To study how stressful environments early in life influence long-term behavioral patterns across and within all developmental stages we continuously tracked the behavior of individual animals following a transient period of starvation early in development. Imaging was performed using a custom-made multi-camera imaging system across the full developmental trajectory of C. elegans individuals (55 hr), at high spatiotemporal resolution (3 fps, ~10 µm) and in a tightly controlled environment (Stern et al., 2017). First larval stage (L1) animals that hatch into an environment that completely lacks a food source do not grow and their development is arrested (Greenwald and Horvitz, 1982; Johnson et al., 1984; Baugh, 2013). Following L1 arrest, when animals encounter food, they resume their normal developmental trajectory to reach adulthood. This early-stress paradigm allows us to maintain a homogeneous stress environment across individuals at their earliest stage of development, immediately after hatching.

We continuously monitored single N2 wild-type individuals grown in isolation from their first larval stage to 16 hr of adulthood (n = 456) on defined concentrations of UV-killed OP50 bacteria, following periods of stress ranging from 1 to 4 days of early starvation (Figure 1A, B). In parallel, we tracked the behavior of individuals grown continuously on food, without experiencing starvation (Figure 1A, B). Animals exposed to early starvation required more time to complete their development (Figure 1—figure supplement 1A). To align developmental trajectories of different individuals in time, we age-normalized individuals by dividing each developmental stage, detected by the lethargus period during molting (Cassada and Russell, 1975), into 75 time windows (Figure 1—figure supplement 1C; Stern et al., 2017). Whilst growing in a food environment, C. elegans shifts between two behavioral states called roaming and dwelling that last seconds to minutes (Ben Arous et al., 2009; Flavell et al., 2013; Fujiwara et al., 2002; Stern et al., 2017). During a roaming episode animals explore a large area by high-speed forward movements, while in the dwelling episode they show dramatically less exploration due to low-speed movements coupled with frequent reorientations (Figure 1—figure supplement 1B). We quantified long-term patterns of locomotory behavior shown by individuals throughout development by measuring two behavioral parameters: fraction of time spent roaming, and speed during roaming episodes.

Unstressed individuals hatching in a food environment show dynamic behavioral structures of roaming activity across development, as was previously shown (Stern et al., 2017; Figure 1B–D). We found that a transient exposure to early-life starvation generates alterations in long-term behavioral patterns throughout development that were distinct and discontinuous across and within developmental stages. A short early-starvation period of 1 day strongly decreased average roaming activity levels during the L1 and adult stages compared to unstressed individuals (Figure 1C, E; Figure 1—figure supplement 1D). In contrast, while 1 day of early starvation modified within-stage temporal behavioral structures by shifting roaming activity peaks to later time windows during the L2 and L3 stages (Figure 1C; Figure 1—figure supplement 1E), average roaming activity was not decreased during these stages (Figure 1E). Similarly, we found that average roaming activity level was also maintained during the L4 stage, following 1-day starvation (Figure 1C, E; Figure 1—figure supplement 1D). Interestingly, animals exposed to longer starvation periods of 3 and 4 days further showed strong roaming decrease during L1 and adulthood, but exhibited only minor effects on average roaming activity within the intermediate L2, L3, and L4 larval stages (Figure 1D, E; Figure 1—figure supplement 1D). Furthermore, while the average size of animals that experienced early starvation was slightly decreased (~10%), a comparison of starved and unstarved individuals within the same size range (size-matched) showed similar stage-specific effects on roaming activity during L1 and adulthood following early starvation (Figure 1—figure supplement 1F). These results indicate that, while the memory of early starvation is maintained throughout development to expose strong decrease in average roaming behavior during L1 and adulthood, behavioral effects are buffered across intermediate development times.

Similar to the stage-specific effects of early starvation on the fraction of time spent roaming, instantaneous speed during roaming episodes in individuals exposed to early starvation was affected more strongly at the L1 and adult stages, compared to the intermediate stages (Figure 1—figure supplement 1G). In summary, transient early-life starvation discontinuously reshapes long-term behavioral patterns across development time by exposing strong behavioral alterations at early and late developmental stages and buffered effects during intermediate stages.

The longitudinal measurements in single animals across development allow us to further quantify long-term inter-individual diversity within stressed and unstressed populations. Following early starvation, different individuals show substantial variation in long-term behavioral responses. For instance, during L1 and adulthood, a fraction of wild-type individuals that were exposed to early stress show 8- to 10-fold decrease in roaming activity relative to the average roaming level of the unstressed population, while other stressed individuals show roaming activity which is indistinguishable from unstressed animals (Figure 1—figure supplement 1D). Behavioral individuality is usually defined as a consistent tendency of an individual to show the same behavioral bias relative to the population across long time-periods (Bierbach et al., 2017; Buchanan et al., 2015; Kain et al., 2012; Stern et al., 2017; Schuett et al., 2011). However, individuals may also show alternative patterns of temporal behavioral biases relative to the population that are not random and represent more complex structures of individual biases over time. Here, we extend the ‘classic’ analysis of individuality and ask if alternative individuality dimensions coexist within C. elegans populations across development.

To analyze long-term individual biases in behavior we first systematically rank individuals based on their roaming activity compared to all other individuals within the same experiment across developmental windows (50 time bins) (Figure 2A, B). The rank approach allows us to homogeneously compare between individuals at each developmental window. To take an unsupervised approach for detecting temporal patterns of individual biases that are dominant within stressed and unstressed populations we performed principal component analysis (PCA) of the temporal behavioral ranks of all wild-type individuals (n = 456). Following PCA, each individual within the population is represented by its score (value) in each of the PC dimensions. We identified statistically significant PC dimensions by comparing the variances in PC scores within each dimension to those obtained from PCA of a randomly shuffled rank dataset (Figure 2C; Figure 2—figure supplement 1A) or to score variances of a shuffled rank dataset within the same PCA space (Figure 2—figure supplement 1B).

We found that among the significant PC dimensions (Figure 2D–F; Figure 2—figure supplement 1F), the three major PCs (PC1–3) captured three distinct dimensions of temporal individuality patterns within stressed and unstressed populations (Figure 2D–F). PC1, which explained the majority of temporal variation in individual biases over time had eigenvector components of the same sign, indicating an individuality dimension of animals that consistently roam more or less than the population homogeneously across all developmental stages (Figure 2D). The individuality dimension identified by PC1 unbiasedly recaptured a known mode of consistent individuality that was previously identified using a pre-defined index of long-term behavioral consistency across development (Stern et al., 2017; Figure 2—figure supplement 1C, D). This was further verified by the high correlation between the pre-defined consistency index and scores of PC1 across individuals (R = 0.9) (Figure 2—figure supplement 1E). Interestingly, other PC dimensions identified uncharacterized individuality patterns. PC2, which had opposite signs of eigenvector components before and after mid-development captured an individuality dimension that includes individuals that switch their behavioral bias once, during the L3 stage, from roaming more to roaming less than the population and vice versa (Figure 2E). In addition, PC3, which had signs of eigenvector components that switch twice during development (at the end of L1 and L4), identified individuals that show the same behavioral bias during L1 and adulthood, which is opposite to their behavioral bias during intermediate stages (Figure 2F). Other significant PC dimensions showed more complex dynamics of temporal individual biases across development, displaying multiple bias switching within developmental stages (Figure 2—figure supplement 1F). PC individual scores in these alternative PC dimensions did not correlate with the pre-defined consistency index (R = 0.003–0.09) (Figure 2—figure supplement 1E), further indicating that they indeed represent uncharacterized modes of temporal individuality.

Inter-individual variation in PC scores within a specific PC dimension reflects how extreme individuals are within a population toward the identified individuality dimension. We found that wild-type populations with different early-life experiences show extreme inter-individual variation in multiple PC dimensions, compared to a randomly shuffled rank dataset within the same PCA space or within a PCA space generated from the shuffled rank dataset (Figure 2G–J; Figure 2—figure supplement 1G, H), indicating the coexistence of these alternative individuality dimensions. Altogether, these results demonstrate the use of unsupervised analysis for identifying multiple individuality dimensions across development and suggest a broad individuality space within isogenic populations.

Neuromodulatory pathways are known to establish internal behavioral states and modify them based on the environmental context (Harris-Warrick and Marder, 1991; Bargmann, 2012; Marder, 2012; Kennedy et al., 2014; Taghert and Nitabach, 2012; Nusbaum and Blitz, 2012). In particular, the bioamine dopamine was implicated in controlling a wide array of behavioral outputs at various timescales, ranging from minutes and hours, to long-term behavioral patterns that are regulated across life stages (Marella et al., 2012; Omura et al., 2012; Sawin et al., 2000; Cermak et al., 2020; Stern et al., 2017). In C. elegans, dopamine is produced in a specific set of neuronal sites and its effects are known to be mediated by dopamine receptors that are localized to responding neurons (Sulston et al., 1975; Lints and Emmons, 1999; Chase et al., 2004; Tsalik et al., 2003).

To ask if dopamine acts across different developmental stages to shape the discontinuous pattern of long-term behavioral responses to early stress and to dissect its temporal requirement, we tracked the behavior of dopamine-deficient cat-2 animals following exposure to L1 starvation (Figure 3A; Figure 3—figure supplement 1A). When continuously grown in a food environment, cat-2 individuals show a long-term roaming activity pattern that is similar to the wild-type population (Stern et al., 2017; Figure 1; Figure 3). However, we found that in contrast to stressed wild-type individuals that show buffering of behavioral responses during the L2, L3, and L4 intermediate stages, cat-2 individuals that were exposed to early starvation show reduction in average roaming activity across all developmental stages, including during the intermediate stages (Figure 3B, C). The behavioral effects of early starvation during mid-development in cat-2 individuals were not only restricted to animals that were exposed to long starvation periods, as 1 day of early starvation was sufficient to induce a strong reduction in roaming activity during the L2–L4 intermediate stages (Figure 3D; Figure 3—figure supplement 1B–D). Interestingly, during the L2 intermediate stage the effects on roaming activity patterns were more pronounced during earlier time windows of the stage, suggesting a potential within-stage regulation of behavioral response by dopamine (Figure 3B, C; Figure 3—figure supplement 1C). However, behavioral effects during L1 and adulthood following early stress were similar in cat-2 and wild-type (Figure 3B–D), implying that dopamine function is mainly required during intermediate developmental stages to buffer alterations in roaming activity in response to a transient early stress.

It was previously shown that during L2 to adulthood, cat-2 animals have higher instantaneous speed during roaming episodes (Stern et al., 2017; Sawin et al., 2000; Figure 3—figure supplement 1E). We found that unlike stressed wild-type individuals in which roaming speed was decreased mainly during L1 and adulthood (Figure 1—figure supplement 1G), cat-2 mutants show lower speed also across the L2 and L3 stages following stress (Figure 3—figure supplement 1F).

To further ask if dopamine supplementation can restore roaming activity following stress, during intermediate developmental stages, we exposed cat-2 individuals to exogenous dopamine (Figure 4A–C; Figure 4—figure supplement 1A–E). We found that supplementing dopamine was sufficient to increase only the roaming activity following stress in cat-2 individuals exposed to 1 and 3 days of starvation, during the L2, L3, and L4 stages (Figure 4A–C; Figure 4—figure supplement 1B, C, F, G). In contrast, following 1 day of starvation exogenous dopamine did not restore roaming activity during L1 and adulthood (Figure 4A, C; Figure 4—figure supplement 1A, B) and following 3 days of starvation it only slightly increased roaming in the L1 stage and did not affect roaming activity during adulthood (Figure 4B, C; Figure 4—figure supplement 1A, C). Overall, these results show that following early and transient starvation, dopamine acts to restrict long-term behavioral alterations in roaming activity, specifically during intermediate developmental windows.

The buffering of behavioral effects during the L2, L3, and L4 intermediate developmental stages by dopamine led us to explore the temporal contribution of specific dopamine receptors during these development times. The C. elegans dopamine receptor DOP-1 is a D1-like receptor which signal through Gα_s/olf to activate adenylyl cyclase and DOP-2 and DOP-3 receptors are D2-like receptors which signal via Gα_i to suppress adenylyl cyclase (Chase et al., 2004; Sanyal et al., 2004; Sugiura et al., 2005; Suo et al., 2003).

To study the independent function of dopamine receptors we analyzed the long-term behavioral effects of early starvation in animals mutant for each of the single dopamine receptors. These analyses showed that each receptor has a different temporal effect on behavioral responses within the intermediate L2–L4 stages (Figure 4D–I; Figure 4—figure supplement 2). In particular, following 3 days of early starvation, dop-2 animals showed strong roaming decrease during the L2 and L4 stages, but not during the L3 stage, compared to wild-type (Figure 4D, G; Figure 4—figure supplement 2). Interestingly, similar to cat-2 mutants, roaming was mainly decreased in dop-2 individuals during early time windows within the L2 stage. In addition, dop-1 individuals showed a roaming decrease during the L2 stage and opposite effects during L3 and L4 (Figure 4E, H; Figure 4—figure supplement 2) and dop-3 animals showed weaker overall roaming response during the L2 stage (Figure 4F, I; Figure 4—figure supplement 2).

Previously, DOP-2 and DOP-3 were shown to function cooperatively (Cermak et al., 2020; Suo et al., 2009). Therefore, we sought to test if simultaneous alteration of both dopamine receptors will recapitulate the full long-term behavioral effect during intermediate developmental stages, as shown in cat-2 mutants. We found that following early starvation, dop-2;dop-3 double mutants showed a decreased roaming activity across all intermediate stages (Figure 4J, K; Figure 4—figure supplement 3). These results imply that a deficiency in both DOP-2 and DOP-3 receptors is sufficient to recapitulate the behavioral effects in dopamine-deficient individuals during mid-development and suggest that buffering of long-term behavioral responses by dopamine is temporally regulated by the modular function of specific dopamine receptors.

To ask if the stage-specific effects of early-life stress on developmental patterns of behavior are an integration of multiple temporal responses that are mediated by different neuromodulators, we also examined serotonin function in shaping long-term behavior following early starvation (Figure 5A; Figure 5—figure supplement 1A). Under normal growth conditions, serotonin-deficient tph-1 individuals roam more than wild-type across all developmental stages (Flavell et al., 2013; Stern et al., 2017; Figure 5—figure supplement 1B). We found that contrary to the effects of dopamine on the buffering of behavioral responses during intermediate stages, tph-1 individuals that were exposed to 1 day of early starvation maintained their roaming activity during L1 and adulthood (Figure 5B, D; Figure 5—figure supplement 1C), compared to the strong roaming decrease generated in the wild-type population during these early and late stages. In addition, no significant difference in roaming response was shown during the L2–L4 intermediate developmental stages in the tph-1 population following 1 day of early starvation (Figure 5B, D; Figure 5—figure supplement 1C), compared to wild-type.

To test if longer starvation periods early in life will establish behavioral effects during L1 and adulthood we further exposed tph-1 animals to 3- and 4 days of early starvation. We found that long starvation periods led to a reduction in roaming activity in the L1 stage of tph-1 animals. However, during the adult stage, tph-1 individuals were still less responsive to early stress, compared to the strong decrease in roaming in the wild-type (Figure 5C, D; Figure 5—figure supplement 1C, D). In addition, behavioral responses to early stress were similar in tph-1 and wild-type individuals during the intermediate L2–L4 stages, indicating that serotonin effects are specific to shaping behavioral responses during L1 and adulthood.

These results show that dopamine and serotonin functions are opposite and segregated across developmental stages in regulating long-term roaming behavior following stress. While dopamine buffers behavioral modifications during intermediate stages of development, serotonin functions to promote behavioral sensitivity to early starvation during the early L1 stage and adulthood. Interestingly, functional segregation among dopamine and serotonin regulation is behavior specific, as the long-term effects of early stress on roaming speed were similar in tph-1 and cat-2 individuals (Figure 5—figure supplement 1E; Figure 3—figure supplement 1F).

To ask if early-life experience and neuromodulatory pathways affect specific individuality dimensions to reshape inter-individual variation within populations, we performed the PCA on pooled data across the wild-type, cat-2, and tph-1 populations (Figure 6—figure supplement 1A–C, E). We then directly compared inter-individual variation in PC scores within specific individuality dimensions between stressed and unstressed populations of wild-type and neuromodulatory mutant individuals. We found that early starvation modified inter-individual variation in specific individuality dimensions and that wild-type and neuromodulatory mutant populations showed both shared and unique effects on inter-individual variation following the same stressful condition (Figure 6; Figure 6—figure supplement 1F–H). In particular, both the wild-type and dopamine-deficient cat-2 populations showed an increase in inter-individual variation in PC scores within the PC3 individuality dimension (double bias switching across development) following starvation, compared to the unstressed population (Figure 6B, D, F). In contrast, inter-individual variation in scores within the PC3 dimension was not significantly altered in serotonin-deficient tph-1 individuals following the same stressful experiences (Figure 6B, E). Furthermore, we found that inter-individual variation in scores within the PC6 individuality dimension which captured multiple bias switching of individuals within developmental stages (L2-Adulthood) was strongly increased following all stressful conditions in the wild-type population, but was decreased after 3 days of early starvation in the tph-1 population (Figure 6C–E). While inter-individual variation in PC6 was affected in opposite directions in the wild-type and tph-1 populations following stress, cat-2 individuals did not show altered PC6 inter-individual variation after early-life stress (Figure 6C, F). In addition, it was previously shown that unstarved tph-1 individuals show low levels of behavioral consistency across development compared to the wild-type population, as quantified by the pre-defined consistency index (Stern et al., 2017). Similarly, we found that unstarved tph-1 individuals show lower levels of inter-individual variation in PC scores within the PC1 dimension, which represents individuals with homogenous consistent bias across development, compared to the wild-type population (Figure 6—figure supplement 1D; Figure 6E). Interestingly, we found that 3- and 4-day starved tph-1 individuals showed an increase in inter-individual variation in scores within the PC1 individuality dimension, compared to unstressed individuals, while in the wild-type and cat-2 populations there was no significant change following starvation (Figure 6A, D–F). Early starvation affected only a fraction of the identified individuality dimensions as inter-individual variation in scores was not significantly altered following early stress within the PC2, PC4, and PC5 individuality dimensions in all wild-type and neuromodulatory mutant populations (Figure 6—figure supplement 1F–H). Overall, these results imply that inter-individual variation in a spectrum of individuality dimensions may be dynamically structured by the early experience of the population and be further modified by its neuromodulatory state.

C. elegans worms were maintained on NGM agar plates, supplemented with E. coli OP50 bacteria as a food source. For behavioral tracking, we imaged single individuals grown in custom-made laser-cut multi-well plates. Each well (10 mm diameter) was seeded with a specified amount of OP50 bacteria (10 µL of 1.5 OD) that was UV-killed before the experiment to prevent bacterial growth. For the starvation experiments, eggs were collected from isogenic populations using a standard bleaching protocol, into an agar plate without OP50 bacteria. Newly hatched L1 larvae were starved for a specified time window (L1 arrest of 1, 3, or 4 days) before being transferred to the imaging multi-well plates. For tracking behavior without early L1 starvation, animals were monitored immediately after hatching in the multi-well plates. For DA supplementation experiments, animals were exposed during starvation to 15 mM DA for 1 day before transferring them to OP50 bacterial food supplemented with 600 mM DA.

Strains used in this study:

Longitudinal behavioral imaging was performed using custom-made imaging systems. Each imaging system consists of an array of six 12 MP USB3 cameras (Pointgrey, Flea3) and 35-mm high-resolution objectives (Edmund optics) mounted on optical construction rails (Thorlabs). Each camera images up to six wells, each containing an individual grown in isolation. Movies are captured at 3 fps with a spatial resolution of ∼9.5 µm. For uniform illumination of the imaging plates we used identical LED backlights (Metaphase Technologies) and polarization sheets. To tightly control the environmental parameters during the experiment, imaging was conducted inside a custom-made environmental chamber in which temperature was controlled using a Peltier element (TE technologies, temperature fluctuations in the range of 22.5 ± 0.1°C). Humidity was held in the range of 50 ± 5% with a sterile water reservoir and outside illumination was blocked, keeping the internal LED backlights as the only illumination source. Movies from the cameras were captured using commercial software (FlyCapture, Pointgrey) and saved on two computers (3 cameras per computer; each computer has at least 8-core Intel i7/i9 processor and 64 GB RAM).

To extract behavioral trajectories of animals across the experiment, captured movies were analyzed by custom-made script programmed in MATLAB (Mathworks, version 2019b) (Stern et al., 2017). In each frame of the movie and for each behavioral arena, the worm is automatically detected as a moving object by background subtraction, and its XY position is logged (center of mass). In each experiment, 600,000–1,000,000 frames per individual are analyzed using ~50 processor cores in parallel to reconstruct the full behavioral trajectory of individuals over days of measurements across development. The total time of image processing was 3–7 days per experiment. Egg hatching time of each individual in the experiment is automatically marked by the time when activity can be detected in the behavioral arena. The middle of the lethargus periods, in which animals stop their locomotion and molt, were defined as the transition points between different stages of development (based on 10 s timescale speed trajectories over time, smoothed over 300 frames). To synchronize temporal behavioral trajectories of different individuals we age-normalized individuals by dividing the behavioral trajectory of each life stage into a fixed number of time windows.

For each individual, we differentiate between roaming and dwelling states by averaging speed (μm/s) and angular velocity (absolute deg/s) over 10 s using a rolling time window, and generating a 2D probability distribution of these two behavioral parameters for all intervals in each time bin along the experiment (50 × 50 bins distribution, speed bin size: 7.59 µm/s, angular velocity bin size: 3.6 deg/s) (Stern et al., 2017). Drawing a diagonal through the probability distribution separated roaming and dwelling states, such that intervals in the distribution bins below the diagonal were classified as roaming intervals and intervals in bins above the diagonal were classified as dwelling intervals (Ben Arous et al., 2009; Flavell et al., 2013; Stern et al., 2017). The behavior of each animal over time could be quantified as a sequence of roaming and dwelling intervals. The fraction of time spent roaming of the individual in a time bin represents the fraction of these intervals classified as roaming states within a given time bin. For each developmental stage, we examined the two-dimensional probability distribution of the whole population and changed the slope of the diagonal to classify roaming and dwelling appropriately (slopes: 5, 2.5, 2.3, 2, and 1.5 for the L1, L2, L3, L4, and adult stages, respectively). Based on these roaming and dwelling classifications we further quantified the average instantaneous speed of the animal during roaming episodes (μm/s).

To measure the worm’s size in each frame, a cropped image of size 151 by 151 pixels around the detected center of mass was used. First, background subtraction was performed in each frame, using the same method as in Stern et al., 2017: The typical background was estimated in each input video (approx. 11.5 min) by averaging 8 equally spaced sample frames. Each frame f in the ith video, with grayscale levels represented as 8-bit values in the range 0–255, was normalized as f~=(f-bi+8+100)/256, where bi+8 is the background estimate for the (i + 8)th video. The worm’s contour was then found in each background-subtracted frame using a fixed grayscale threshold of 0.34, and the number of pixels enclosed by the contour was computed.

For a more robust estimate, a running median was applied with a window length of 10 min (301 frames). Specifically, the area estimate in the ith frame was obtained as the median of raw pixel counts in frames i-150 to i+150, disregarding any frames within this range where worm detection had failed or where contour computation produced no closed contour or multiple closed contours. This step helps smooth out temporary changes to the worm’s apparent size due to imaging noise or changes in posture, as well as errors due to the worm being partially outside the imaging area. For comparing size-matched individuals we quantified statistical difference in roaming parameters within 20 running size windows (width: 10% of range).