times in high-stakes applications, domain experts already have a list of concepts that they care about. We see this repeat over and over again in our medical applications at Google Brain. They don’t want to be given a set of concepts — they want to tell the model the concepts that they are interested in. We worked with a doctor who treats diabetic retinopathy, which is an eye disease, and when we told her about TCAV, she was so excited because she already had many, many hypotheses about what this model might be doing, and now she can test those exact questions. It’s actually a huge plus, and a very user-centric way of doing collaborative machine learning.

You believe that without interpretability, humankind might just give up on AI technology. Given how powerful it is, do you really think that’s a realistic possibility?

Yes, I do. That’s what happened with expert systems. [In the 1980s] we established that they were cheaper than human operators to conduct certain tasks. But who is using expert systems now? Nobody. And after that we entered an AI winter.

Right now it doesn’t seem likely, because of all the hype and money in AI. But in the long run, I think that humankind might decide — perhaps out of fear, perhaps out of lack of evidence — that this technology is not for us. It’s possible.