Saturday, March 23, 2024

A 25-year quest for the Holy Grail of evolutionary biology

When I started my postdoc in 1998, I think it is safe to say that the Holy Grail (or maybe Rosetta Stone) for many evolutionary biologists was a concept called the Adaptive Landscape. The reason for such exalted status is that the adaptive landscape was then – and remains – the only formal quantitative way to predict and interpret an adaptive radiation of few organisms into many. I was heavily indoctrinated into this framework - as my postdoc was at UBC during precisely the time when Dolph Schluter was writing his now-classic book The Ecology of Adaptive Radiation.

Adaptive landscapes come in several forms (e.g., genotype based, allele frequency based, phenotype based), and the one we are concerned with here is the “phenotypic adaptive landscape.” Perhaps the first serious description of this landscape was the one presented by George Gaylord Simpson in his 1953 book The Major Features of Evolution. In essence, Simpson’s landscape depicts phenotype combinations (e.g., population average values for two important traits, such as beak size and beak shape for birds) and the expected fitness associated with a population having those mean values. The resulting landscapes are expected to be “rugged,” with various peaks of high fitness separated by valleys of low fitness. The idea is that a lineage (e.g., of birds) is expected to diversify phenotypically into multiple species that each occupy one of those different high-fitness peaks, with few individuals having phenotypes in the low-fitness combinations between the peaks. From one species thus comes many in a predictable way by different species matching their traits to particular resources or habitats that represent high-fitness peaks on the adaptive landscape,

This adaptive landscape idea remained largely conceptual, or heuristic, until Lande (1976, 1979) figured out how to represent it mathematically. Then a bit later, Lande and Arnold (1983) showed how to link the adaptive landscape to formal estimates of natural selection at the phenotypic level – thus providing a means for formally estimate the landscape. This theoretical work spurred several decades of intense interest in attempting to quantify adaptive landscapes – or parts of them – for particular adaptive radiations. Yet the effort required turned out to be rather extreme for several reasons. Most critically, perhaps, the full landscape can be specified only if one knows the fitness of individuals across the entire range of phenotypic space (e.g., from small beaks to large beaks for every possible beak shape value from pointy to blunt). Yet, almost by definition, individuals with phenotypes of low fitness should be rare in nature -  because they are constantly selected against; and so the fitness of phenotypes between species tend to be unknown in nature.

The Holy Grail

Owing to this problem of “missing phenotypes,” as well as other difficulties, I would argue that no formal adaptive landscape – in the Lande sense – existed for any natural radiation of organisms at the time I started my postdoc. Yet its predictive promise made it the Holy Grail of the time.

Although no adaptive landscapes had been formally estimated, some studies got part way there (see the Appendix below on “other landscapes I have known and loved”). As one example, Benkman (2003) measured the performance of crossbills on different types of conifer cones and used these estimates and other information to construct an “individual fitness landscape” spanning the range of phenotypes – that is, the expected fitness of an individual having each possible combination of beak trait values. As had been predicted, the landscape had peaks of high fitness near the average phenotypes of the different crossbill types – and those peaks were separated by phenotypes with low fitness – yet something was missing.

This figure is from my 2017 book.

In particular, a formal adaptive landscape relates MEAN phenotypes for a population (assuming a particular phenotypic variance) to the expected MEAN fitness of that population, which requires a particular conversion – as the gif below (made by Marc-Olivier Beausoleil) illustrates. When this conversion takes place, the fitness peaks tend to sink and the fitness valleys tend to rise – because the adaptive landscape averages fitness across all of the phenotypes in a population, which inevitably span a range of fitnesses on either side of a peak (or valley). That is, because a population adapted to a fitness peak will have some variation in phenotypes, the off-peak individuals will reduce mean population fitness relative to a population where all individuals were identical and had phenotypes EXACTLY on the peak. The inevitable is that adaptive landscapes based on population means are always smoother than those based on individual fitness values. As an example, the individual fitness peaks for crossbills shown above tend to disappear if converted to a formal adaptive landscape based on population means (C. Benkman pers. comm.)

So, as time went on, a formal adaptive landscape for an adaptive radiation in nature remained elusive.

My Galapagos Dreams

In 2002, I had the opportunity to visit Galapagos at the behest of Jeff Podos (UMASS Amherst) who had just received an NSF Career Grant that could fund my participation. I was really excited and read all of the classic books by Lack (1947), Bowman (1961), and Grant (1998). Yet, I still had no real practical experience with Galapagos or even with bird research. Thus, on my first visit, I decided to not actually conduct any of my own research but rather learn about the system – both by helping Jeff with his research and also simply walking around in the field making natural history observations that might motivate future experiments. A highlight from that year was the afternoon I spent in the field with the deacons of all things finch, Peter and Rosemary Grant.

One of the major efforts of Jeff’s team was to capture finches at a focal site (El Garrapatero), measure their beak traits, band them, and then relate those traits to song features and mating. Within a year, I had encouraged an increasing emphasis on using the banded birds to track inter-annual survival. When it was clear that this effort would be productive, I set the goal of – once and for all – estimating the adaptive landscape – in the formal way – for Darwin’s finches at this site.

My first step in this effort was to use the extremely-variable population of Geospiza fortis (the medium ground finch) at El Garrapatero to measure selection across their phenotypic distribution. This population had long been suggested to be bimodal in its beak size distribution, with somewhat distinct “large” and “small” beak morphs. My idea was that the fitness landscape should show a valley between the two morphs, such that intermediate birds have lower fitness – a selective function called disruptive selection. Using two years of data, low-and-behold, disruptive selection between the beak modes (and stabilizing selection around each mode) is what we found. This initial discovery was exhilarating but (1) we had only two years of data, (2) we had a potentially imprecise measure of fitness (survival over one year), and (3) we were looking at only one species.

This figure is from my 2017 book.

It took another 10 years of accumulated data – contributed by many team members and collaborators working at El Garrapatero – to solve the first of those problems. In particular, Marc-Olivier Beausoleil (graduate student of Rowan Barrett at McGill) was able to compile years of the most intensive data collection for G. fortis to show that, yes, disruptive selection was always working to maintain some separation of the two morphs (large and small) – and that the intensity of this selection could be explained by weather conditions. Specifically, disruptive selection was strongest when dry periods followed wet periods – probably because many fledglings were produced during wet periods which then increased competition (and hence mortality) during subsequent dry periods.

But these landscapes were still at the level of the individual (rather than population mean), and they still showed only one species, whereas the Geospiza radiation at this location has four species: the small ground finch (Geospiza fuliginosa), the medium ground finch (Geospiza fortis), the large ground finch (Geospiza magnirostris), and the cactus finch (Geospiza scandens). When would we have enough data for this more comprehensive effort.

After 17 years of data collection that included more than 3400 individual birds, we set out to give it a try. For a fitness surrogate, we chose longevity, which Peter and Rosemary had previously shown was strongly correlated with total fitness (note: we could not track reproductive output, such as the number of fledglings, in our population). Marc-Olivier Beausoleil led this effort and first calculated an individual fitness landscape, relating the fitness of individual birds to their individual phenotypes across the entire dataset. As expected, and as shown in other studies (crossbills, African seedcrackers, and more – see Appendix), the individual fitness landscape was rugged, showing clear peaks and valleys. Further, the average phenotypes of each species (and the two morphs of G. fortis) were situated fairly close – in phenotypic space – to the estimated peaks of high fitness.

Would this reasonable and logical landscape hold up when the individual fitness landscape was converted to a formal population mean adaptive landscape. We held our breadth – and I was skeptical given how no study of an adaptive radiation had been comfortable doing so before. As expected, the landscape was smoother and a peak or two sank to the point of obscurity: yet, remarkably, the adaptive landscape still had peaks and valleys and the mean phenotypes of each species (or G. fortis morph) was reasonably close to a peak of mean fitness. IT WORKED – and it only took 20 years of effort!

So, what did we learn from all this work. Well, for starters, one can – with enough effort – estimate a formal adaptive landscape for a real adaptive radiation in nature. Second, these landscapes do have the expected peaks and valleys even when using MEAN phenotypes and fitness – they are rugged indeed! Third, evolution of the various species and morphs seems to follow the estimated landscape, with about as many phenotypic modes (species or morphs) as estimate peaks and with the mean trait values of each close to a different peak. Yet some deviations from these expectations were also seen – note in the figure aboe how the circles (mean phenotypes) are always displaced a bit off the peak. At this point, we expect the deviations arise from our incomplete fitness surrogate: longevity. Perhaps, for instance, the birds that live the longest had the fewest offspring (or, stated the other way around, birds that have the fewest offspring live the longest) – as has been documented in some species.

Was it worth it?

The first ten years of effort were accompanied by high optimism that “this could work.” But, then, as various constraints and funding issues came into play, fatalism set in: “well, it was a nice idea but impractical.” Then I kind of forgot about it for several years while the data stream stayed alive – until Marc-Olivier came in and cleaned up the database and applied his statistical wizardry to G. fortis. Then hope sprang again that we could do the full adaptive landscape and here we are. Despite the effort, I think the accomplishment by the entire team of field organizers, crews, funders, and analyzers is quite remarkable given that – to my knowledge – this is the first formal adaptive landscape estimated for an natural radiation of local species. It has been 25 years coming for me – but we made it.

I should note in closing that the adaptive landscape is more like a Rosetta Stone than a Holy Grail. First, different evolutionary biologist would likely search for different Holy Grails – but, of course, there can be only one. Second, the phenotypic adaptive landscape links phenotypes, fitness, and evolution – and thus is something of a translation between these traits.

Finally, I suppose the real Holy Grail is to link not just phenotypes and fitness across all species in a radiation (or part of a radiation) but to also integrate individual genotypes. Did I mention that Rowan Barrett’s team (with Marc-Olivier) have more than 400 individual whole genomes for these birds – via collaborators in Switzerland (Daniel Wegmann). Stay tuned to this channel!


Appendix: Some other landscapes I have known (and loved) for vertebrates:

1.  1. Smith (1993) estimated a single-trait individual fitness landscape for African seedcrackers, which showed disruptive selection between two beak size morphs – and thus inspired my initial analysis testing for the same thing in G. fortis at El Garrapatero.

2.  2. Schluter and Grant (1984) used seed size distributions on different Galapagos islands to estimate the beak sizes of Darwin’s finches that would be expected to evolve. This landscape was inspiring for our own efforts on finches – in part because it examined all of the ground finch species – as opposed to only single species.

Martin and Wainwright (2013) used hybridization between pupfish species to increase phenotypic variation and then estimate individual fitness landscapes in natural environments – albeit in enclosures. Fortunately, hybridization in our finches inflates the variation naturally – precluding the need to perform artificial crosses (which would be impossible anyway).


4. 4. Arnegaard et al. (2014) performed a similar study the pupfish work by hybridizing threespine stickleback populations and testing their fitness in artificial ponds. The additional innovation in this second study was to link the traits in question to genomic regions.

5.   5. Stroud et al. (2023) estimated fitness landscapes through time for a community (although not a radiation) of Anolis lizards on a very small island.

Friday, February 23, 2024

Predicting Speciation?

(posted by Andrew on behalf of Marius Roesti)

Another year is in full swing. What will 2024 hold for us?

Nostradamus, the infamous French astrologer, is well known for his projections into the future. Some 500 years ago, he also made a series of predictions for 2024. He wrote, for instance, “Red adversary will become pale with fear, putting the great Ocean in dread”, and “The dry earth will grow more parched, and there will be great floods when it is seen".1 These vague predictions that lack a rigorous rationale and are quite open to interpretation (have a go at it!) are probably better called “prophecies”. In contrast, the scientific enterprise has led to a very different quality of predictions. For example, today's astronomy predicts with stunning precision a total solar eclipse for parts of North America on 8 April this year. According to NASA, "... the first location in continental North America that will experience totality is Mexico’s Pacific coast at around 11:07 a.m. PDT. (...) The eclipse will exit continental North America on the Atlantic coast of Newfoundland, Canada, at 5:16 p.m. NDT". 2 Based on NASA's past predictive success we can confidently mark this eclipse in our calendars. As of today, however, meteorology cannot predict whether the weather conditions will indeed allow us to witness the eclipse from Earth. In this respect, a lot of uncertainty remains.

In many life situations, we base our decisions and actions on their predicted consequences. Yet, predictions are also indispensable in fundamental sciences where interests center on understanding how our universe works, rather than assisting with practical life problems. We recently asked some colleagues why we want to predict things in science and often heard something like "because this will tell us whether we got it right or not". – This made us think. If true, what does this mean for a field of research still struggling with making accurate predictions?

(And now for some shameless self-promotion!...)

This question marked the starting point for our essay3, which now appeared online and will be part of a special volume on "Speciation" in CSH Perspectives in Biology. To start, it seems essential to distinguish between two types of scientific predictions because they fundamentally differ in their function and value: in fundamental research, we are mainly interested in "causal predictions", not "correlational predictions". We then identify the three fundamental challenges for making accurate causal predictions in speciation research and discuss which of them are theoretically surmountable. Don't despair, there is hope! We also outline how these and further insights (more in the paper!) could shape future speciation research – namely toward a Standard Model – as well as related research in ecology and evolution.

Indeed, although tailored to speciation research, our essay connects with and easily translates to scientific disciplines beyond this field. We believe there are benefits for many of us in a deep consideration of scientific predictability, whether empiricists, theoreticians, or folks drawn to the philosophy of science. Perhaps even Nostradamus would have found interest in this read. While the fundamental topics we discuss in the essay are not new, we use an integrative and somewhat unorthodox approach – including a thought experiment with an Orrery and a Speciation Machine (see figure below) – to hopefully stimulate not only vivid and fun but also fruitful discussions. In fact, our (many) discussions on predictability have led us to organize a symposium on this topic at this year's joint ESEB/Evolution conference in Montreal (stay tuned for it!).

Our paper well complements and extends Andrew Hendry's recent paper on "Prediction in Ecology and Evolution"4, and these two papers may provide a nice back-to-back read for students, journal clubs, or coffee-break discussions. Below is the abstract of our paper. Should you have trouble accessing it in full, don't hesitate to let me know and I can email you a copy (you should also be able to access the full paper via my personal homepage).

And folks in North America, don't forget to mark 8 April in your calendars for an astronomical spectacle because it won’t happen there again for the next 9 years (3278 days, to be precise)!


1. Nostradamus, M. Les Proph├ęties. Lyon, 1555.

2. From (accessed 28 January      2024)

3. Roesti M, Roesti H, Satokangas I, Boughman J, Chaturvedi S, Wolf JBW, Langerhans RB. Predictability, an Orrery, and a Speciation Machine: Quest for a Standard Model of Speciation. Cold Spring Harbor Perspectives in Biology 2024 Feb 12:a041456. doi: 10.1101/cshperspect.a041456. Epub ahead of print. PMID: 38346860.

4. Hendry A. Prediction in ecology and evolution, BioScience, Volume 73, Issue 11, November 2023, Pages 785–799,

Tuesday, February 20, 2024

Sticklestock center

"There are two kinds of readers.  Those who have read the Lord of the Rings.  And those who are going to."

There are two kinds of biologists.  Those who have worked with stickleback. And those who are going to.  

If you count yourself in the former category, we have good news. You can now buy stickleback through a non-profit center.  The Stickleback Stock Center (aka Sticklestock) at the University of Connecticut provides eggs, juveniles, adults, cell cultures, and associated microbes for research and education. We can do microinjection of CRISPR so you don't have to. You can place orders here:

If you count yourself in the latter category, even better news: not only can you buy stickleback for research, but the website provides detailed protocols to help you get started with husbandry, breeding, field work, lab work, spatial transcriptomics, and more, through our protocols page:

Our goal is to facilitate adoption of this useful research model organism, both by supplying animals (or tissues), and advice to ease the transition.

Why have a stock center? A number of reasons motivated us to start this initiative. 

First, stickleback are highly seasonal animals that breed in mid-spring (depending on where you go). If you want embryos another time of year, you are out of luck. By offering a stock center we hope to provide a year-round source of samples.  

Second, field work is expensive and time-consuming. You need flights, lodging, rental cars, permits, traps. You need to ship fish which can turn into a major hassle when crossing international borders.  We typically do field work at just one location at a time (e.g., this summer I'll go to Vancouver Island for two weeks), which makes it hard to create crosses between geographically far-flung populations. A stock center provides a cost effective alternative, saving researchers time, expense, and reducing carbon footprints.

Third, when we study wild populations we are studying a moving target. For example, my lab has worked on Gosling Lake on Vancouver Island for 20 years now. During that time the population has seen a major evolutionary change leaving genome-wide alterations in allele frequencies, and large allele frequency change at the gene I care most about (spi1b). A stock center population may also evolve, adapting to culture enviroments in aquaria, but provides a stable genotype for future study. Moreover, multiple labs can readily study the sample genotypes facilitating collaboration and replication.

Fourth, we rarely have the option to work with populations that have genome sequences before we ever begin. The stock center plans to work primarily with populations where we already have some whole genome sequence data, or acquire some early on during culture. This allows us to offer our customers genotype-guided choices of populations to work with, an unusual opportunity for a wild vertebrate model system.

We hope the stock center grows and provides a sustained resource for stickleback researchers, current and future. If you are thinking of trying out stickleback research, get in touch! Only through sustained orders will we convince funding agencies that we have the capacity to become a self-sustaining program that can provide long-term support to the community. 

A 25-year quest for the Holy Grail of evolutionary biology

When I started my postdoc in 1998, I think it is safe to say that the Holy Grail (or maybe Rosetta Stone) for many evolutionary biologists w...