NHANES data from 2003–2004 and 2005–2006 were used in this study: https://wwwn.cdc.gov/nchs/nhanes/. Accelerometer data was available in the PAXRAW_C and PAXRAW_D files, and the demographic data was available in DEMO_C and DEMO_D files. The nutritional data was accessed using the DR1TOT_D file and use of prescription drugs were accessed with the RXQ_RX_D file. For use in model building, the 2003–2004 accelerometer data was downloaded from the NHANES repository, filtered to ensure that data was present for each minute of the seven days, devices were calibrated, daily data contained sufficient accelerometer entries (10% data >0), and participants were aged 18 and older. This resulted in high quality dataset of 7 days of accelerometer data for 2,634 adults. For use in model validation, the 2005–2006 NHANES accelerometer data was preprocessed identically and resulted in data for 2,505 adults. Linked mortality data on the participants from NHANES was accessed at the link below. See supplementary materials for extended methods on data access and processing.

ftp://ftp.cdc.gov/pub/Health_Statistics/NCHS/datalinkage/linked_mortality/

The final 2003–2004 accelerometer data was split into training (70%) and testing (30%) datasets using the caret package in R (Kuhn et al., 2016). A random forest model was generated on centered and scaled data from the training dataset using the randomForest package in R (Breiman and Cutler, 2018). Model parameters were assessed using the randomForest and randomForestExplainer packages in R (Breiman and Cutler, 2018; Paluszynska et al., 2020). The model was validated using the 2005–2006 validation dataset. Finally, to ensure a linear relation between predicted and chronological age, an individual’s predicted age was normalized by dividing by the median predicted ages of individuals of similar chronological ages and multiplying again by the individual’s actual chronological age. Age acceleration and deceleration was then evaluated by subtracting an individual’s chronological age from their normalized biological age prediction. See supplementary materials for extended methods.

For correlating nutritional intake to biological age, the 2005–2006 nutrition survey data was downloaded from the NHANES repository and correlations and significant association to deltaAges was assessed using Pearson’s product moment correlation coefficient of binned age groups. Results were clustered using the hclust function in R based on Euclidean distance. Comparisons were further performed between either very high (>10 years) or very low (<10 years) biological age differences (deltaAge) using Student’s t-test. See supplementary materials for extended methods. To screen for drugs that are associated to lower biological aging, the 2005–2006 prescription medication data was downloaded from the NHANES repository and individuals of an advanced age (70–85+). For each drug, the deltaAges of users were compared to the deltaAges of all individuals to identify a significant difference using the non-parametric Kolmogorov-Smirnov test. p values were corrected for using the Benjamini and Hochberg method. See supplementary materials for extended methods.

The R programming environment was used for all data processing steps and statistics in this study (R Core Team, 2013). For assessing the random forest model, RMSE was calculated using the Metrics package (Hammer et al., 2018). Pearson’s product moment correlation was used to assess the random forest model and nutritional impact on age acceleration/deceleration. Aging acceleration/deceleration was calculated considering the predicted age minus the normalized biological age. Comparisons of the ratios of accelerated, normal, or decelerated aging relative to mortality was performed using a 3-sample test for equality of proportions without continuity correction. Students t-test was used for two sampled comparisons. The non-parametric test of Kolmogorov-Smirnov was used for comparison of deltaAges of drug users. Throughout this study, p-values less than 0.05 were considered significant. When multiple comparisons were performed in large number multiple hypothesis testing corrections were applied to p-values using the Benjamini and Hochberg method.

N2 Bristol C. elegans were obtained and maintained as previously described (Liu et al., 2020). For mobility crawling measurements, doxazosin mesylate was dissolved in DMSO and added to plates before pouring at a concentration of 33 µM. Worms were synchronized and grown to day 9 or 10 of adulthood. Crawling speed was measured and analyzed as previously described (McIntyre et al., 2021). See supplementary materials for extended methods. Statistical analysis compared treated and untreated conditions using a Mann-Whitney U test. Mobility assays were performed at least twice, one of which is represented in the data shown. Statistics for mobility experiments and replicates are represented in Supplementary Table 3.

Lifespan and motility assays were prepared and analyzed on a microfluidic chip (Infinity Chips, NemaLife Inc., TX, United States) and Infinity Code Software as previously described (Rahman et al., 2020). Doxazosin was tested at 3.3 and 33 μM, as drugs added in microfluidic systems have been shown to act more potently than drugs mixed in agar (Hewitt et al., 2018). Each microfluidics assay was conducted in triplicate (three biological replicates), and each biological replicate consisted of two technical replicates. One technical replicate is a population of ∼60 animals in a microfluidic growth chamber. See supplementary materials for extended methods. Kaplan-Meier curves from the lifespan assays were generated using GraphPad Prism. Log-rank testing was used to compare the survival curves between the non-exposed control and doxazosin-treated populations. Statistics for all lifespan experiments and replicates are represented in Supplementary Table 4. Statistical comparisons of motility were performed in GraphPad Prism using two-way ANOVA. Statistics for motility experiments and replicates are represented in Supplementary Table 5.

To obtain datasets containing both in-depth characterization of demographic information from individuals and also continuous movement data from a wearable device, we turned to the NHANES data repositories (https://wwwn.cdc.gov/nchs/nhanes/). Data from a single individual can include what drugs they have taken, foods eaten, along with demographic data including age. For earlier survey periods of NHANES, there is also well-documented mortality information for the individuals. The NHANES study periods of 2003–2004 and 2005–2006 are particularly interesting in this regard, as these two study rounds additionally requested individuals to wear a device around their upper thigh measuring their activity levels for a week (ActiGraph AM-7164). Accordingly, we downloaded the accelerometer and demographic data from these study periods, which consisted of several thousands of individuals for each respective study year. We used the 2003–2004 data for model building and testing, then the 2005–2006 data for external validation and exploration of possible effectors influencing the biological aging rate (Figure 1B).

The accelerometer data consisted of a seven-day measurement period with daily entries, and we filtered the data to include only individuals with calibrated instruments that included data for the whole seven days of the study period. Together, this resulted in high quality datasets of 2,634 and 2,505 adults for the 2003–2004 and 2005–2006 datasets, respectively, which covered a broad range of ages from 18 to 85+ (Figure 1C). These data included intensity values at the resolution of minutes for differently aged individuals (Figure 1D), so we aimed to focus the dataset using summary statistics. We reasoned that maximum intensity for each hour could capture elements of vitality in the individual, and the variance of the data for each hour could capture the diversity of movements performed by an individual. Indeed, we found clear differences between these two parameters when assessing different age groups (Figure 1E). This effectively served to reduce our datasets to a summarized form to more efficiently build our MoveAge predictor.

To build the biological age predictor, we split the 2003–2004 dataset into separate training and testing subsets in a 70:30 ratio, and, using the training subset, trained a random forest model to predict an individual’s age from their summarized accelerometer data. We used random forest based on our and others’ previous experiences building age-predictive models (Janssens et al., 2019; Schultz et al., 2020; Shokhirev and Johnson, 2021). Exploring the random forest model’s parameters, we found both maximum activity and variance to have strong predictive abilities for age, with variance throughout the hour having greater influence on the predictive ability of the model (Figure 1F). The time of day also had an influence on the predictor’s strength in the model, with afternoon variance throughout each day of the week having greatest influence (Figure 1F).

Finally, we assessed our model on the 30% validation dataset reserved for this purpose, and found a strong correlation of 0.75 (p < 0.01) between predicted and chronological age to exist, with a root mean square error (RMSE) of 13.58 years (Figure 1G). Noting that the model tended to slightly overestimate the ages of younger individuals while underestimating the ages of older individuals, we performed a final normalization with a priori knowledge of the chronological age of an individual. Namely, an individual’s predicted biological age was normalized relative to other predicted biological ages of people with the same chronological age. This normalization allowed us to use a more accurate comparison of age acceleration and deceleration per individual, and logically improved our predicted vs. chronological age correlation to 0.94 (p < 0.01), with a RMSE of 7.54 years (Figure 1G). As our RMSE values ranged between the RMSE errors of DNA methylation-based epigenetic aging clocks (∼2.9–6 RMSE) (Galkin et al., 2020) and blood-based biomarkers of aging from NHANES (∼14–17.5 RMSE) (Pyrkov and Fedichev, 2019), we reasoned that our model was sufficiently trained to proceed with our study. We termed our final model MoveAge.

To further validate our model, and explore the relevance of MoveAge, we turned to the 2005–2006 NHANES accelerometer dataset and proceeded to predict the biological ages of individuals based on their movement patterns. Here we confirmed the accuracy of our model, with predicted ages correlating with actual chronological ages at 0.69 (p < 0.01), with an RMSE of 14.52 years before normalization. After normalization, our model achieved a correlation of 0.93 (p < 0.01) and an RMSE of 8.02 years (Figure 2A).

To further assess how age acceleration or deceleration predicted by our model correlated to biological aging, we related our predictions to the available mortality data within NHANES. Because the 2005–2006 NHANES surveying took place over a decade ago, mortality is recorded in the older age groups of the population (Figure 2B). We then calculated the deltaAge for each individual as the difference between actual chronological age and our predicted biological age, so as to evaluate its relationship to mortality (Supplementary Table 1).

With these elements in hand, we selected the individuals above the age of 60, where mortality becomes more prominent in the data (Figure 2B), and compared the quartile distributions of deltaAges to define groups that had accelerated, normal, or decelerated biological aging (Figure 2C). Comparing the ratio of live to dead from each of the three groups, we found a significant difference whereby individuals with accelerated aging had a greater proportion of mortality (0.47), individuals with decelerated aging had a lower proportion of mortality (0.34) and individuals without altered biological aging had a proportion somewhere in between (0.41) (p = 0.018) (Figure 2D). This suggested that our predictions based on movement data captured a biological aging process.

Having observed that higher mortality was present in elderly individuals with accelerated aging, we asked how deltaAge relates to mortality for each decade-sized age bin in our age distributions. For individuals of ages ranging from 18–29, and 30–39, we found no differences in mortality between individuals with age acceleration (Figure 2E). This is likely due to the fact that very few individuals had died in the follow up period, and the few that had were unlikely to be attributable to age-related causes (Figure 2E). However, starting from the 4th decade of life, we found a significant difference, whereby individuals with accelerated aging as predicted from MoveAge had higher incidences of mortality (p = 0.011, Figure 2E). This significance was maintained at the 5th (p = 0.023), 6th (p < 0.01), and 7th (p < 0.01) decades of life, with a trend still remaining, though not significant, in the 8th decade (Figure 2E). This suggested that starting from the 40s and proceeding into the next three life decades, accelerated aging plays an increasingly prominent role in determining an individual’s remaining lifespan.

Finally, we asked whether certain causes of death were more linked to accelerated aging as predicted by our model. To address this, we compared the causes of death for each mortality entry to our deltaAge measures. Here we found certain diseases associated with accelerated aging as captured by our model, such as chronic lower respiratory disease, though none were significant with the population size assessed (Figure 2F). Altogether, we conclude that our MoveAge biological age prediction captures the biological aging process in individuals, with higher mortality in individuals with accelerated aging (Figure 2G).

One of the main aims of developing our MoveAge model was to explore nutritional and pharmacological trends that are associated with aging deceleration. To address the nutritional aspect of this, we accessed the dietary intake data available for each NHANES participant. This information is derived from questionnaires that are used to estimate intakes of nutrients, and macromolecule food components. We reasoned that there might be nutritional components whose greater abundance is associated with either aging acceleration or deceleration, and that this could be detected by comparing an individual’s intake to their calculated deltaAge across the population (Figure 3A). Doing so for each decade of life would allow detection of when temporally a food component might affect biological age.

To explore the possibility that dietary components are associated to aging, we proceeded to calculate for each dietary component, at each decade of life, the correlations between deltaAge and the abundance of intake (Supplementary Table 2). We clustered the data to identify patterns occurring across the decades of life (Figure 3B). This revealed how food components had greater or lesser importance for age deceleration. Notably, we found that most food groups had little association with decelerated aging at early ages (Figure 3C). However, a cluster of food components tended to peak after increasing steadily in importance throughout life (Figure 3C, cluster 1). Higher intakes of these specific foods were associated to younger biological ages, growing in importance from the 4th and 5th decade, reaching maximal association at the 6th.

We further explored the components of this cluster, and found the three most significant food components associated with decelerated aging at the 6th decade of life were fiber (r = −0.18, p < 0.01), magnesium (r = −0.17, p < 0.01), and vitamin E (r = −0.13, p < 0.01) (Figures 3D–F, left panels). Higher intake of these foods was associated to decelerated aging. This could further be exemplified by comparing individuals with either very high (>10 years) or very low (<10 years) deltaAges. Doing so revealed a significant association with decelerated aging for all three food components (Figures 3D–F, right panels). Together, these findings are in line with general observations known to benefit health in humans, such as how higher fiber intake is linked to lower mortality (Gopinath et al., 2016), magnesium deficiency is associated to age-related diseases in the elderly (Barbagallo and Dominguez, 2018), and vitamin E may benefit the lifespan of certain human populations (Hemilä and Kaprio, 2011). We therefore conclude that our MoveAge model is useful for the potential identification of factors associated with decelerated aging.

Our final aim was to identify if specific drugs that individuals in our dataset were taking may also be associated to decelerated aging. This approach would, in effect, constitute a human in vivo screening for compounds promoting healthy aging. While false positives may result from this approach in the form of compounds only associated, rather than causing, age deceleration (e.g., antibiotics or antihistamines, which are the main drugs that otherwise healthy individuals may take), it could nonetheless reveal compounds that contribute to healthy aging (Figure 4A).

To identify compounds associated to decelerated aging, we focused on the 70+ population who was both elderly and most likely to be taking pharmaceuticals (Figure 4B). Then, for each drug, we compared the distribution of deltaAges of users of that drug, relative to the distribution of deltaAges of others in this same demographic. This allowed us to generate both a significance level comparing these distributions, and a median fold change in delta ages that could be represented as a volcano plot (Figure 4C). With these metrics we could rank the compounds based on their p-values (Figure 4D), and, a non-stringent significance criteria (p < 0.05) could generate a list of compounds that were most associated with decelerated aging (doxazosin, amoxicillin, verapamil) or accelerated aging (omerprazole, captopril) (Figure 4E). While no drug of the 500 + FDA approved drugs we assessed passed significance following a correction for multiple hypothesis testing (Benjamini and Hochberg), one compound—doxazosin—passed significance using a stringent unadjusted p-value cutoff (p < 0.01) (Figure 4E). Users of doxazosin were more likely to have negative deltaAges, representing a possible age-deceleration due to the drug, when compared to all others in the same demographic population (Figure 4F).

As stated above, the association between doxazosin and decelerated age does not necessarily mean it causally benefits healthy aging. We therefore turned to C. elegans to test if doxazosin was causally geroprotective. We used C. elegans because it is a well-described aging model organism, and has been used to identify and understand other lifespan extending drugs, such as metformin and others (Cabreiro et al., 2013; Ye et al., 2014). We based our original screen on movement data, so we aimed to determine if doxazosin could improve the worms’ healthspan, which we initially tested using a measure of age-related mobility later in life. When worms were grown on agar plates that contained 33 μM doxazosin, a dose previously shown to extend lifespan in worms (Ye et al., 2014), they crawled across the plate at significantly higher speeds than those treated with the vehicle at day 9 or 10 of adulthood (p < 0.001) (Figure 4G; Supplementary Table 3). We then used a microfluidics platform totest lifespan and healthspan. Here we used multiple doses of doxazosin (3.3 and 33 μM) as previous observations show that microfluidics devices provide more direct drug contact than agar plates, and therefore a lower dose is needed (Hewitt et al., 2018). Indeed, we saw the greatest beneficial effects at 3.3 μM (Figures 4H,I; Supplementary Tables 4, 5). We confirmed that doxazosin extended worm lifespan (p < 0.0001) (Figure 4H; Supplementary Table 4) and simultaneously studied worm motility so as to measure healthspan via locomotion. With automated locomotion measurements, we calculated what percentage of worms remained highly active (defined as moving a distance greater than their body length in a 30 s window). Upon doxazosin treatment, we observed a higher percentage of highly active animals throughout life, particularly in the later ages tested, confirming their improved healthspan (Figure 4E; Supplementary Table 5). Altogether, we conclude that doxazosin extends both lifespan and healthspan in C. elegans.