Abstract
Several studies have documented a global pattern of phenological advancement across multiple taxa that is consistent with ongoing climate change1–3. However, the magnitude of these phenological shifts is highly variable across taxa and locations2–4. This variability of phenological responses disrupts species interactions under climate change5–9, but has been difficult to explain mechanistically10–13. To understand how climate change could evoke such variable responses in different groups of organisms, we constructed a model for the evolution of phenological cueing strategies using historic climate data from 78 locations in North America and Hawaii. Here we show how phenological cueing strategies can evolve in predictable ways, but still express highly variable responses to climate change. Across locations, organisms in our model evolved diverse strategies that reflected geographic differences in the reliability of different environmental cues for predicting future conditions. Within locations, a wide range of evolved strategies showed similar emergence phenotypes under historical conditions. However, these same strategies revealed previously hidden and variable responses under novel climatic conditions, with strong fitness consequences. These cryptic differences in cueing strategies evolved under historical conditions because epistasis and non-additive genotype × environment interactions among years resulted in weak selection gradients across an extensive region of trait space. These findings show how the evolution of integrated phenological cueing strategies can explain observed variation in phenological shifts and unexpected responses to climate change.
Recent years have seen increasing interest in the study of phenological shifts. While organisms around the world have generally shown a “global coherent fingerprint” of advancing phenology with climate change1–3, several studies also point to substantial unexplained variation in phenological shifts2–4. This variation in responses to climate change is an important factor driving phenological mismatch and the disruption of species interactions8, 9. It has become increasingly clear that understanding how organisms integrate multiple environmental cues will be necessary to explain and predict phenological shifts10–13.
In order to examine the causes of variation in phenological shifts, we developed a model that simulates how different environmental histories shape the evolution of phenological cueing strategies. We hypothesized that organisms experiencing different environmental histories would evolve different phenological strategies, caused by consistent differences in the reliability of predictive information provided by different kinds of environmental cues. We further hypothesized that these variable strategies create variation in phenological responses to climate change.
Our model simulates the evolution of a generalized organism in a simplified environment defined by daily maximum temperature, total daily precipitation, and day of the year (hereafter, temperature, precipitation and day). These environments are drawn from historic climatic records representing an average of 98 years from each of 78 locations in North America and Hawaii. In our model, these real-world environmental data provide cues to anticipate future environmental conditions each year, and also determine the fitness of individuals in the population (Extended Data Fig. 1). The environmental cues (E) on each day are cumulative annual daily maximum temperature (γtemp), cumulative annual daily precipitation (γprecip), and day of year (γday):
The use of cumulative annual temperature and precipitation is based on the assumption that organisms are aware of and can be influenced by past environmental conditions, consistent with degree-day models of development and phenology. Each individual has a genotype (G) defined by three traits (τ), which reflect its sensitivity to the three environmental cues:
Each day of the simulation, each individual combines its cues and genotype into a weighted sum, which represents the emergence signal (S):
When this signal crosses the threshold S≥1, the organism makes an irreversible decision to emerge. Its fitness is then dependent on daily temperature and moisture conditions over a fixed window, beginning one day after the threshold is crossed. We defined fitness as a function of daily temperature and moisture using a multivariate skew-normal distribution with optimal temperature and moisture values set to the 90th percentile of all values observed in a given location. This assumes that fitness is a fixed, asymmetric function of environmental conditions and that physiological performance is adapted to past local conditions. For each location, we ran 60 simulations with the same parameters. Each simulation included 1000 randomly resampled years of climatic data and a population of 500 individuals whose initial genotypes were drawn from a broad uniform random distribution (Extended Data Fig. 2).
The results of this model suggest two key findings. First, the mean evolved strategies of each location in our model are spatially autocorrelated, indicating that similar strategies evolved in locations that experienced similar climates, while individuals from different climates consistently evolved different strategies (Fig. 1, Fig. 2, Extended Data Fig. 3 and Extended Data Fig. 4). These evolved differences reflect geographic differences in the relative reliability of temperature, precipitation and day cues for predicting future fitness outcomes. This finding was qualitatively robust across model variants that used different cues, fitness functions, and emergence durations (Extended Data Fig. 5). We evaluated several climatic and location-based variables in order to explain the pattern of evolved cueing strategies, but most showed small or counterintuitive effects (Extended Data Fig. 6). In general, these evolved strategies did not conform to predictions based on simple assumptions about latitude, climatic variability, predictability or seasonality14, 15; instead, they illustrated the complexity of the location-specific climatic regimes and fitness landscapes that affect organisms in this model. Selection favored cues based on their ability to predict future environmental conditions – both the ability to consistently trigger emergence ahead of favorable conditions, and the ability to avoid triggering emergence ahead of unfavorable conditions. Although the simplified environmental data in our model showed unexpected complexity, the spatial autocorrelation of evolved strategies in our model suggests that information constrains the evolution of phenological cueing strategies in predictable ways.
The second key finding of this model is that repeated simulations from the same location produced diverse strategies that expressed similar phenotypes under historical conditions (Figs. 2 and 3, Extended Data Fig. 3), but showed strong phenotypic and fitness differences under simulated climate change (Fig. 4, Extended Data Fig. 7). This finding was an unanticipated consequence of cue integration, which allows a wide range of strategies to be almost equally effective under the same environmental conditions. This pattern emerges in our model because the function combining environmental conditions and traits is non-injective: multiple combinations of traits can yield the same expressed phenotype. In our model, multiple cues are combined to yield a single emergence signal; in this context, epistatic interactions between traits mean that increased sensitivity to one cue can compensate for reduced sensitivity to another. Moreover, the expressed phenotype results from a non-additive genotype × environment interaction which allows genotypes to show different reaction norms across years16, driven by the complex structure of real-world climate data. On average, genotype × environment interactions explained 29% of observed fitness variation across all locations (Extended Data Fig. 8). As a result, the phenotypic and fitness differences between evolved genotypes are inconsistent from year to year in a way that reduces the long-term mean fitness differences between genotypes (Extended Data Fig. 9). The outcomes of these processes are weak selection gradients across a wide range of trait space (Extended Data Fig. 10). The resulting diversity of phenological cueing strategies could contribute to observed variation in phenological responses to climate change2, 3, 17 and the evolution of cryptic genetic variation18, 19.
Climate change had a strong effect on both phenotypes and fitness in most locations (Fig. 4, Extended Data Fig. 7). We modeled two simple climate change scenarios. In the first (“shift”) scenario, we advanced both temperature and precipitation regimes by 5 d. In the second (“warming”) scenario, we increased all daily maximum temperatures by 3°C. With each changed climate regime, the mean genotypes that evolved under historical conditions in 30 simulations were evaluated against all available years. In both scenarios, populations advanced their phenology and generally showed reduced fitness (Fig. 4). These effects were non-random; for example, organisms in the shift scenario that relied more on day cues were less likely to advance their emergence timing on pace with the changed climate (χ2(1)=466.7, p<0.00001), and generally showed negative fitness consequences relative to individuals that favored temperature or precipitation cues (χ2(1)=11.1, p=0.0009, Fig. 4). This result is consistent with expectations about the costs of relying on invariant day-of-year cues under a climate change scenario20, 21, and the potential for adaptive plasticity in response to changing climates. By comparison, under the warming scenario, organisms with greater reliance on day cues also showed reduced phenological advancement (χ2(1)=35.5, p<0.00001), but generally showed higher fitness than organisms that were more reliant on climatic cues (χ2(1)=9.0, p=0.0026). This unexpected result occurs because the warming scenario broke the historic correlation structure between temperature- and precipitation-based factors in most climates, causing many strategies to demonstrate maladaptive plasticity in a novel climate22, 23.
The key prediction of this model is that evolved phenological cueing strategies will show hidden variation in their responses to climate change. In our model, environmental history shaped the evolution of phenological cueing strategies in ways that reflected local differences in environmental conditions.
However, instead of favoring a single climatically-determined optimum strategy in each location, selection produced substantial variation in the ways individuals combined multiple cues to determine their emergence phenology. Importantly, the fundamental mechanisms driving this finding will emerge from any reasonable model of cue integration where multiple cues are combined to inform phenological decisions. This underlying principle likely contributes to the observed variability of phenological responses to novel climates around the world, and challenges our ability to predict phenological responses and fitness consequences under climate change. Developing a stronger understanding of phenological cueing mechanisms may improve our ability to understand and predict the ecological effects of climate change.
Methods
Environmental data
All available years of daily maximum temperature (degrees Celsius) and daily precipitation (mm rainfall equivalent) data were obtained from the NOAA Climate Data Online portal24 for 82 locations in North America and Hawaii. Years with <325 daily temperature and precipitation observations were excluded from further analysis, and four locations with <50 years of data remaining were also excluded. For each of the remaining 78 locations, the interquartile range (IQR) was calculated as the difference between first quartile (Q1) and third quartile (Q3) observations. Temperature observations less than (Q1−4*IQR) and greater than (Q3+4*IQR) were identified as extreme outliers likely resulting from measurement error, and were excluded; such outliers were a small proportion (0.0012%) of the overall dataset. Missing observations in the remaining dataset (less than 1% of observations) were imputed using an expectation-maximization with bootstrapping (EMB) algorithm. Imputed values for temperature and precipitation were bounded by the observed minimum and maximum values of each location, and informed by priors based on the means and standard deviations of each location. This procedure imputed a complete dataset of daily maximum temperatures and daily precipitation for each location, with an average duration of 98 years (SD = 18.9 years).
Organisms used cumulative temperature and precipitation as climatic cues, and their fitness was affected by daily temperature and moisture during their emergence period. Temperature for each location was shifted so that the minimum transformed temperature for that location was zero. This meant that cue values were always non-negative. Environmental moisture (m) was calculated based on daily precipitation totals (p) in the dataset using a formula that includes a proportional retention constant (α, set to 0.8) to represent the partial retention of moisture over time, as well as the input of new precipitation each day (p). Changing the retention constant did not qualitatively change the model.
Day of the year was represented as an integer value reflecting the number of days since January 1 of each year inclusive. The 366th day was truncated from leap years in the dataset. Day of year provides a proxy for a consistent and non-climatic environmental cue akin to photoperiod, implicitly assuming physiological mechanisms in each organism that are able to infer the day of the year from photoperiodic dynamics with equal accuracy across all locations. This simplifying assumption is supported by studies showing that although the amplitude photoperiodic changes is larger at higher latitudes, tropical species are able to detect the extremely small changes in photoperiod that occur near the equator25, 26. This assumption also allows us to infer the relative information content of day as a cue across multiple locations, separate from the effect of increasing photoperiodic amplitude at higher latitude. In examining spatial variation in the evolution of phenological strategies, day of year conservatively assumes that the phenological information available to organisms is unaffected by latitude. Using cumulative photoperiod produced qualitatively similar results (e.g., Extended Data Figure 5).
Emergence signal
Organisms in our model accrue fitness for 10 days after the emergence signal exceeds 1. To facilitate interpretation, this emergence signal model uses linear coefficients equal to the inverse of the trait value, so that genotypic traits are represented in same units as the cue itself, and trait values indicate the critical cue value that would trigger emergence in the hypothetical absence of other cues. Thus, large trait values correspond to low sensitivity and small trait values correspond to high sensitivity to the corresponding cue.
Fitness and reproduction
Individuals that emerge reproduce at a rate proportional to the cumulative fitness they accrue over their lifespan. The fitness gained on any given day is the product of two skew-normal function outputs – one based on temperature, the other on moisture. The thermal performance curves of ectotherms are generally asymmetrical, with a sharp decline above their optimal temperature and a more gradual decline below it27, 28. For simplicity, we used the same skew normal functional form (with a skew parameter of −10) for both temperature and moisture. This function was parameterized separately for each location for both temperature and moisture, such that the peak for each occurred at the 90th percentile of temperature and of moisture of all daily observations for a given location, and the function had a value that was 10 percent of the peak when the cue was at the 10th percentile of all daily observations. This parameterization assumes that organisms are physiologically adapted to the prevailing conditions at each location, with an expected 10% of temperature and moisture values exceeding the optimal values. To evaluate the robustness of observed results, we tested this model at a range of alternative fitness parameterizations with qualitatively identical results (see below). Because the skew normal function does not have a simple mathematical relationship between its parameters and the location of the peak, we fit parameter values by minimizing the sum of squared errors. After both skew normal functions were parameterized, we calculated the daily fitness payoff for each day in each year of each climate as the product of the two skew normal function outputs. The fitness of each emerged individual (Wi) was calculated as the sum of daily fitness payoffs over its lifespan. Organisms that did not emerge by the end of the year received zero fitness. Individuals reproduced asexually with mutation (see below), with population size held constant at 500 individuals and expected realized fitness of each individual proportional to its calculated relative fitness. Reproduction was implemented as a lottery model to allow for demographic stochasticity. For each evolved strategy in the final generation, we calculated its geometric mean fitness across all years.
Heritability and mutation
Offspring genotypes reflected the trait values of their parent, modified by mutation. We modeled mutation by adding random numbers drawn from a normal distribution with mean 0 and a small standard deviation to the traits of all individuals in each generation. We set the standard deviation of mutation to be 0.5 percent of the overall cue range in order to produce mutation distributions with the same expected effect size in each location. In the case of the day cue, we used 360 as the maximum, leading to a standard deviation of 1.8 for mutation rate of the day trait in all locations.
Initialization and execution
For each simulation, each individual in the initial generation was assigned uniform random trait values between 0 and 4*(the maximum cue value in that location, or 360 in the case of the day cue). This results in an initial population of individuals with emergence phenotypes ranging between emerging on day 1 and never emerging. Each simulation run experienced 1000 years of environmental conditions drawn by year with replacement from the set of available environmental data for that location (e.g., Extended Data Figure 2).
Assessing realized relative cue use
We define the “trait effect” (Τ) as a metric of proportional cue use in order to assess the relative degree to which an organism’s emergence decision was affected by each environmental cue. This metric quantifies each organism’s realized reliance on different cues represented by the proportion of the emergence signal that is contributed by each term on the day emergence is triggered. This metric allows the relative contribution of each cue type to be compared across locations. The trait effect metric also allows realized relative cue use to be analyzed as a dataset of mathematical compositions, and thus plotted on ternary plots. Because of the discontinuous nature of daily cues, individuals might have an S > 1 when they emerged; we rescaled the trait effect values for each individual so that they sum to 1. For each individual in the final population at the end of each simulation, we calculated the Aitchison compositional mean3 of the trait effects that would have been realized for each year of actual climate data used. This compositional mean represents the expected relative cue use of that genotype in the historic environment.
Sensitivity analyses
We conducted sensitivity analyses using a range of values for lifespan, lag, moisture decay parameter α, and normal and skew-normal fitness distributions using different optimal temperature and moisture values. All parameters and distributions yielded qualitatively identical results, and we present a representative set of model parameters. We also tested additional climatic cues, including daily temperature and moisture, cumulative photoperiod, and quadratic formulations of temperature, moisture and day. These models showed results consistent with those presented here.
Analysis of explanatory factors
We conducted analyses to examine correlations between evolved strategies and three sets of potential explanatory variables. The first set of analyses considered five location variables that provide a broad biogeographic description of each location: distance to coast, elevation, latitude, mean annual precipitation and mean annual temperature. The second set of analyses focused on six variables that quantify climatic variance and predictability: the mean annual coefficient of variation for daily maximum temperature, the mean annual coefficient of variation for daily precipitation total, the coefficient of variation for annual mean daily maximum temperatures, the coefficient of variation for annual mean daily precipitation totals, the lag=1 autocorrelation coefficient for daily maximum temperature, and the lag=1 autocorrelation coefficient for daily precipitation totals. The first two of these variables provide metrics of intra-annual climatic variation, the second two provide metrics of inter-annual climatic variation, and the last two provide metrics of short-term predictability. The third set of analyses used six published metrics of climatic predictability, variability and seasonality12, 30: Lisovski et al.’s predictability and seasonality metrics for temperature and precipitation, and Pau et al.’s variance metrics for temperature and precipitation. Because the Pau et al. and Lisovski et al. metrics of seasonality were highly correlated, we did not also include the Pau et al. metrics of seasonality.
These analyses used a dataset composed of the mean strategies that evolved in each simulation conducted in each location. For each analysis set, we used linear mixed models including all a priori explanatory variables as fixed factors for each trait effect dimension, with an additional random factor to allow intercepts to vary by location. We find qualitatively identical results using logit-transformed trait effects, but present analyses of untransformed data here so that the effect sizes are reported in interpretable units (Extended Data Fig. 6).
Climate change scenarios
We examined how the individuals from the final generation of each simulation performed in novel climate regimes using two simple climate change scenarios. In the “shift” scenario, we advanced the historic temperature and precipitation regime in each year by 5 days. In the “warming” scenario, we increased all daily temperatures by 3 degrees; the precipitation regime was unchanged. In both scenarios, we calculated the emergence, fitness, and Aitchison compositional mean 29 of the trait effects that would have been realized for each individual in each unique year of the modified climate regime. This allowed us to assess how climate change affected the phenotype and fitness consequences of each genotype that evolved under historical conditions.
We assessed correlations between each trait effect (Τ) and the change in emergence timing, and between each trait effect and geometric mean fitness for each evolved genotype, in all cases using linear mixed models with location as a random factor allowing intercepts and slopes to vary.
Author Contributions
Both authors contributed equally to this work.
Author Information
We do not have any competing financial and/or non-financial interests in relation to the work described.
Acknowledgements
We thank Stephen Ellner, Andy Sih, Sebastian Schreiber, Jay Rosenheim, and Jaime Ashander for comments on the development of this work. CBE was supported by an NSF Graduate Research Fellowship, and LHY was supported by NSF DEB-1253101.