Article activity feed

  1. Author Response

    Reviewer 1

    In this article Farrell et al. leverage existing datasets which measure frailty longitudinally in mice and humans to model 'robustness' (the ability to resist damage) and 'resilience' (the ability to recover from damage), their dynamics across age, and their relative contributions to overall frailty and mortality. The concept of separating damage/robustness from recovery/resilience is valid and has many important applications including better assessment and prediction of effective intervention strategies. I also appreciate the authors' sophisticated attempts to effectively model longitudinal data, which is a challenge in the field. The use of human and mouse data is another strength of the study, and it is quite interesting to see overlapping trends between the two species.

    While I find the rationale sound and appreciate the approach taken at a high level, there are a few key considerations of the specific data used which are lacking. The authors conceptualize resilience based on studies which primarily use short time scales and dynamic objective measures (ex. complete blood cell counts in Pyrkov et al.) often in conjunction with an acute stress stimulus. For example, they heavily cite Ukraintseva et al. who define resilience as "the ability to quickly and completely recover after deviation from normal physiological state or damage caused by a stressor or an adverse health event."

    Resilience and robustness are typically studied at short time-scales, with small numbers of continuous health attributes. We study transitions of binary health attributes, which we call damage and repair, and which we suggest should be thought of as resilience and robustness. Our approach is well suited for studying large numbers of binary health attributes over long time-scales without acute external stimuli. How resilience and robustness in these limits (binary, large numbers, long times, intrinsic dynamics) compare with resilience and robustness as has been typically measured (continuous, short times, acute stimuli) is an interesting and important question that arises from our work.

    Given these definitions, the human data used seem to fit within this framework, but we should carefully consider the mouse data. The mouse frailty index is a very useful tool for efficiently measuring the organismal state in large cohorts. A tradeoff for quickly measuring a broad range of health domains is that the individual measurements are low resolution (categorical) and involve inherent subjectivity (which may be considered part of the measurement error). Some transitions in individual components are due to random measurement error and I believe this is especially likely with decreases (or 'resilience' transitions).

    The reason I think the resilience transitions are subject to high measurement error is that I am skeptical as to whether many of the deficits in the mouse index are reversible under normal physiologic conditions. For example, it is exceptionally unlikely for a palpable/visible tumor to resolve in an aged mouse over the time scales studied here, thus any reversal that was observed is very likely due to random measurement error. Other components which I have doubts about reversibility are alopecia, loss of fur color, loss of whiskers, tumors, kyphosis, hearing loss, cataracts, corneal capacity, vision loss, rectal prolapse, genital prolapse.

    In summary, I applaud the authors' efforts in generating complex models to better understand longitudinal aging data. This is an important area that needs further development. I appreciate their conceptualization of resilience and robustness and think this framework has an important place in aging research. I also appreciate their cross-species approach. However, the authors may have over-conceptualized and made some assumptions about the mouse data which may not be valid. It will be important to assess the results with careful consideration of the time scales of the underlying biology and the resolution and measurement error inherent to these tools.

    For each of our mouse attributes, there are published studies demonstrating reversibility (see our new Supplementary Table 1). Nevertheless, we cannot distinguish what causes the observed discrete transitions (measurement error, stochastic fluctuations in underlying organismal features, or logisticlike continuous transitions in underlying continuous variables). We analyze the discrete data as given.

    The question of time-scale is interesting. From survival curves of individual binarized attributes, we obtain reasonable fits to exponential models (i.e. a single timescale) see Fig 5 supplement 1 and 2. For the human data there are a broad range of timescales for both robustness and resilience. For the mouse data there appears to be a similarly broad range (note the logarithmic scale) though with considerable uncertainty. We work with the data we have, so we are unable to probe shorter timescales than the measurement interval (months for mice, and years for humans). We have reinforced this caveat in the discussion.

    Reviewer 2

    This study uses repeated measurements of the frailty index (FI), composed of multiple binary parameters. It is posited that newly detected changes in the number of these parameters represent damage and that the parameters that have previously been detected but are not detected currently represent damage repair. Statistical treatment then follows, deriving resilience and robustness and their changes over time. This is an interesting idea. Strengths of the study include analyses across species (mice and humans), including multiple datasets in mice.

    To be clear, our data analysis is on the binary health attributes that are used in the FI. By considering the damage/repair (binary transitions) of individual attributes, we can obtain the aggregate damage/repair rates.

    What are the elements of FI that increase at each period of life, and what are those that decrease? For example, humped phenotype or alopecia are more likely to appear in old mice and are essentially irreversible, whereas weight loss due to infection may be more common in young mice and is reversible. Therefore, the choice of health deficits would affect the model and, for example, may artificially lead to a decreased value of what the authors call damage repair.

    More generally, information on the frailty index lacks sufficient details. I doubt that this method has sufficient accuracy to draw conclusions from as little as 32 female mice (21 + 11 animals in datasets 1 and 2) and 63 males (13 + 6 + 44 animals in datasets 1, 2 and 3). Also, only 25 enalapril-treated mice of each sec were analyzed, and only 17 exercised mice (11 females and 6 males). The number of human participants is large, but the total follow-up period is not shown, and the subjects were assessed based on 23 parameters only.

    We have not examined other choices of health attributes. While we picked standard sets from available data, we do not know whether other attributes would behave differently. It would be difficult to do our detailed modelling on single attributes in the mouse data, since the data is so sparse. Our approach was developed specifically to be able to draw conclusions from limited mouse data. Where possible we aggregate the individual mice, sex, health attributes, studies, and measurement times. The analysis of human data shows that the approach generalizes.

    While we have mostly not studied individual attributes (we have considered survival times, but without age or time effects), we would expect that some of them may have behavior that qualitatively differs from our aggregate results. If attribute selection was biased towards (or away from) qualitatively distinct behaviors that would, of course, be reflected in aggregate results. We suspect that this would be unlikely, but that any such distinctive behavior would be interesting and important to identify and understand. We have added some discussion on this point, since we cannot exclude this possibility.

    A key assumption in this work is that increased FI is equivalent to the rise in damage. However, the relationship between changes in FI and damage is unknown. One can imagine a situation when damage increases, but protection also increases. In this case, fitness may increase, decrease or remain unchanged. What is the basis for calling an increased number of health deficits damage? Is there a more reliable method to measure damage that could support the authors' claims?

    See also discussion point #1 in essential revisions. We call binary states 0 “healthy” and 1 “damaged”, but we could instead say “more healthy for most individuals” and “less healthy for most individuals” – where “healthy” means associated with desirable (low FI and low mortality) health outcomes. We have not explored other measures of organismal damage. We have not explored how interactions between variables could affect resilience or robustness for individuals. We do not think that alternative approaches would be easy to study without much more data (for mice) that is more finely resolved in time (for mice and humans). We are quite happy to have found an approach to use with binarized data, but would welcome viable alternative approaches to compare with.

    Reviewer 3

    In this work, the authors aimed at investigating two related components of aging-related processes of health deficits accumulation in mice and humans: the processes of damage (representing the robustness of an organism) and repair (corresponding to resilience), and at determining how different interventions (the angiotensin-converting enzyme inhibitor enalapril and voluntary exercise) in mice and a representative measure of socio-economic status (household wealth) in humans affect the rates of damage and repair. Two key elements in this study allowed the authors to achieve their goals: 1) the use of relevant data containing repeated measurements of health deficits from which they were able to compute the cumulative indices of health deficits in mice and humans and which are also necessary to evaluate the processes of damage and repair; 2) the methodological approach that allowed them to formulate the concepts of damage and repair, model them and estimate from the available data. This methodological framework coupled with the data resulted in important findings about the contribution of the age-related decline in robustness and resilience in health deficits accumulation with age and the differential impact of interventions on the processes of damage and repair. This provides important insights into these key components of the process of aging and this research should be of interest to both lab researchers who plan experimental studies with laboratory animals to study potential mechanisms and interventions affecting health deficits accumulation as well as researchers working with human longitudinal studies who can apply this approach to further investigate the impact of different factors on robustness and resilience and their contribution to the overall health deterioration, onset of diseases and, eventually, death.

    The key strength of this work is a rigorous analytic approach that includes joint modeling of longitudinal measurements of health deficits and mortality (in mice). This approach avoids biased inference which would be observed if longitudinal data were analyzed alone, ignoring attrition due to mortality. Another strength is a comprehensive analysis of both laboratory animal data that allows exploring the impact of different interventions on the processes of damage and repair and human data that allows investigating disparities in these processes in individuals with different socioeconomic backgrounds (represented by household wealth).

    One weakness (which is commonplace for human studies) is self-reported data on health deficits in humans which makes it difficult to compare with lab data where deficits are assessed objectively by lab researchers. The subjective nature of health deficits measurements complicates the interpretation of findings, especially about repairs of deficits. In addition, it is not clear whether the availability/absence of caregivers at different exams/interviews factors into the answers on difficulty/not difficulty with specific activities constituting health deficits and, respectively, into their change over time reflected in damage/repair estimates.

    Variability of the evaluator is expected in any longitudinal study, and amounts to a variety of measurement error. The question of whether there are age-effects in the measurement error, such as bias or age-dependent variability is interesting. For the mouse data, evaluator training is designed to minimize such errors and inter-evaluator differences are not large (Feridooni et al, 2015; Kane et al, 2017). For the human self-report data any such age-effects are unavoidable.

    Read the original source
    Was this evaluation helpful?
  2. Evaluation Summary:

    The key contribution of this paper is to measure frailty longitudinally in mice and humans to model 'robustness' (the ability to resist damage) and 'resilience' (the ability to recover from damage). To model these concepts, a frailty index (FI) composed of multiple binary parameters is calculated, but with the novel contribution that newly detected changes represent damage and that the parameters that have previously been detected but are not detected currently represent damage repair. Statistical steps then derive resilience and robustness and their changes over time. The sophisticated attempts to effectively model longitudinal data and rigorous analytic approach are strengths, as is the use of both human and animal species and intervention studies. A few overarching concerns were raised, primarily pertaining to the potential risk of over-conceptualized links between deficit index and biologic constructs of 'damage' and 'repair', but it nonetheless advances a growing field interested in measuring the longitudinal change in biologic age.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #3 agreed to share their name with the authors.)

    Read the original source
    Was this evaluation helpful?
  3. Reviewer #1 (Public Review):

    In this article Farrell et al. leverage existing datasets which measure frailty longitudinally in mice and humans to model 'robustness' (the ability to resist damage) and 'resilience' (the ability to recover from damage), their dynamics across age, and their relative contributions to overall frailty and mortality. The concept of separating damage/robustness from recovery/resilience is valid and has many important applications including better assessment and prediction of effective intervention strategies. I also appreciate the authors' sophisticated attempts to effectively model longitudinal data, which is a challenge in the field. The use of human and mouse data is another strength of the study, and it is quite interesting to see overlapping trends between the two species.

    While I find the rationale sound and appreciate the approach taken at a high level, there are a few key considerations of the specific data used which are lacking. The authors conceptualize resilience based on studies which primarily use short time scales and dynamic objective measures (ex. complete blood cell counts in Pyrkov et al.) often in conjunction with an acute stress stimulus. For example, they heavily cite Ukraintseva et al. who define resilience as "the ability to quickly and completely recover after deviation from normal physiological state or damage caused by a stressor or an adverse health event."

    Given these definitions, the human data used seem to fit within this framework, but we should carefully consider the mouse data. The mouse frailty index is a very useful tool for efficiently measuring the organismal state in large cohorts. A tradeoff for quickly measuring a broad range of health domains is that the individual measurements are low resolution (categorical) and involve inherent subjectivity (which may be considered part of the measurement error). Some transitions in individual components are due to random measurement error and I believe this is especially likely with decreases (or 'resilience' transitions).

    The reason I think the resilience transitions are subject to high measurement error is that I am skeptical as to whether many of the deficits in the mouse index are reversible under normal physiologic conditions. For example, it is exceptionally unlikely for a palpable/visible tumor to resolve in an aged mouse over the time scales studied here, thus any reversal that was observed is very likely due to random measurement error. Other components which I have doubts about reversibility are alopecia, loss of fur color, loss of whiskers, tumors, kyphosis, hearing loss, cataracts, corneal capacity, vision loss, rectal prolapse, genital prolapse.

    In summary, I applaud the authors' efforts in generating complex models to better understand longitudinal aging data. This is an important area that needs further development. I appreciate their conceptualization of resilience and robustness and think this framework has an important place in aging research. I also appreciate their cross-species approach. However, the authors may have over-conceptualized and made some assumptions about the mouse data which may not be valid. It will be important to assess the results with careful consideration of the time scales of the underlying biology and the resolution and measurement error inherent to these tools.

    Read the original source
    Was this evaluation helpful?
  4. Reviewer #2 (Public Review):

    This study uses repeated measurements of the frailty index (FI), composed of multiple binary parameters. It is posited that newly detected changes in the number of these parameters represent damage and that the parameters that have previously been detected but are not detected currently represent damage repair. Statistical treatment then follows, deriving resilience and robustness and their changes over time. This is an interesting idea. Strengths of the study include analyses across species (mice and humans), including multiple datasets in mice.

    What are the elements of FI that increase at each period of life, and what are those that decrease? For example, humped phenotype or alopecia are more likely to appear in old mice and are essentially irreversible, whereas weight loss due to infection may be more common in young mice and is reversible. Therefore, the choice of health deficits would affect the model and, for example, may artificially lead to a decreased value of what the authors call damage repair.

    More generally, information on the frailty index lacks sufficient details. I doubt that this method has sufficient accuracy to draw conclusions from as little as 32 female mice (21 + 11 animals in datasets 1 and 2) and 63 males (13 + 6 + 44 animals in datasets 1, 2 and 3). Also, only 25 enalapril-treated mice of each sec were analyzed, and only 17 exercised mice (11 females and 6 males). The number of human participants is large, but the total follow-up period is not shown, and the subjects were assessed based on 23 parameters only.

    A key assumption in this work is that increased FI is equivalent to the rise in damage. However, the relationship between changes in FI and damage is unknown. One can imagine a situation when damage increases, but protection also increases. In this case, fitness may increase, decrease or remain unchanged. What is the basis for calling an increased number of health deficits damage? Is there a more reliable method to measure damage that could support the authors' claims?

    Read the original source
    Was this evaluation helpful?
  5. Reviewer #3 (Public Review):

    In this work, the authors aimed at investigating two related components of aging-related processes of health deficits accumulation in mice and humans: the processes of damage (representing the robustness of an organism) and repair (corresponding to resilience), and at determining how different interventions (the angiotensin-converting enzyme inhibitor enalapril and voluntary exercise) in mice and a representative measure of socio-economic status (household wealth) in humans affect the rates of damage and repair. Two key elements in this study allowed the authors to achieve their goals: 1) the use of relevant data containing repeated measurements of health deficits from which they were able to compute the cumulative indices of health deficits in mice and humans and which are also necessary to evaluate the processes of damage and repair; 2) the methodological approach that allowed them to formulate the concepts of damage and repair, model them and estimate from the available data. This methodological framework coupled with the data resulted in important findings about the contribution of the age-related decline in robustness and resilience in health deficits accumulation with age and the differential impact of interventions on the processes of damage and repair. This provides important insights into these key components of the process of aging and this research should be of interest to both lab researchers who plan experimental studies with laboratory animals to study potential mechanisms and interventions affecting health deficits accumulation as well as researchers working with human longitudinal studies who can apply this approach to further investigate the impact of different factors on robustness and resilience and their contribution to the overall health deterioration, onset of diseases and, eventually, death.

    The key strength of this work is a rigorous analytic approach that includes joint modeling of longitudinal measurements of health deficits and mortality (in mice). This approach avoids biased inference which would be observed if longitudinal data were analyzed alone, ignoring attrition due to mortality. Another strength is a comprehensive analysis of both laboratory animal data that allows exploring the impact of different interventions on the processes of damage and repair and human data that allows investigating disparities in these processes in individuals with different socioeconomic backgrounds (represented by household wealth).

    One weakness (which is commonplace for human studies) is self-reported data on health deficits in humans which makes it difficult to compare with lab data where deficits are assessed objectively by lab researchers. The subjective nature of health deficits measurements complicates the interpretation of findings, especially about repairs of deficits. In addition, it is not clear whether the availability/absence of caregivers at different exams/interviews factors into the answers on difficulty/not difficulty with specific activities constituting health deficits and, respectively, into their change over time reflected in damage/repair estimates.

    Read the original source
    Was this evaluation helpful?