A science advocate responds to the criticism that “evolution is just a theory,” explaining that, for scientists, a “theory” is a thoroughly tested, well established explanation. In some venues, this is a familiar trope. Yet, not only is this completely, utterly wrong, but anyone with a brain and an internet connection can verify that it’s wrong in minutes. Instead, scientists for 200 years have used “theory” for conjectures that are not well established, even ones whose truth is rejected, e.g., Lamarck’s theory. What is going on?
This post started out as a wonky rant about why a particular high-profile study of laboratory adaptation was mis-framed as though it were a validation of the mutational landscape model of Orr and Gillespie (see Orr, 2003), when in fact the specific innovations of that theory were either rejected, or not tested critically. As I continued ranting, I realized that there was quite a bit to say that is educational, and I contemplated that the reason for the original mis-framing is that this is an unfamiliar area, such that even the experts are confused— which means that there is a value to explaining things.
The crux of the matter is that the Gillespie-Orr “mutational landscape” model has some innovations, but also draws on other concepts and older work. We’ll start with these older foundations.
Origin-fixation dynamics in sequence space
First, the mutational landscape model draws on origin-fixation dynamics, proposed in 1969 by Kimura and Maruyama, and by King and Jukes (see McCandlish and Stoltzfus for an exhaustive review, or my blog on The surprising case of origin-fixation models).
In origin-fixation models, evolution is seen as a simple 2-step process of the introduction of a new allele, and its subsequent fixation (image). The rate of change is then characterized as a product of 2 factors, the rate of mutational origin (introduction) of new alleles of a particular type, and the probability that a new allele of that type will reach fixation. Under some general assumptions, this product is equal to the rate of an origin-fixation process when it reaches steady state.
Probably the most famous origin-fixation model is K = 4Nus, which uses 2s (Haldane, 1927) for the probability of fixation of a beneficial allele, and 2Nu (diploids) for the rate of mutational origin. Thus K = 4Nus is the expected rate of changes when we are considering types of beneficial alleles that arise by mutation at rate u, and have a selective advantage s. But we can adapt origin-fixation dynamics to other cases, including neutral and deleterious changes. If we were applying origin-fixation dynamics to meiotic bursts, or to phage bursts, in which the same mutational event gives rise immediately to multiple copies (prior to selection), we would use a probability of fixation that takes this multiplicity into account.
In passing, note that origin-fixation models appeared in 1969, and we haven’t always viewed evolution this way. The architects of the Modern Synthesis rejected this view— and if you don’t believe that, read The Shift to Mutationism is Documented in Our Language or The Buffet and the Sushi Conveyor. They saw evolution as more of an ongoing process of “shifting gene frequencies” in a super-abundant “gene pool” (image). Mutation merely supplies variation to the “gene pool”, which is kept full of variation. The contribution of mutation is trivial. The available variation is constantly mixed up by recombination, represented by the egg-beaters in the figure. When the environment changes, selection results in a new distribution of allele frequencies, and that’s “evolution”— shifting gene frequencies to a new optimum in response to a change in conditions.
This is probably too geeky to mention, but from a theoretical perspective, an origin-fixation model might mean 2 different things. It might be an aggregate rate of change across many sites, or the rate applied to a sequence of changes at a single locus. The mathematical derivation, the underlying assumptions, and the legitimate uses are different under these two conditions, as pointed out by McCandlish and Stoltzfus. The early models of King, et al were aggregate-rate models, while Gillespie, (1983) was the first to derive a sequential-fixations model.
Second, the mutational landscape model draws on Maynard Smith’s (1970) concept of evolution as discrete stepwise movement in a discrete “sequence space”. More specifically, it draws on the rarely articulated locality assumption by which we say that a step is limited to mutational “neighbors” of a sequence that differ by one simple mutation, rather than the entire universe of sequences. The justification for this assumption is that double-mutants, for instance, will arise in proportion to the square of the mutation rate, which is a very small number, so that we can ignore them. Instead, we can think of the evolutionary process as accessing only a local part of the universe of sequences, which shifts with each step it takes. In order for adaptive evolution to happen, there must be fitter genotypes in the neighborhood.
This is an important concept, and we ought to have a name for it. I call it the “evolutionary horizon”, because we can’t see beyond the horizon, and the horizon changes as we move. Note two things about this idea. The first is that this is a modeling assumption, not a feature of reality. Mutations that change 2 sites at once actually occur, and presumably they sometimes contribute to evolution. The second thing to note is that we could choose to define the horizon however we want, e.g., we could include single and double changes, but not triple ones. In practice, the mutational neighbors of a sequence are always defined as the sequences that differ by just 1 residue.
Putting these 2 pieces together, we can formulate a model of stepwise evolution with predictable dynamics.
Making this into a simulation of evolution is easy using the kind of number line shown at left. Each segment represents a possible mutation-fixation event from the starting sequence. For instance, we can change the “A” nucleotide that begins the sequence to “T”, “C” or “G”. The length of each segment is proportional to the origin-fixation probability for that change (where the probability is computed from the instantaneous rate). To pick the next step in evolution, we simply pick a random point on the number line. Then, we have to update the horizon— recompute the numberline with the new set of 1-mutant neighbors.
Where do we get the actual values for mutation and fixation? One way to do it is by drawing from some random distribution. I did this in a 2006 simulation study. I wasn’t doing anything special. It seemed very obvious at the time.
Surprisingly, researchers almost never measure actual mutation and selection values for relevant mutants. One exception is the important study by Rokyta, et al (2005), who repeatedly carried out 1-step adaptation using bacteriophage phiX174. The selection coefficients for each of the 11 beneficial changes observed in replicate experiments are shown in the rightmost column, with the mutants ranked from highest to lowest selection coefficient. Notice that the genotype that recurred most often (see the “Number” column) was not the alternative genotype with the highest fitness, but the 4th-most-fit alternative, which happened to be favored by a considerable bias in mutation. Rokyta, et al didn’t actually measure specific rates for each mutation, but simply estimated average rates for different classes of nucleotide mutation based on an evolutionary model.
Then Rokyta, et al. developed a model of origin-fixation dynamics using the estimated mutation rates, the measured selection coefficients, and a term for the probability of fixation customized to account for the way that phages grow. This model fit the data very well, as I’ll show in a figure below (panel C in the final figure).
The mutational landscape model
Given all of that, you might ask, what does the mutational landscape model do?
This is where the specific innovations of Orr and Gillespie come in. Just putting together origin-fixation dynamics and an evolutionary horizon doesn’t get us very far, because we can’t actually predict anything without filling in something concrete about the parameters, and that is a huge unknown. What if we don’t have that? Furthermore, although Rokyta, et al implicitly assumed a horizon in the sense that they ignored mutations too rare to appear in their study, they never tackled the question of how the horizon shifts with each step, because they only took one step. What if we want to do an extended adaptive walk? How will we know what is the distribution of fitnesses for the new set of neighbors, and how it relates to the previous distribution? In the simulation model that I mentioned previously, I used an abstract “NK” model of the fitness of a protein sequence that allowed me to specify the fitness of every possible sequence with a relatively small number of randomly assigned parameter values.
Gillespie and Orr were aiming to do something more clever than that. Theoreticians want to find ways to show something interesting just from applying basic principles, without having to use case-specific values for empirical parameters. After all, if we insert the numbers from one particular experimental system, then we are making a model just for that system.
The first innovation of Orr and Gillespie is to apply extreme value theory (EVT) in a way that offers predictions even if we haven’t measured the s values or assumed a specific model. If we assume that the current genotype is already highly adapted, this is tantamount to assuming it is in the tail end of a fitness distribution. EVT applies to the tail ends of distributions, even if we don’t know the specific shape of the distribution, which is very useful. Specifically, EVT tells us something about the relative sizes of s as we go from one rank to the next among the top-ranked genotypes: the distribution of fitness intervals is exponential. This leads to very specific predictions about the probability of jumping from rank r to some higher rank r’, including a fascinating invariance property where the expected upward jump in the fitness ranking is the same no matter where we are in the ranking. Namely, if the rank of the current genotype is j (i.e., j – 1 genotypes are better), we will jump to rank (j + 2)/4.
That’s fascinating, but what are we going to do with that information? I suspect the idea of a fitness rank previously appeared nowhere in the history of experimental biology, because rank isn’t a measurement one takes anywhere other than the racetrack. But remember that we would like a theory for an adaptive walk, not just 1-step adaptation. If we jump from j to k = (j + 2)/4, then from k to m = (k + 2)/4, and so on, we could develop a theory for the trajectory of fitness increases during an adaptive walk, and for the length of an adaptive walk— for how many steps we are likely to take before we can’t climb anymore.
The barrier to solving that theory is solving the evolutionary horizon problem. Every time we take a step, the horizon shifts— some points disappear from view, and others appear (Figure). We might be the 15th most-fit genotype, but at any step, only a subset of the 14 better genotypes will be accessible, and this subset changes with each step: this condition is precisely what Gillespie (1984) means by the phrase “the mutational landscape” (see Figure). In his 1983 paper, he just assumes that all the higher-fitness mutants are accessible throughout the walk. Gillespie’s 1984 paper entitled “Molecular Evolution over the Mutational Landscape” tackles the changing horizon explicitly. He doesn’t solve it analytically, but uses simulations. I won’t explain his approach, which I don’t fully understand. Analytical solutions appeared in later work by Jain and Seetharaman, 2011 (thanks to Dave McCandlish for pointing this out).
The third and fourth key innovations are to (3) ignore differences in u and (4) treat the chances of evolution as a linear function of s, based on Haldane’s 2s. In origin-fixation dynamics, the chance of a particular step is based on a simple product: rate of origin multiplied by probability of fixation. Orr’s model relates the chances of an evolutionary step entirely to the probability of fixation, assuming that u is uniform. Then, using 2s for the probability of fixation means that the chance of picking a mutant with fitness si is simply si / sum(s) where the sum is over all mutants (the factor of 2 cancels out because its the same for every mutant). Then, by applying EVT to the distribution of s, the model allows predictions based solely on the current rank.
A test of the mutational landscape model?
As noted earlier, this post started as a rant about a study that was mis-framed as though it were some kind of validation of Orr’s model. In fact, that study is Rokyta, et al., described above. Indeed, Rokyta, et al. tested Orr’s predictions, as shown in the left-most panel in the figure below. The predictions (grey bars) decrease smoothly because they are based, not on the actual measured fitness values shown above, but merely on the ranking. The starting genotype is ranked #10, and all the predictions of Orr’s model follow from that single fact, which is what makes the model cool!
If they did a test, what’s my objection? Yes, Rokyta, et al. turned the crank on Orr’s model and got predictions out, and did a goodness-of-fit test comparing observations to predictions. But, to test the mutational landscape model properly, you have to turn the crank at least 2 full turns to get the mojo working. Remember, what Gillespie means by “evolution over the mutational landscape” is literally the way evolution navigates the change in accessibility of higher-fitness genotypes due to the shifting evolutionary horizon. That doesn’t come into play in 1-step adaptation. You have to take at least 2 steps. Claiming to test the mutational landscape model with data on 1-step adaptation is like claiming to test a new model for long-range weather predictions using data from only 1 day.
The second problem is that Rokyta, et al respond to the relatively poor fit of Orr’s model by successively discarding every unique feature. The next thing to go was the assumption of uniform mutation. As I noted earlier, there are strong mutation biases at work. So, in the middle panel of the figure above, they present a prediction that depends on EVT and assumes Haldane’s 2s, but rejects the uniform mutation rate. In their best model (right panel) they have discarded all 4 assumptions. They have measured the fitnesses (Table 1, above), and they aren’t a great fit to an exponential, so they just use these instead of the theory. Haldane’s 2s only works for small values of s like 0.01 or 0.003, but the actual measured fitnesses go from 0.11 to 0.39! Rokyta, et al provide a more appropriate probability of fixation developed by theoretician Lindi Wahl that also takes into account the context of phage (burst-based) replication. To summarize,
Assumption 1 of the MLM. The exponential distribution of fitness among the top-ranked genotypes is tested, but not tested critically, because the data are not sufficient to distinguish different distributions.
Assumption 2 of the MLM. Gillespie’s “mutational landscape” strategy— his model for how the changing horizon cuts off previous choices and offers a new set of choices at each step— isn’t tested because Rokyta’s study is of 1-step walks.
Assumption 3 of the MLM. The assumption that the probability of a step is not dependent on u, on the grounds that u is uniform or inconsequential, is rejected, because u is non-uniform and consequential.
Assumption 4 of the MLM. The assumption that we can rely on Haldane’s 2s is rejected, for 2 different reasons explained earlier.
I’m not objecting so much to what Rokyta, et al wrote, and I’m certainly not objecting to what they did— it’s a fine study, and one that advanced the field. I’m mainly objecting to the way this study is cited by pretty much everyone else in the field, as though it were a critical test that validates Orr’s approach. That just isn’t supported by the results. You can’t really test the mutational landscape model with 1-step walks. Furthermore, the results of Rokyta, et al led them away from the unique assumptions of the model. Their revised model just applies origin-fixation dynamics in a realistic way suited to their experimental system— which has strong mutation biases and special fixation dynamics— and without any of the innovations that Orr and Gillespie reference when they refer to “the mutational landscape model.”
The interesting things at Sandwalk always seem to happen when I’m not looking. On Sunday, while I was out west taking the offspring to start university at UBC, Larry Moran posted a blog on Constructive Neutral Evolution that has elicited almost 200 comments. Alas, many of the comments are not particularly useful, as Sandwalk is home to an ongoing pseudoscientific debate on intelligent design.
The one point that I would like to make about CNE is that it was not proposed as some kind of law or tendency (i.e., not like “Biology’s First Law” of McShea and Brandon). Some other people treat CNE as the manifestation or the realization of some kind of intrinsic tendency to complexification. If this were the case, then examples of reductive evolution (e.g., cases involving viruses and intracellular parasites) would raise a question about the generality of the idea. Obviously evolutionary change occurs in both reductive and constructive modes. Bateson and Haldane each speculated that reductive evolution would be common because it is so easy to accomplish.
From my perspective, CNE is not a theory about a general tendency of evolution. Instead it is a schema for generating specific testable hypotheses of local complexification.
One also can imagine a mode of Reductive Neutral Evolution in which simplification occurs. It is simply a matter of the local position of the system relative to the spectrum of mutational possibilities.
The return of mutationism to mainstream evolutionary biology is evident in the way mainstream articles now describe the role of mutation in evolution, in our reliance on mathematical models that evoke a mutationist view, and in evo-devo research programs that focus on identifying causative major-effect mutations.
This shift has happened in a kind of sub-conscious way, without commentary or reflection. I’ll comment below on the reasons for that.
My main purpose here is to contrast way that the neo-Darwinian and mutationist views refer implicitly to two different regimes of population genetics evoked in two styles of self-service restaurant: the buffet and the sushi conveyor.
Over at Sandwalk, Larry Moran posted some interesting bits rrom his molecular evolution class exam, including a passage from Mike Lynch arguing for his claim that “nothing in evolution makes sense except in the light of population genetics”. In this passage, which I’ll quote below, Lynch says that evolution is governed by 4 fundamental forces.
The idea that evolution is governed by population-genetic “forces” is common but fundamentally mistaken. I wish we could just put this to rest.
To understand the potential role of mutation in evolution, it is important to understand the enormous range of rates for different types of mutations. If one ignores this, and thinks of “the mutation rate” as a single number, or if one divides mutation into point mutations with a characteristic rate, and other mutations that are ignored, one is going to miss out on how rates of mutation determine what kinds of changes are more or less common in evolution. It would be like allowing a concept of fitness, then subverting its utility by distinguishing only two values, viable and inviable. 
Last year Sahotra Sarkar published a paper that got me thinking. His piece entitled “The Genomic Challenge to Adaptationism” focused on the writings of Lynch & Koonin, arguing that molecular studies continue to present a major challenge to the received view of evolution, by suggesting that “non-adaptive processes dominate genome architecture evolution”.
The idea that molecular studies are bringing about a gradual but profound shift in how we understand evolution is something I’ve considered for a long time. It reminds me of the urban myth about boiling a frog, to the effect that the frog will not notice the change if you bring it on slowly enough. Molecular results on evolution have been emerging slowly and steadily since the late 1950s. Initially these results were shunted into a separate stream of “molecular evolution” (with its own journals and conferences), but over time, they have been merged into the mainstream, leading to the impression that molecular results can’t possibly have any revolutionary implications (read more in a recent article here).
Debates on “gradualism” in evolutionary biology address the size distribution of evolutionary changes. The classical Darwinian position, better described as “infinitesimalism”, holds that evolutionary change is smooth in the sense of being composed of an abundance of infinitesimals (not one infinitesimal at a time, but a blending flow of infinitesimals). The alternative is that evolution sometimes involves “saltations” or jumps, i.e., distinctive and discrete steps. The dispute between these two positions has been a subject of acrimony at various times in the 20th century, with several minor skirmishes, and a larger battle with at least one genuine casualty (image).
Today, over a decade into the 21st century, we have abundant evidence for saltations, yet the term is virtually unknown, and we still seem to invoke selection under the assumption of gradualism. Are we saltationists, or not? I’m going to offer 3 answers below.
But first, we need to review why the issue is important for evolutionary theory.
In a previous post called “The revolt of the clay“, I described four different ways to think about the role of variation as a process with a predictably non-random impact on the outcome of evolution. The main point was to draw attention to my favorite idea, about biases in the introduction of variation as a source of orientation or direction, and to provide a list of what (IMHO) represents the best evidence for this idea. I gave anecdotes from four categories of evidence
- mutation-biased laboratory adaptation
- mutation-biased parallel adaptation in nature
- recurrent evolution of genomic features
- miscellaneous evo-devo cases such as worm sperm
With the publication of a study by Couce, et al., 2015 (a team of researchers from Spain, France and Germany), the first category just got stronger. Couce, et al did an experiment that I’ve been trying to talk experimentalists into doing for a long time: they directly compared the spectrum of lab-evolved changes between two strains with a known difference in mutation spectrum.
In brief, they exposed 576 different lines of E. coli to a dose of cefotaxime that started well below the MIC and doubled every 2 days for 12 iterations. The evolved lines fell into 6 groups depending on 3 choices of genetic background— wild-type, mutH, or mutT— and 2 choices of target gene copy number— just 1 copy of the TEM-1 gene, or an extra ~18 plasmid copies.
The ultimate dose of antibiotic was so extreme that most of the lines went extinct. The figure shows how many of the 96 lines (for each condition) remain viable at a given concentration.
The mutators were chosen because they have different spectra. The “red” mutator (mutH) greatly elevates the rate of G –> A transitions, and also elevates A –> G transitions. The “blue” mutator (mutT) elevates A –> C transversions. The figure above indicates that the blue mutator fared slightly better. A much stronger effect is that the lines with multiple copies of the TEM-1 gene did better. Very few single-copy strains survived the highest doses of antibiotic.
The authors then did some phenotypic characterization that led them to suspect that it was not just the TEM-1 gene, but other genes that were important in adaptation. In particular, they suspected the gene for PBP3, which is a direct target of cefotaxime.
They sequenced the genes for TEM-1 and PBP3 from a large sample of the surviving strains above, with results shown in this figure:
The left panel, i.e., figure 3a, shows that evolved strains of mutH (red), mutT (blue), and wild-type (black) sometimes have one of the well-known mutants in the TEM-1 gene encoding beta-lactamase. The predominance of red means that more mutH strains evolved the familiar TEM-1 mutations, which makes sense because these mutations happen to be the G–>A and A–>G transitions favored by mutH.
Panel 3b, which refers to mutants in the gene encoding PBP3, is a bit complicated. As before, the histogram bars are colored based on the background in which the particular mutation listed on the x axis is fixed. However, rather than showing all the lines together in one histogram, they are arranged into 3 different histograms depending on whether the mutational pathway matches the pathways preferred by mutT (left sub-panel), by mutH (center sub-panel), or neither (right sub-panel). For instance, blue appears in 16 columns, so there are 16 different mutant PBP3 types that evolved in mutT backgrounds, and 14 of them are in the left sub-panel indicating mutT-favored mutational pathways. Red appears in 18 columns, 15 corresponding to mutH-favored pathways.
Importantly, the density in the columns doesn’t overlap much: nearly all of the blueness is in the left histogram (~50/52), and nearly all the redness is in the center histogram (~26/31). This means that evolved mutT strains overwhelmingly carried mutations favored by the mutT mutators, and likewise for mutH strains and mutH mutators. If adaptation were unaffected by tendencies of mutation, as the architects of the Modern Synthesis believed, then the colors would be randomly scrambled between the 3 histograms. Because it is not, we know that the regime of mutation imposes a predictable bias on the spectrum of adaptive changes.
The main weakness of the paper, perhaps, is that they did not do reconstructions to show causative mutations. Instead they make plausibility arguments based on having seen some of these in other studies of laboratory or clinical resistance. I think the paper is pretty strong in spite of this. Recurrence of a mutation in independent lines is a very strong sign that it is causative, whereas singletons might be hitch-hikers with no causative significance. Yet if we were to eliminate the 17 non-recurring mutants, we would still have a striking contrast. Of ~43 mutT mutants with recurrently fixed mutations, ~41 are for mutT-favored pathways; of the ~23 mutH mutants with recurrently fixed mutations, ~18 are for mutH-favored pathways.
Unfortunately, due to the language and framing of this paper, casual readers might not recognize that it is actually about mutation-biased adaptation. The term “mutation bias” only occurs once, in the claim that “Mutational biases forcibly impose a high level of divergence at the molecular level.” I’m not sure what that means. What I would describe as “evolution biased by mutation” is described by Couce, et al as “limits to natural selection imposed by genetic constraints”. Sometimes the language of my fellow evolutionists seems like a strange and foreign language to me. But I digress.
To summarize, this work clearly establishes that 2 different mutator strains adapting to increasing antibiotic concentrations exhibit strikingly different spectra of evolved changes that match their mutation spectra, consistent with the a priori expectation of mutation-biased adaptation.
A “chance” encounter
Earlier this month I was contacted by a reporter writing a piece on the role of chance in evolution. I responded that I didn’t work on that topic, but if he was interested in predictable non-randomness due to biases in variation, then I would be happy to talk. We had a nice chat last Friday.
I’m only working on the role of “chance” in the sense that, in our field, referring to “chance” is a placemarker for the demise of an approach based implicitly on deterministic thinking— evolution proceeds to equilibrium, and everything turns out for the best, driven by selection. This justifies the classic view that “the ultimate source of explanation in biology is the principle of natural selection” (Ayala, 1970). Bruce Levin and colleagues mock this idea hilariously in the following passage from an actual research paper:
To be sure, the ascent and fixation of the earlier-occurring rather than the best-adapted genotypes due to this bottleneck-mutation rate mechanism is a non-equilibrium result. On Equilibrium Day deterministic processes will prevail and the best genotypes will inherit the earth (Levin, Perrot & Walker, 2000)
Wait, I’m still laughing. (more…)