Introduction

No two organisms of the same species, even when genetically identical, behave precisely the same. Individuality has been observed and measured in organisms ranging from bacteria(1) to plants(2), flies(3–7) to humans(8), even in the absence of genetic or environmental differences. This variability poses two major questions in biology—how does it arise and what, if any, evolutionary role does it serve.

Heritable phenotypic variation in a population allows for adaptive tracking, i.e., when alleles change in frequency as the selective pressure of the environment changes (9). A complementary strategy for fluctuating environments is phenotypic plasticity, in which organisms change their behavior (or morphology) in direct response to environmental changes. These two adaptive strategies share a basic outcome: the phenotype they produce matches the current or recent environment. Organisms that fail to match the environment can suffer deadly consequences.

Random differences in phenotype may reflect a “bet-hedging” strategy for species to deal with unpredictability in their environment (2, 10–16). Under this theory, variability allows some individuals to survive no matter what future environment arrives, increasing the odds that a population avoids extinction. Thus, bet-hedging species accept a lower arithmetic mean fitness for a higher geometric mean fitness (which equals zero if there is a single generation of no fitness). Theoretical work shows that bet-hedging strategies relying on random non-heritable variation outperform adaptive tracking in some environments (6, 16, 17). In particular, bet-hedging outperforms adaptive tracking when the environment fluctuations on timescales similar to the lifespan, as adaptive tracking requires multiple generations to respond to environmental cues (15, 18), and can lag behind fluctuating selective pressures. While phenotypic plasticity occurs on faster timescales, the ability to adapt and learn can be metabolically costly and provides limited buffer against sudden changes in selective pressures (19). Intrinsic variability in genetic control (1) has been identified as a source of phenotypic variation in microbes. In multicellular organisms, stochastic processes in development underlie variation (20), and, in the case of behavior, stochastic neuronal wiring (4, 21–23) has been shown to predict persistent individual differences.

Individuals, however, do not express a singular behavioral phenotype across their life. Individual biases in multiple behavioral settings are only partially consistent over time (3, 5, 24–27), showing spontaneous changes even in the absence of macroscopic cues that could trigger plasticity. These observations raise two questions: to what degree does individuality arise due to developmental differences versus spontaneous fluctuations within the lifetime of the animal, and do fluctuations across the lifetime of an animal provide a evolutionary benefit? To answer these questions, we used isogenic animals to measure 1) the extent of variation present at the start of adulthood (“bet-hedging”) and 2) the amount behavioral phenotypic change over time (“drift”). We use locomotor handedness in Drosophila melanogaster as a model to measure how behavior drifts over time and show that the extent of drift is influenced by genes and the neuromodulator serotonin. We then use a life-history model to test the hypothesis that drift is adaptive and assess whether real world environmental fluctuations may drive the evolution of drift as a phenotypic strategy. Taken together, these experiments and analyses suggest that phenotypic drift is plausibly an adaptive strategy to cope with environmental fluctuations within an organism’s lifespan.

Results

Individual behavioral biases drift over time

To investigate the timescales on which individual preference spontaneously changes, we placed 252 flies in individual circular arenas and continuously tracked them for up to 30 days. At each time point, we computed the direction of circling (Figure 1A), a measure that exhibits idiosyncratic variation, when averaged over long timescales(3). To look for slow changes in bias, we examined the turning bias of individual flies at different timescales: first, we averaged their direction of circling over each hour and second we lowpass filtered their continuous turning behavior to discard frequency components faster than 24hrs (Figure 1B). Comparing both of these measures to the experiment-wide average preference for each fly shows substantial long-duration changes from their experiment-wide circling tendency. To quantify this across all flies we calculated the average power spectrum of turning across all measured flies (Figure 1C), showing substantial power in lower frequencies corresponding to hours- to days-long fluctuations in turning preferences. Interestingly, we did not see peaks at any specific frequencies (e.g., circadian), though there was a broad shoulder of power between 10⁻⁴ and 10⁻² Hz. The overall trend across six orders of magnitude bore some resemblance to a power-law relationship between frequency and power. We saw similar patterns in other measures of behavior from this experiment, including speed, heading velocity, and distance from the center of the arena (Fig S1A-D).

Fig. 1.

Genes regulate the rate of behavioral drift

The extent of behavioral variability in a population differs between genotypes (28–30). We next looked to see if the rate of behavioral drift also differs between genotypes. This is a prerequisite if drift can evolve as a trait under natural selection. We used 10 different lines from the Drosophila Genetic Resource Panel (DGRP) (31). Using a Y-maze assay (Figure 1D) (3), we measured the locomotor handedness of flies in each of these lines three times weekly for three weeks. As in the circling assay (Figure 1A-C), individual flies exhibited substantial variation in their daily average turn biases (consistent with drifting biases) in all genotypes, as measured by the standard deviation in the daily fraction of right turns for each fly, showing substantial variation from their experiment-wise average (Figure 1E). We used a Bayesian autoregressive model (Figure 1F) to estimate the initial variability in turn bias for each genotype (σ_Bet-Hedging), the extent to which turn bias changed each day (σ_Drift), and the tendency of turn bias to revert to the mean (φ). We observed large differences in all three parameters between genotypes, as evidenced by non-overlapping 95% posterior distributions (Figures 1G, S1E). Interestingly, we saw independent variation across genotypes in the posteriors for σ_Bet-Hedging and σ_Drift, suggesting that different genetic mechanisms regulate bet-hedging and drift.

Serotonin modulates the rate of behavioral drift

Neuro-modulators have been shown to play an important role in regulating individuality in a large range of organisms (8, 32–35), including flies (5, 36, 37). To test whether serotonin affects the rate of drift (and bet-hedging), we fed isogenized Oregon-R flies food supplemented with either the serotonin synthesis inhibitor aMW or the serotonin precursor 5HTP and measured their locomotor handedness for 3 days each week for 3 weeks, as in the previous experiment. While σ_Bet-Hedging was similar in all three treatment groups, both pharmacological manipulations of serotonin decreased the stability of behavior over time, increasing σ_Drift and decreasing φ (Figure 1H, S1F). This was also evident in decreased correlation coefficients between handedness on successive days (Figure S1H-J) compared to control flies.

To assess the role of serotonin as a regulator of drift by a second approach, we generated a constitutive mutation in the tryptophan hydroxylase gene (trh), which is involved in the synthesis of serotonin. This approach also controls for off-target pharmaceutical effects. We used in vivo CRISPR (38) to knock out trh in an isogenized Oregon R background, yielding control flies of REFREF genotype with a closely matched genetic background. Knocking out trh led to an increase in both σ_Bet-Hedging and σ_Drift and decrease in φ compared to controls (Figure 1I, SG), as well as a decrease in correlation between control and trh null flies (Figure S1K-M).

A theoretical adaptive advantage for drift

The variation we see in σ_Drift among genotypes suggests that drift can potentially evolve to provide fitness advantages in some situations. To build an intuition for the ideal drift strategy and explore possible theoretical foundations, we formulated a simple two-state model (Figure 2A). Imagine an organism in an environment that shifts probabilistically between two states. The organism can exhibit two behavioral phenotypes. If the chosen phenotype matches the environment, the organism survives; if not, it dies. What is the optimal strategy for changing its behavior to survive probabilistic shifts in the environment that occur with probability p? We can prove, through methods analogous to classic findings on optimal betting solutions (39, 40) and previous analytical work on generational bet-hedging (41, 42), that in this two-state model, the optimal fraction of the population that should shift their behavioral phenotype in order to maximize long-term population growth equals p. This result extends to both the case of an arbitrary number of possible states as well as a continuous distribution of environmental states and behavioral phenotypes for all cases where the fitness narrowly depends on a tight match to a given environment (see Supplementary Materials for proofs). Interestingly, this holds true even when the fitness associated with successfully matching phenotype to environment A vs differs from matching environment B – in other words, matching the environment as much as possible is the most important thing, independent of the quality of specific environments.

Fig. 2.

A flexible model of drift shows a potential adaptive role under some patterns of environmental change

To model the effect of more realistic environmental fluctuations, we added dynamic environments, fitness effects and life history to the auto-regressive model we used to fit the experimental behavioral data (Figure 2B). In this model, individuals, tracked at the population level, exhibit different behavioral preferences that can potentially vary over their lifespans. The environment fluctuates, and the survival probability of an individual increases as the difference between their preference and the environment decreases. Two parameters determine the phenotypic strategy: σ_Bet-Hedging, captures the initial variability of the population, and σ_Drift, captures the amount each fly’s preference changed per day. By integrating this model over time, we estimate the final population size of populations with varying levels of σ_Bet-Hedging and σ_Drift (Figure 2C-D, S2), thereby determining the optimal variability strategy for a given pattern of environmental fluctuations. Sampling a few randomly generated environments quickly revealed that different patterns of fluctuation favored different strategies (Figure 2E-F), so we began to systematically assess what characteristics of the environment and life history might favor drift, bet-hedging or combinations of both.

Drift is an effective response at shorter timescale fluctuations than bet-hedging

To characterize the adaptive value of bet-hedging and drift strategies as a function of statistics of environmental fluctuations, we created randomized environmental dynamics by filtering white noise in the time domain, leading to random fluctuations with a given temporal frequency content. We normalized and scaled the resulting environmental patterns, and used them in simulations of population survival for a range of bet-hedging (σ_Bet-Hedging) and drift (σ_Drift) values. From these simulations we constructed a fitness landscape over the four dimensions of σ_Bet-Hedging, σ_Drift, environmental fluctuation frequency and environmental fluctuation amplitude. At very high frequency fluctuations the most successful strategy was σ_Bet-Hedging = σ_Drift = 0, reflecting an optimal strategy of all individuals tightly matching the average conditions of the environment (Figure 3C). Similarly, low amplitude fluctuations favored low bet-hedging and low drift. These findings make sense: low amplitude fluctuations are effectively a static environment that does not require a variable phenotypic strategy, and very rapid fluctuations changing within the timescale of organismic response are averaged away (6, 15, 18). However, as the amplitude of the fluctuations increased and the frequency decreased, the optimal strategy showed increased amounts of drift and bet-hedging, as producing individuals with a preference far from the environmental mean became increasingly necessary for survival. While both drift and bet-hedging are preferred over no variability in these situations, higher frequency fluctuations favor drift, while lower frequency fluctuations favor bet-hedging. Increasing the time to sexual maturity increases the advantage of drift, suggesting that varying a phenotype dynamically provides a mechanism for organisms to survive long enough to reproduce. These results suggest that the relative benefit of a drift or bet-hedging strategy depends on the rate of fluctuation of the environment compared to the development time of an organism. These relationships hold for a wide range of β (Figure S3A) and a_min (Figure S3B-C).

Fig. 3.

Real world environmental time series may favor drift

To predict the adaptive value of drift strategies in somewhat less synthetic circumstances, we collected 43 types of environmental time series from publicly available NOAA (43) and NEON (44) datasets across 118,618 sites. We randomly sampled contiguous 1000-day time series from each of these datasets from the past 20 years, and normalized each sample’s mean and standard deviation, scaled by σ_mean to represent different amplitudes of selective pressure. We then used these time series as model inputs to calculate the optimal amounts of drift and bet-hedging as a function of σ_mean and a_min (Figure 4A-B). Different patterns of environmental fluctuation, across differing measurement types and locales, showed differing optimal amounts of drift and bet-hedging. As in our previous analyses, increasing σ_mean increased the optimal σ_Drift and σ_Bet-Hedging (Figure 3), and increasing the a_min increased the optimal magnitude of drift (Figure 4A, S4B,D).

Fig. 4.

To relate optimal strategies of drift and bet-hedging to climate and environmental factors, we used principle components analysis (PCA) to study variation in fitness landscapes over σ_Bet-Hedging and σ_Drift (Figure 4C). The first principle component of variability strategy landscapes had the vast majority of the variance (97.8%) and encoded the extent of bet-hedging; the second component (1.9%) encoded the extent of drift. Across time series, we observed a variety of optimal amounts of drift and bet-hedging. A substantial minority (12%) of environmental signals favored non-zero drift (Figure 4D, E), particularly those reflecting temperature and wind-speed measurements and a subset of locations with tropical, sub-tropical, Mediterranean and continental climates (Figure 4F, G). These results depend on the a_min in our model, with later ages favoring drift (Figure S4).

Discussion

We found that individual behavioral biases are not stable over an animal’s lifetime, and that the degree of stability in behavioral bias is influenced by genetics and neuromodulation. Inspired by these findings, we provide a theoretical rationale for how instability in preferences could be adaptive for the long-term survival of individuals and species.

The behavioral preferences of individual flies change continuously over their life on days-long timescales (Figure 1 A-C; S1A-D). While indications of this phenomenon have been seen previously (3, 5, 24, 45–47), these studies examined a small number of time points or recorded behavior continuously for up to a week. By measuring behavior continuously for up to four weeks (roughly the lifespan of a fly), we were able to estimate the complete power spectrum of behavior: the continuous distribution of timescales at which it changes. Across multiple behaviors (Figure 1C, Figure S1A-D), we show that there is no specific timescale over which behavioral preferences change (e.g., between bouts of movement), and suggests that individuality both quickly and slowly (and on intermediate timescales).

Examining several different inbred strains derived from a natural fly population, we found that the stability of preferences depends on genotype (Figure 1H). This implies that there is natural allelic variation that affects the rate of behavioral drift. In turn, this suggests that drift is not strongly maladaptive, in which case we would expect alleles involved in promoting drift to have been selected out of the population. Genetic variation for drift implies that the rate of drift could evolve to increase the fitness of animals subject to different kinds of environmental fluctuations. However, genetic variation for drift does not, per se, demonstrate that drift is subject to selection; drift could be a neutral epiphenomenon of some other trait under selection.

Importantly, some lines with differing magnitudes of initial preference variability (σ_bet-hedging) exhibited similar amounts of preference stability (σ_drift) (e.g., lines DGRP 819 and 426). This suggests that mechanisms that drive differences in population preferences developmentally incompletely overlap with mechanisms that drive stability in preferences as adults.

Serotonin has been implicated as a regulator of the extent of in individuality in fly behavior (5, 36, 37) and stability of idiosyncratic preferences in C. elegans (33, 48). Bidirectional pharamacological perturbation of serotonin signaling (increasing it with serotonin precursors, or decreasing it with synthesis inhibitors or genetic perturbations) decreases the stability of preferences over time (or conversely increases the amount of drift) (Figures 1H,I, S1F-M). Constitutive mutations in the serotonin synthesis pathway increase both the amount of drift and the amount of variability we observe, suggesting that serotonin has a role in regulating preferences both during and after development.

The combination of genetic and neuromodulatory control suggests that extent of behavioral drift could evolve if it confers a fitness advantage in natural settings. The bethedging framework provides a theoretical basis for the adaptive value of stable idiosyncrasy in behavior(2, 49): namely, that variability in progeny phenotypes increases the chances that at least some offspring are fit when the environment fluctuates unpredictably. We hypothesize that drift may be advantageous for similar reasons, just operating within the lifespan of each individual. We provide two arguments in support of this hypothesis. First, we show in several analytically tractable cases (Fig 2A, Supplemental Materials) that it is evolutionarily optimal for an individual to switch behavioral preference phenotypes with a probability equal to the probability the environment changes.

This analytical result matches what we find using a more biologically realistic computational model of drifting individual preferences in fluctuating environments. We find that drift is adaptive when the environment changes faster than the generation time of an animal. This finding suggests that behaviors that confer fitness with respect to aspects of the environment that change quickly should correspondingly change more quickly. Conversely, stable behavioral preferences may confer fitness with respect to stable aspects of the environment. This is borne out by simulating populations in different environments: quickly changing environmental parameters such as relative humidity (Figures 4A, S4D) favor a drift strategy, while environments that vary at longer (e.g., seasonal) time scales, such as average daily temperature, favor a bet-hedging strategy (Figures 4B, S4D). A key parameter of this model is the age of reproductive maturity (a_min). Drift provides a mechanism for organisms with slow development times to survive fluctuating environments long enough to reproduce. Thus, a key quantity is the time scale of environmental fluctuations relative to maturation time. We predict that organisms with more delayed onsets of reproduction will be more likely to exhibit drift compared to organisms with more rapid development (subject to the same environmental fluctuations).

Organisms employ many strategies to survive changing environments. While our study focused on the respective advantages of bet-hedging (stable variability at the individual level) and drift (variability within an individual’s lifespan), these findings complement previous work comparing bethedging and adaptive tracking (selection-induced changes in allelic frequency over time(50)). Previous work on thermal preference in flies suggested that adaptive tracking offers an advantage when the environment fluctuates on approximately years-long time scales, whereas bet-hedging offers an advantage for months-long fluctuations (6, 30). These time scales roughly correspond to several fly lifespans and one lifespan, respectively. This study finds that drift may be an adaptive strategy for environmental fluctuations within a lifespan. Thus, adaptive tracking, bet-hedging and drift appear to be complementary strategies for the challenges of environmental fluctuation over a wide range of time scales.

Importantly, random phenotypic drift and bet-hedging comes at potentially high costs. If organisms could detect (or predict) environmental fluctuations and deterministically change their phenotype to be optimal for the realized environment, they could avoid the losses of randomly choosing the wrong phenotypes. Nonetheless, drift may play an adaptive role in some systems for two key reasons. Firstly, sensing and predicting future environments may be metabolically costly (19), especially for all possible environments. Random changes in preferences may therefore be more efficient than evolving the ability to reliably adapt to any fluctuation.

Second, random strategies may be inherently optimal. In a game theoretic context, finite games always have an optimal strategy (a Nash equilibrium), but this strategy may have to be random. For instance playing randomly in Rock Paper Scissors is optimal in so far as your opponent cannot learn to predict your play and exploit it. The randomness of stochastic evolutionary strategies, such as bet-hedging and drift, may thus be truly optimal(41, 42). This may particularly likely if the fitness effect of a particular behavioral phenotype depends on fluctuating game-theoretic interactions.

This paper characterizes changes in individual behavior within a fly’s lifetime, and proposes a theoretical framework in which such changes are evolutionarily adaptive. Together, these results motivate continued study of the biological mechanisms underpinning behavioral drift as well as empirical studies to test the hypothesis that behavioral drift helps organisms survive rapidly changing environments.

Materials and methods

See SI Materials and Methods for details. All Raw Data, Data Acquisition Software, and analysis scripts are available at http://lab.debivort.org/drift-in-individual-preference/ and https://zenodo.org/doi/10.5281/zenodo.13698148. Analysis scripts are available at https://github.com/Maloney-Lab/Drift-in-Individual-Preference.git

Fly Care

Flies were raised at room temperature and maintained at room temperature. Flies were fed cornmeal/dextrose medium, as previously described (5). Flies were maintained in communal housing before the first time point (Age 2-5), and then maintained in individual housing to maintain identity at later timepoints. Flies were kept on normal 12:12h light dark cycles.

Behavior

All behavioral experiments were tracked using MARGO, as described previously (51).

Continuous Circling Experiments

For the 24hr continuous circling experiments (Fig 1 A-C, Fig S1A-D), flies were placed in a 28mm radius circular arena filled with standard fly allowing 3mm of space between food floor and lid. Arenas were illuminated with white light from 9am to 9pm.

Handedness

For y-maze experiments, flies were placed in y-mazes as described previously (3, 51). Flies were allowed to roam freely in the y-maze for two hours, and the number of left and right turns were computed for each fly. Y-maze experiments were done illuminated in white light.

Statistics and Modeling

Data were analyzed in Matlab and Python. Bayesian analysis was performed using the STAN programming language (52). All simulations were done in Python.

Real World Data

Data was collected from publicly available sources (NEON(44) and NOAA(43)) for 43 environmental variables across 118,618 sites sampled from the last 20 years.

Acknowledgements

We thank M. Miyagi, S. Lavopulo, S. Lall, D. Lavrentovich, and other members of the de Bivort lab for helpful comments on this manuscript. This work was funded by NIH R01NS121874-01 to BdB and a Harvard Brain Institute Postdoctoral Pioneers grant to RM.