To identify genes whose expression early in life is predictive of future lifespan, we selected 22 integrated miRNA fluorescent reporters (PmiRNA::GFP) from a larger library of 73 transgenic strains (Martinez et al., 2008), using as the only selection criterion the ability to detect reporter fluorescence with short (<100 ms) exposure times at ×5 magnification on our imaging system. Most of the miRNAs corresponding to the selected reporters have no reported function, and in particular, no lifespan phenotype upon knockout (Table 1).

Each reporter was crossed into the temperature-sensitive sterile strain spe-9(hc88) and examined in a high-density single animal culture device previously developed by our lab (Figure 1a; Zhang et al., 2016; Pittman et al., 2017) maintained at 25°C. For each PmiRNA::GFP;spe-9(hc88) strain, we collected bright-field and fluorescence images of each individual every 4 hr from hatch until death. Using an in-house image analysis pipeline, the C. elegans in each bright-field image was automatically identified, defining a mask that separates ‘worm pixels’ from the image background. We converted images of GFP fluorescence, which contain many thousands of pixel intensities within the image region corresponding to a single animal, into a single summary statistic. This allowed us to quantify PmiRNA::GFP expression over time within a single individual (Figure 1b, left; see Materials and methods for details). We then averaged the fluorescence measurements across all individuals to capture the overall population trend in expression of each reporter over time (Figure 1b, right, and 1 c), which is consistent with individual expression trends (Figure 1—figure supplement 2). For some reporters, such as Pmir-788::GFP, expression analysis was limited to shortened time windows, as late in life, reporter fluorescence becomes indistinguishable from background autofluorescence (see Table 1).

We observed several categories of temporal dynamics among the reporters. Expression of some, such as Plet-7::GFP, Pmir-788::GFP, Pmir-79::GFP, Pmir-85::GFP, and Pmir-84::GFP, peak during young adulthood and sharply decline thereafter, in some cases becoming undetectable with age. Expression of others, like Pmir-793::GFP, Pmir-246::GFP, and Pmir-228::GFP, also decrease with age but exhibit a slower decline or plateau after young adulthood. Overall, the majority of PmiRNA::GFP reporters decreased with age, in agreement with previously published microarray (Ibáñez-Ventoso et al., 2006) and small-RNA sequencing studies (de Lencastre et al., 2010; Kato et al., 2011). Unusual for miRNAs, Pmir-51::GFP, Pmir-242::GFP, and Pmir-240–786::GFP show an increase in expression over time (Figure 1c). This increase was also observed with small RNA sequencing (de Lencastre et al., 2010; Kato et al., 2011), indicating that the PmiRNA::GFP reporters generally reflect age-related trends in endogenous transcription of miRNAs. In contrast to sequencing, however, this work provides a more detailed picture of expression trends over time.

In order to visualize the relationship between PmiRNA::GFP expression and longevity (if any), we binned populations of the PmiRNA::GFP reporter strains into cohorts based on eventual lifespan (Figure 2a) and plotted average reporter expression for each lifespan-cohort over time (Figure 2b and d). For a subset of PmiRNA::GFP reporters, the average expression level and rate of change of expression appeared to differ substantially between cohorts in mid- to late- adulthood, indicating that reporter expression may be predictive of an individual’s future lifespan. To test this quantitatively, we performed a multivariate regression on each individual animal’s average level of GFP expression and the trend in GFP expression (positive or negative slope) against future lifespan (Figure 2c and d; Table 2). Because the temporal dynamics of the different PmiRNA::GFP reporters vary substantially, we used a sliding time window to determine the optimal span of expression data to include in the regression for each reporter strain. We examined GFP expression bounded between 3 days post-hatch (dph; approximately the first day of adulthood) and the time of 90% population survival. Note that that the time of 90% survival varied among strains, due to per-strain differences and, to a lesser extent, batch and seasonal effects (Table 2). We performed regressions using GFP levels at all possible windows within the overall bounds, by independently moving beginning and ending timepoints of the window at 12 hr intervals. Individuals that died before or within the chosen window were censored from analysis to avoid truncation effects (these short-lived individuals necessarily have fewer expression measurements, which can confounding analysis). The window that resulted in the maximum correlation between expression and lifespan was selected (Table 2). Similar results obtain using a simple, fixed time-window between 3 days post-hatch and the 90% survival timepoint, or repeating the variable-window analysis bounded between three dph and the 95% survival timepoint (tables S1 and S2 in Supplementary file 1).

Expression of 10 of the 22 tested PmiRNA::GFP reporters exhibited an ability to reproducibly predict lifespan with a joint correlation coefficient (R²) of at least 0.15: the reporters for miRNAs lin-4, mir-47, mir-60, mir-85, mir-90, mir-228, mir-240–786, mir-243, mir-246, and mir-793 (Table 2). The R² value in this case represents the fraction of total inter-individual variation in lifespan that can be accounted for by the inter-individual variation in measured GFP levels of a single reporter.

The ability to predict lifespan does not appear to be a generic property of promoter::GFP constructs, as over half of the examined reporters did not correlate substantially with lifespan. Furthermore, two non-miRNA reporters we tested, Pmyo-2::GFP and Pcpna-2::GFP, were not substantially predictive of lifespan (Figure 2—figure supplement 1 and table S3 in Supplementary file 1). Moreover, the optimal time window for regression (i.e. the time window of expression data producing the maximum correlation coefficient value) varied between reporters, suggesting that the reporters may provide readouts of different phases of the aging process. Some reporters exhibited broad (at least 2 days) optimal time windows spanning young adulthood (Pmir-243::GFP, Pmir-85::GFP) or mid-to-late adulthood (Plin-4::GFP, Pmir-47::GFP, Pmir-228::GFP, Pmir-246::GFP, Pmir-793::GFP). Other reporters like Pmir-60::GFP, Pmir-90::GFP, and Pmir-240–786::GFP were optimally predictive in narrower time windows centered at late adulthood. For the latter set of reporters, the average expression value tended to be more predictive of future lifespan than the slope of an individual’s expression trendline. For the earlier predictors, the slope of the trendline tended to be as or more predictive of lifespan than the the average expression value (Table 2). Neither the maximum expression value an individual achieved nor the timepoint at which maximum expression was reached were substantially predictive of future lifespan for any of the PmiRNA::GFP reporters (table S4 in Supplementary file 1). A positive correlation between Pmir-246::GFP expression and lifespan has been previously published and lin-4 and mir-228 are known to play roles in longevity and the aging process; however, to our knowledge, an association between lifespan and expression of the other miRNAs in unperturbed C. elegans has not been reported (Table 1).

In each case, extended lifespan was correlated with retention of young-adult gene-expression levels and/or trends. Specifically, all reporters except for Pmir-240–786::GFP both decreased in expression throughout aging and correlated positively with future lifespan. Conversely, Pmir-240–786::GFP increases in expression with age (Figure 1) and was negatively correlated with future lifespan. In other words, for reporters where expression levels start high and decrease over time, long lifespan is predicted by high (‘youthful’) expression levels and a delay in their decline. By the same token, for Pmir-240–786::GFP, with low young-adult expression that increases with age, extended periods of ‘youthful’ low expression predict extended lifespan.

We next examined how the lifespan-predictive power of the PmiRNA::GFP reporters changes throughout adulthood. For every 24 hr period post-hatch, we jointly regressed each individual’s average expression and the slope of the expression trend over those 24 hr against that individual’s eventual lifespan (Figure 3a). We plotted the resulting correlation coefficient over time to illustrate how the predictive power of the reporter changes with aging (Figure 3a and b). Expression of predictive PmiRNA::GFP reporters generally begins to correlate with future lifespan when measured at mid to late adulthood, becoming more predictive over time. Comparing the correlation plot with the population survival curve revealed that reporter predictivity peaks simultaneously with the beginning of population die-off, continuing to significantly correlate with lifespan as more individuals in the population perish (Figure 3a and b). However, reporters do not simply correlate with imminent death, as for this analysis we excluded individuals that perished within the 24 hr time window. Moreover, excluding even those that perish within the subsequent 24 or 48 hr does not abolish predictive power for most reporters (Figure 3—figure supplement 1). None of the reporters we examined showed substantial correlation with future lifespan at timepoints less than 5 days post-hatch (roughly young adulthood). This result would be expected if reporter expression levels were identical across individuals early in life, and individuals’ expression levels only diverged later in adulthood. However, we found that the degree of inter-individual variability in reporter levels is largely consistent throughout life (Figure 1—figure supplements 2 and 3). This suggests that only at mid-adulthood do pre-existing inter-individual differences in PmiRNA::GFP expression become coupled to the aging process and future lifespan.

Interestingly, we found that autofluorescence intensity, a phenomenological measure that negatively correlates with lifespan (Pincus et al., 2016), is predictive of future longevity at approximately the same time in adulthood as are most of the PmiRNA::GFP reporters we examined (Figure 3—figure supplement 2). However, PmiRNA::GFP expression and autofluorescence provide largely non-redundant information with respect to future lifespan, based on the near-additivity of correlation coefficients in single vs. joint regression (table S5 in Supplementary file 1). This suggests that although both act as biomarkers of future lifespan, they likely report on distinct aspects of the aging process.

Expression of the PmiRNA::GFP constructs can only correlate with future lifespan if there is some biological process which both (a) determines individual lifespan and (b) directly or indirectly regulates transcription of the GFP reporters. To learn more about the nature of this process, we investigated the genetic requirements for the correlation between lifespan and GFP levels in the two most lifespan-predictive reporters, Pmir-47::GFP and Pmir-243::GFP.

One straightforward possibility is that the PmiRNA::GFP constructs are honest reporters for their endogenous, cognate miRNAs, which determine individual lifespan by variability in their own expression (‘Direct Causation’ model; Figure 4a). Alternately, the endogenous miRNAs themselves may have nothing to do with the true lifespan-determining pathway; in this scenario, reporter expression is a downstream readout of pathways, processes and/or transcription factor activity that affect lifespan but do not require the activity of the endogenous miRNA (‘Bystander Correlation’ model; Figure 4a). Under the former model, activity of the cognate miRNA is a critical causal link to lifespan; absent that miRNA, variability in GFP would be decoupled from variability in lifespan. Under the latter model, loss of the miRNA would not influence the relationship between GFP expression and lifespan.

To distinguish these two possibilities, we tested whether Pmir-47::GFP and Pmir-243::GFP were still predictive of lifespan in a genetic background lacking the cognate endogenous miRNAs (mir-47(gk167) and mir-243(n4759), respectively). As before, we performed regressions on average expression and the slope of the expression trendline in windows between three dph and the 90% survival timepoint; we also measured the lifespan-predictive value of the reporters over time in discrete 24 hr time periods. The correlation of Pmir-47::GFP expression with lifespan was somewhat reduced in the mir-47(gk167) mutant compared to mir-47(+) (R² = 0.22 vs. 0.36) and exhibited a slightly different temporal window for lifespan prediction (7–9 dph vs. 6–8 dph) (Figure 4b). The correlation of Pmir-243::GFP expression with lifespan was actually increased in the mir-243(n4759) mutant compared to mir-243(+) (R² = 0.257 vs. 0.193), and exhibited a different temporal window (6–7 dph vs. 7–8 dph) (Figure 4c). Overall lifespans were unaffected in mir-47(gk167) and mir-243(n4759) compared to strains wild-type at those loci (median lifespan = 10.8 vs. 10.2 days and 8.9 vs. 9.6 days, respectively). Because both reporters correlated positively with lifespan, a short-lived phenotype might be expected if either miR-243 or miR-47 played a functional role in determining lifespan. The lack of substantial lifespan difference further suggests that neither miRNA plays a direct functional role in determining lifespan.

Overall, the correlation of Pmir-47::GFP and Pmir-243::GFP with lifespan was not abolished in the absence of the endogenous miRNAs. This provides evidence for the ‘Bystander Correlation’ model of Figure 4a, implying that lifespan may be largely determined by transcriptional regulators of the PmiRNA::GFP transgenes (and presumably of the endogenous miRNAs as well), or regulatory pathways/processes even further upstream, rather than via the activity of the miRNAs themselves.

As the abundance and/or activity of miR-47 and miR-243 does not appear to directly affect lifespan, Pmir-47::GFP and Pmir-243::GFP must instead serve as markers of other pathways or processes that influence lifespan. We wondered if these might include insulin/insulin-like growth factor (IGF-1) signaling (IIS), a canonical aging pathway in C. elegans that is highly conserved across taxa (Kenyon, 2010). The primary effector of the IIS pathway in C. elegans is the FOXO transcription factor DAF-16, which modulates the expression of a number of downstream genes that influence aging and lifespan (Kenyon, 2010; Murphy et al., 2003). DAF-16 activity is a common genetic requirement for genes, pathways, and processes that have been reported to increase lifespan in C. elegans, including several microRNAs (Pincus et al., 2011; Boehm, 2005; Boulias and Horvitz, 2012). It is therefore plausible that inter-individual variability in insulin signaling or signal-responsiveness leads to variability in DAF-16 activity and thus to inter-individual differences in subsequent lifespan. Indeed, two previously characterized lifespan-predictive gene-expression reporters, Psod-3::GFP and Pmir-71::GFP, both require functional DAF-16 to correlate with future lifespan. This suggests those transgenes predict future lifespan by reporting on inter-individual variability in DAF-16 (and presumably IIS) activity. To test if Pmir-47::GFP and Pmir-243::GFP similarly report on DAF-16 activity, we examined whether GFP levels still correlated with lifespan in the absence of DAF-16 (Figure 5a). Specifically, we crossed each reporter into a daf-16(mu86);spe-9(hc88) background and assayed lifespan-predictive ability in side-by-side experiments with the reporter in the reference background (spe-9(hc88) only).

Regression on average expression and the slope of the expression trendline between 3 dph and the 90% survival timepoint showed that the correlation of Pmir-47::GFP or Pmir-243::GFP expression with lifespan was not substantially suppressed in the absence of daf-16 (bottom left panels of Figure 5b and c, respectively). For Pmir-47::GFP, the R² values vs. lifespan were 0.24 for daf-16(+) and 0.37 for daf-16(-); for Pmir-243::GFP, the R² values were 0.37 for daf-16(+) and 0.18 for daf-16(-). Similarly, plotting the lifespan-predictive value of Pmir-47::GFP or Pmir-243::GFP over time shows that the predictive power of each reporter is not reduced in the daf-16(mu86) background (bottom right of Figure 5b and c, respectively). As expected, the predictive windows are slightly shifted between daf-16(mu86) and wild-type backgrounds due to the short-lived phenotype of daf-16(mu86) (top right of Figure 5b and c). Interestingly, peak levels of Pmir-243::GFP and Pmir-47::GFP were slightly but consistently reduced in the daf-16(mu86) background, suggesting that DAF-16 may play a minor role in the transcriptional regulation of these reporters.

The lifespan-predictive PmiRNA::GFP reporters we identified cannot all provide independent information. Because the correlation coefficients with future lifespan (Table 2) sum to more than one, at least some subset of these reporters must necessarily be at least partially redundant with one another. This raises an obvious question: to what degree are the reporters independent? At one extreme, all 10 reporters might reflect the activity of a single lifespan-determining pathway. Alternately, there may be a small handful of such pathways that are reported on by different sets of the PmiRNA::GFPs.

If two reporters provide information about distinct, independent biological processes, then when both reporters are measured in the same individual, the lifespan estimate using both reporters will be better than when using either individually. More specifically, if the information that each reporter provides about future lifespan is completely independent, then the R² of a joint regression using both measurements to predict future lifespan will be the sum of the R²-values from the two single regressions. Alternatively, if two PmiRNA::GFP transgenes both report on the same process, then expression of one transgene will be correlated with the other, and, equivalently, the joint regression will not be substantially more predictive of lifespan than either single regression alone (Figure 6a).

To begin to count the number of lifespan-determining processes reported on by the PmiRNA::GFPs we identified, we investigated the degree of redundancy among Pmir-47::GFP, Pmir-793::GFP and Pmir-240–786::GFP. We chose these reporters because they are among the strongest predictors of future lifespan, and can be easily spatially resolved because they are primarily expressed in distinct tissues. We thus constructed three dual-reporter strains comprising all pairs of these three transgenes, and for each strain measured GFP levels of both transgene independently. In specific, we manually annotated fluorescence images to delineate tissues in which each transgene was specifically expressed, and extracted our GFP measurements from those regions only, rather than the whole worm (see Materials and methods and Figure 6b–d).

To determine whether each PmiRNA::GFP in the dual-reporter strain provided independent information about longevity, we regressed each reporter individually against future lifespan and compared the resulting R²-values to that produced by multivariate regression of both reporters against lifespan. As a complementary analysis, we also calculated the semipartial correlation coefficient between each single reporter and future lifespan, statistically controlling for the expression of the other reporter. Specifically, the R² from semipartial correlation between reporter A and lifespan, controlling for reporter B, represents the fraction of variation in lifespan that is uniquely associated with variability in reporter A but not reporter B. In the case that reporter A is fully redundant with reporter B, then reporter A will no longer correlate with future lifespan at all after controlling for the expression levels of reporter B (semipartial R² = 0).

In the dual Pmir-47::GFP; Pmir-793::GFP strain, expression of each reporter is individually predictive of future lifespan, as we previously observed (Figure 6b, R² = 0.37 and 0.23 for Pmir-47::GFP and Pmir-793::GFP respectively). We found that Pmir-793::GFP expression correlates strongly with Pmir-47::GFP expression (Figure 6—figure supplement 1), which suggests that they are co-regulated and thus potentially redundant with respect to future lifespan. Indeed, joint regression incorporating both reporters produces a correlation coefficient no greater than regressing against Pmir-47::GFP alone (Figure 6b, R² = 0.38). This suggests that Pmir-793::GFP levels provide no information about future lifespan that is not also reflected in Pmir-47::GFP expression. As a confirmation, the semipartial correlation between Pmir-793::GFP and lifespan, controlling for Pmir-47::GFP, is (as expected) nearly zero (Figure 6b, left plot). In contrast, Pmir-47::GFP expression remains somewhat predictive of lifespan even after controlling for Pmir-793::GFP. Thus, Pmir-47::GFP and Pmir-793::GFP are not independent predictors of future lifespan and likely report on a single lifespan-determining pathway. Because Pmir-47::GFP expression contains strictly more information about lifespan than Pmir-793::GFP, we infer that Pmir-47::GFP likely reports on activity of that pathway at a somewhat more downstream position with respect to lifespan determination, or integrates additional information from a separate pathway (Figure 6a, center and right).

Similar results obtain for Pmir-47::GFP and Pmir-240–786::GFP. Expression of each reporter was individually predictive of lifespan (Figure 6c, R² = 0.46 and 0.15 for Pmir-47::GFP and Pmir-240–786::GFP, respectively), but expression levels of the reporters are correlated (Figure 6—figure supplement 1) and regression on both reporters again results in a correlation coefficient nearly equal to regressing on Pmir-47::GFP alone (Figure 6c, R² = 0.47). Semipartial correlation analysis was similar to the previous case (Figure 6c), indicating that Pmir-240–786::GFP is wholly redundant with Pmir-47::GFP but not vice-versa. We thus infer that, of the reporters we examined, Pmir-47::GFP expression either integrates information most broadly across lifespan-determining pathways, or reports on a step the farthest downstream in a single lifespan-determining pathway.

Unsurprisingly, Pmir-240–786::GFP and Pmir-793::GFP also were correlated and did not act as independent predictors of lifespan (Figure 6d; R² = 0.13 and 0.24 for Pmir-240–786::GFP and Pmir-793::GFP, respectively, with a joint R² = 0.25, which is close to the figure for Pmir-793::GFP alone and much less than 0.37, the sum of the two individual values). Confirmed by semipartial correlation analysis, this suggests that Pmir-240–786::GFP is almost wholly redundant with Pmir-793::GFP (save approximately 4% of lifespan variation which the former transgene can predict independent of Pmir-793::GFP). We infer that Pmir-793::GFP reports on information that is downstream of (or more broadly integrated than) Pmir-240–786::GFP, with respect to lifespan determination.

Overall, our data is consistent with a model wherein Pmir-240–786::GFP, Pmir-793::GFP, and Pmir-47::GFP expression provide hierarchical information about a lifespan-determining pathway or pathways. While there are multiple scenarios consistent with our results, all reflect the informational hierarchy depicted in Figure 6e (left), where Pmir-47::GFP provides strictly more information about lifespan than Pmir-793::GFP, which itself provides strictly more information than Pmir-240–786::GFP. Such a hierarchy is consistent with a scenario in which Pmir-47::GFP provides the most downstream report on a single lifespan-determination process (Figure 6e, center). It is also consistent with a scenario in which Pmir-47::GFP reports on three distinct lifespan-determination pathways, Pmir-793::GFP reports on two of those, and Pmir-240–786::GFP on a single pathway (Figure 6e, right). Last, it could also be that each transgene reports on the exact same step of a single pathway, but that the different transgenes also reflect different degrees of extraneous, unrelated transcriptional input. Each model has in common, however, the key fact that the bulk of the three transgenes’ correlation with lifespan is due to their mutual transcriptional relationship with a single lifespan-determining process.

It is particularly interesting in this context to recall that each of these reporters is expressed in distinct tissues; moreover, the expression trend of Pmir-240–786::GFP (increasing throughout life) is quite different from the other two reporters (which peak at mid-adulthood and decline). Nevertheless, the levels of these three reporters are all correlated (or anticorrelated) and provide largely redundant information about a single lifespan-determining process. Such redundancy and correlations among reporters expressed in distinct tissues could not exist without information flow between tissues, with the most likely candidate being a cell non-autonomous signaling process. We therefore conclude that Pmir-47::GFP, Pmir-793::GFP, and Pmir-240–786::GFP likely report on a global, cell non-autonomous (and DAF-16-independent) process of lifespan determination, which affects the entire animal’s physiology.

The following strains were obtained from the Caenorhabditis Genetics Center (CGC): VT1735 (Pmir-788::GFP), VT1541 (Pmir-360::GFP), VT1733 (Pmir-60::GFP), VL405 (Pmir-63::GFP), VL440 (Pmir-47::GFP), VT1153 (Plet-7::GFP), VT1072 (Plin-4::GFP), VL412 (Pmir-79::GFP), VT1379 (Pmir-59::GFP), VT1474 (Pmir-243::GFP), VT1189 (Pmir-241::GFP), VT1607 (Pmir-246::GFP), VT1485 (Pmir-228::GFP), VT1600 (Pmir-85::GFP), VT1481 (Pmir-51::GFP), VT2020 (Pmir-793::GFP), VT2021 (Pmir-794::GFP), VL370 (Pmir-240–786::GFP), VT1470 (Pmir-242::GFP), VT1160 (Pmir-84::GFP), VT1589 (Pmir-90::GFP), VT1665 (Pmir-1::GFP), PD4793 (Pmyo-2::GFP; Ppes-10::GFP; F22B7.9::GFP), B010652 (Pcpna-2::GFP) CF1038 (daf-16(mu86)), MT15454 (mir-243(n4759)), VC328 (mir-47(gk167)).

All miRNA reporter (PmiRNA::GFP) strains were selected from a larger library originally created by Martinez et al., 2008. Briefly, miRNA promoters were defined as 300 to 2000 base pair regions upstream of the mature miRNA or stem-loop sequence and cloned via Gateway cloning. The GFP and 3′ UTR for the constructs were cloned from the ‘Fire Lab Vector Kit’ construct pPD95.75, containing a GFP with artificial introns and an unc-54 3′ UTR. The transgenic strains were created via microparticle bombardment to generate low-copy random integrants, with UNC-119 rescue used as a selection marker.

All reporter strains were crossed into BA671 (spe-9(hc88)), a temperature-sensitive sterile mutant with a wild-type lifespan at 25.5 °C (Fabian and Johnson, 1994). Strains were maintained at 20°C and for all assays embryos were shifted to 25°C to prevent reproduction in the single-animal culture apparatus.

For longitudinal analysis, C. elegans were reared in high-density single-animal culture devices (‘worm corrals’), which we have described previously (Zhang et al., 2016; Pittman et al., 2017). In brief, 8-armed PEG-thiol (Jenkem Technology) and PEG-diacrylate (Sigma-Aldrich) are dissolved in a modified nematode growth medium (NGM) at concentrations of 133 mg/mL and 37 mg/mL, respectively, and filter-sterilized. The PEGs are mixed in a 1:1 ratio and cholesterol in ethanol is added to a final concentration of 4 µL/mL. The final media is poured into an aluminum frame adhered to a glass slide with PDMS (Dow Corning) and allowed to cure into a solid hydrogel at room temperature for approximately 2 hr. E. coli OP50-1, resuspended to 50% w/v from an overnight culture, is deposited into an array of 0.4 µL droplets onto the gel. A pretzel-stage embryo is placed into each droplet via eyelash pick. Approximately 1 mL liquid PDMS is poured onto the gel; the PDMS crosslinks to the gel and cures into a solid form within 48 hr. The culture devices were housed on a microscope stage (Leica) in a custom-built climate chamber held at 25°C and 90% relative humidity.

Images of individual animals were automatically acquired every 4 hr at ×5 magnification using custom-built image acquisition and autofocus software. A bright-field, autofluorescence, and fluorescence image were taken of each animal at each timepoint. Fluorescence images were taken with a DAPI/FITC/TRITC filter (Semrock, DA/FI/TX-3X-A-000) and Lumencor Spectra X light source and were used to measure accumulation of autofluorescence material and transgene expression. For autofluorescence, an exposure time of 70 ms and 556/20 nm (center wavelength/bandwidth) excitation filter was used; fluorescence in these red emission wavelengths increases linearly with age and has been shown to correlate well with remaining lifespan (Pincus et al., 2016). An exposure time of 30–70 ms was manually chosen for each reporter strain (as strains vary considerably in brightness), with 40 ms most commonly used. Exposure times were consistent for all images within a given strain. A 480/17 nm excitation filter was used for all GFP imaging. Flat-field images of a fluorescent slide (Chroma) were collected before every image acquisition to control for spatial variation in illumination.

All images were corrected for sensor noise (dark-current) and spatial variation in illumination (flat-field). Images were automatically segmented using a convolutional neural network to determine pixels contained within the worm (Ghiasi and Fowlkes, 2016). Whole-animal summary statistics (e.g. 99th percentile intensity) were calculated from the fluorescence intensity distribution of these pixels. All measurements were performed in this fashion unless specifically indicated. Summary statistics are often generated by simply taking the average fluorescence intensity within the worm region. However, because the reporters we examined vary greatly in spatial and temporal expression pattern as well as absolute intensity, we found that no single summary statistic, such as the mean, was appropriate in all cases to robustly capture expression levels of each different reporter, or to distinguish GFP from background autofluorescence. For example, while taking the mean intensity across all ‘worm pixels’ can often reliably measure the fluorescence of a bright PmiRNA::GFP reporter expressed in a large tissue such as the intestine, the same mean may mostly capture intestinal autofluorescence for a reporter that is dim, or expressed only in a smaller tissue like the pharynx or vulva (Figure 1—figure supplement 1). Furthermore, expression of most reporters decreases with time while autofluorescence increases, further challenging robust, automated detection of reporter expression with measurements like mean intensity. Thus, we calculated a number of potential summary statistics, including 95th percentile intensity, the mean of pixels over the 99th percentile intensity, and maximum intensity (see Materials and methods), and for each reporter selected the one that most reliably captured bona-fide PmiRNA::GFP expression in both young and aged animals (Table 2). Our selection was based on manual comparisons of images of reporter expression with maps identifying the set of pixels measured by each summary statistic (i.e. if the pixels at the 95th-percentile intensity level were never co-localized with reporter-gene expression, the 95th percentile intensity was not considered an appropriate summary statistic). Figure 1—figure supplement 1 shows examples of such maps. Among plausible summary statistics, we selected those which produced the least measurement noise timepoint-to-timepoint.

For some strains with dim, spatially restricted fluorescence, measurements were performed on specific structures (e.g. head) rather than whole-worm images to avoid inadvertently measuring autofluorescence. We manually annotated the center of these structures from brightfield images and extracted summary measurements from fluorescence-image pixels within a circle of defined radius from those center-points. For Pmir-1::GFP, a circle with a radius of 5 pixels (6.5 µm) centered at the pharynx was used. For Pmir-241::GFP, a circle with a radius of 25 pixels (32.5 µm) centered at the pharynx was used. For Pmir-59::GFP, a circle with a radius of 25 pixels (32.5 µm) centered at the vulva was used to extract expression measurements.

For dual reporter strains, Pmir-47::GFP and Pmir-793::GFP expression was extracted from manually annotated circular regions centered at the vulva and pharynx. These are the predominant areas of detectable expression, especially in aged animals. Pmir-240–786::GFP expression was measured from areas of the worm excluding the manually annotated region for Pmir-47::GFP or Pmir-793::GFP. Varying the radius of the circular region did not substantially affect correlation with lifespan (data not shown).

Time of hatch, first egg lay, and death were annotated manually for all animals to determine lifespan. Individuals that hatched prior to the start of the experiment were annotated as hatching at t = 0; this timepoint is within at most 3 hr the actual hatch time (Zhang et al., 2016). No more than 5% of eggs in a given experiment fail to hatch in our system; often all hatch successfully. Of those that do hatch,<5% (typically 1–2%) are excluded from further imaging and analysis due to production of offspring (the majority of the excluded individuals) or larval arrest (a small minority). In a minority of cases, additional individuals were excluded due to localized failure of the ‘worm corral’ device: mold contamination and/or gel desiccation. The remaining individuals were indistinguishable from those in other biological replicates.

All fluorescence measures made for every animal used in all figures and tables (and associated life-stage annotations) are provided in easily parsed text files as Source data 1.

For visualizing population-level or cohort-level expression over time, each biological replicate was rescaled by median expression intensity and median lifespan before pooling. Expression data were fit to trendlines using LOWESS smoothing to facilitate visualization of expression levels over time. For regression and other correlative analyses, all expression values were z-transformed relative to each biological replicate and time rescaled by median lifespan of the replicate to facilitate data pooling and prevent replicate-dependent effects. Regression results from raw data from each biological replicate are consistent with the pooled analyses (table S6 in Supplementary file 1).

Single and multivariable regression of biomarkers versus lifespan was performed using ordinary least-squares regression, and the coefficient of determination (R²) calculated according to the standard formula. We aimed for a minimum sample size of 80 individuals, which leads to a statistical power of 95% to identify a true correlation of R² >0.15 with p<0.05. We have found empirically that below R² = 0.15, correlations with future lifespan are not generally robust across experimental replicates. Multivariable regressions for PmiRNA::GFP reporters were performed using both average expression and the slope of a least-squares linear fit line to an individual’s expression data points in a defined time window. Confidence intervals for correlation coefficients were determined by bootstrap analysis. Each pooled reporter sample was resampled with replacement 1000 times and regressions were performed on the resampled data. The 2.5th and 97.5th percentiles of the resampled correlation coefficients were used to construct the 95% confidence interval.

Individuals were excluded from analysis if they did not survive through the entire time window under consideration. Unless otherwise noted, time windows were chosen such that 10% or less of the total population was excluded.