Statistical inferences vs. biological realities

A 2019 UCLA study introduced a derivative of the epigenetic clock named GrimAge:

“DNAm GrimAge, a linear combination of chronological age, sex, and DNAm-based surrogate biomarkers for seven plasma proteins and smoking pack-years, outperforms all other DNAm-based biomarkers, on a variety of health-related metrics.

An age-adjusted version of DNAm GrimAge, which can be regarded as a new measure of epigenetic age acceleration (AgeAccelGrim), is associated with a host of age-related conditions, lifestyle factors, and clinical biomarkers. Using large scale validation data from three ethnic groups, we demonstrate that AgeAccelGrim stands out among pre-existing epigenetic clocks in terms of its predictive ability for time-to-death, time-to-coronary heart disease, time-to-cancer, its association with computed tomography data for fatty liver/excess fat, and early age at menopause.” “DNA methylation GrimAge strongly predicts lifespan and healthspan”

A miserable attempt at reporting the study’s findings included angles of superstition, fear-of-the-future, and suspicion-by-spurious-association:

“The research has already captured the attention of the life insurance industry. After all, a solid death date could mean real savings when it comes to pricing policies.

The hope is that if and when legitimate anti-aging drugs are developed, GrimAge could be used to test their effectiveness. In a world with functional anti-aging drugs, “doctors could test [your GrimAge number] and say, ‘You know what, you’re aging too quickly. Take this,'” Horvath said.” “A New Test Predicts When You’ll Die (Give or Take a Few Years)”

A detailed blog post from Josh Mitteldorf provided scientific coverage of the study:

“Methylation sites associated with smoking history predicted how long the person would live more accurately than the smoking history itself. Even stranger, the methylation marks most closely associated with smoking were found to be a powerful indication of future health even when the sample was confined to non-smokers.

The DNAm GrimAge clock was developed in two stages, a correlation of a correlation. Curiously, the indirect computation yields the better result.

Horvath’s finding that secondary methylation indicators are more accurate than the underlying primary indicator from which they were derived is provocative, and calls out for a new understanding.” “DNAm GrimAge—the Newest Methylation Clock”

When there are logical disconnects in findings like the above, it’s time to examine underlying premises. As noted in Group statistics don’t necessarily describe an individual, an assumption required by statistical analyses is that each measured item in the sample is interchangeable with the next.

This presumption is often false, producing individually inapplicable results. For example, the review from earlier this week Immune memory vs. immune adaptation included this description of the adaptive immune system:

“To be effective, highly specific immune response requires huge diversity of receptors and antibodies, which is achieved by somatic rearrangement of gene segments. Recombination results in millions of TCR [T cell receptor] and antibody variants able to recognize and neutralize millions of various antigens.”

Standard statistics of millions of T cell receptor and antibody variants won’t represent their individually unique properties. Individual differences are their purpose and benefit to us.

The GrimAge study’s overreach was most apparent in stratifying educational attainment to develop correlations. As mentioned in Does a societal mandate cause DNA methylation? such statistics are poor evidence of each individual’s biological realities.

Neither derivatives of group statistics, nor correlations of correlations, seem to be the techniques needed to understand biological causes of effects. Commentators on the GrimAge study alluded to this point, but glossed it over:

“It remains a mystery why exactly the epigenetic clocks work, and whether age-related changes in DNA methylation contribute to the cause of aging or are a result of it.”


A therapy to reverse cognitive decline

This 2018 human study presented the results of 100 patients’ personalized therapies for cognitive decline:

“The first examples of reversal of cognitive decline in Alzheimer’s disease and the pre-Alzheimer’s disease conditions MCI (Mild Cognitive Impairment) and SCI (Subjective Cognitive Impairment) have recently been published..showing sustained subjective and objective improvement in cognition, using a comprehensive, precision medicine approach that involves determining the potential contributors to the cognitive decline (e.g., activation of the innate immune system by pathogens or intestinal permeability, reduction in trophic or hormonal support, specific toxin exposure, or other contributors), using a computer-based algorithm to determine subtype and then addressing each contributor using a personalized, targeted, multi-factorial approach dubbed ReCODE for reversal of cognitive decline.

An obvious criticism of the initial studies is the small number of patients reported. Therefore, we report here 100 patients, treated by several different physicians, with documented improvement in cognition, in some cases with documentation of improvement in electrophysiology or imaging, as well.” “Reversal of Cognitive Decline: 100 Patients”

The lead author commented on Josh Mitteldorf’s informative post A cure for Alzheimer’s? Yes, a cure for Alzheimer’s!:

  1. “We have a paper in press, due to appear 10.22.18 (open access, JADP, I’ll send a copy as soon as available), showing 100 patients with documented improvement – some with MRI volumetrics improved, others with quantitative EEG improvements, others with evoked response improvements, and all with quantitative cognitive assessment improvement. Some are very striking – 12 point improvements in MoCA, for example – others less so, but all also have subjective improvement. Hopefully this will address some of the criticisms that we haven’t documented improvement in enough people.
  2. We were just turned down again for a randomized, controlled clinical trial, so on the one hand, we are told repeatedly that no one will believe that this approach works until we publish a randomized, controlled study, and on the other hand, we’ve been turned down (first in 2011/12, and now in 2018), with the complaint that we are trying to address more than one variable in the trial (as if AD is a single-variable disease!). Something of a catch-22. We are now resubmitting (unfortunately, the IRBs are not populated by functional medicine physicians, so they are used to seeing old-fashioned drug studies), and we’ll see what happens.
  3. I’ve been extending the studies to other neurodegenerative diseases, and it has been impressive how much of a programmatic response there seems to be in these “diseases.”
  4. I agree with you that there are many features in common with aging itself.
  5. You made a good point that APP is a dependence receptor, and in fact it functions as an integrating dependence receptor, responding to numerous inputs (Kurakin and Bredesen, 2015).
  6. In the book and the publications, we don’t claim it is a “cure” since we don’t have pathological evidence that the disease process is gone. What we claim is “reversal of cognitive decline” since that is what we document.
  7. As I mentioned in the book, AD is turning out to be a protective response to multiple insults, and this fits well with the finding that Abeta has an antimicrobial effect (Moir and Tanzi’s work). It is a network-downsizing, protective response, which is quite effective – some people live with the ongoing degenerative process for decades.
  8. We have seen several cases now in which a clinical trial of an anti-amyloid antibody made the person much worse in a time-dependent manner (each time there was an injection, the person would get much worse for 5-10 days, then begin to improve back toward where he/she was, but over time, marked decline occurred), and this makes sense for the idea that the amyloid is actually protecting against pathogens or toxins or some other insult.
  9. It is important to note that we’ve never claimed that all people get better – this is not what we’ve seen. People very late in the process, or who don’t follow the protocol, or who don’t address the various insults, do not improve. It is also turning out to be practitioner dependent – some are getting the vast majority of people to improve, others very few, so this is more like surgery than old-fashioned prescriptive medicine – you have to do a somewhat complicated therapeutic algorithm and get it right for best results.
  10. I’m very interested in what is needed to take the next step in people who have shown improvement but who started late in the course. For example, we have people now who have increased MoCA from 0 to 9 (or 0 to 3, etc.), with marked subjective improvement but plateauing at less than normal. These people had extensive synaptic and cellular loss prior to the program. So what do we need to raise the plateau? Stem cells? Intranasal trophic support? Something else?
  11. I haven’t yet seen a mono-etiologic theory of AD or a mono-therapeutic approach that has repeatedly positive results, so although I understand that there are many theories and treatments, there doesn’t seem to be one etiology to the disease, nor does there seem to be one simple treatment that works for most. It is much more like a network failure.”

At a specific level:

  • “There doesn’t seem to be one etiology to the disease,
  • nor does there seem to be one simple treatment that works for most.
  • We don’t have pathological evidence that the disease process is gone.”

For general concepts, however:

  • “AD is turning out to be a protective response to multiple insults,
  • It is a network-downsizing, protective response, which is quite effective.
  • The amyloid is actually protecting against pathogens or toxins or some other insult.”

For a framework of an AD cure to be valid, each source of each insult that evoked each “protective response” should be traced.

Longitudinal studies would be preferred inside this framework. These study designs would investigate evidence of each insult’s potential modifying effect on each “protective response” that could affect the cumulative disease trajectory of each individual.

In many cases, existing study designs would be adequate if they extended their periods to the end of the subjects’ natural lifetimes. One AD-relevant example would be extending the prenatally-restraint-stressed model used in:

The framework would also encourage extending studies to at least three generations to investigate evidence for transgenerational effects, as were found in:

An hour of the epigenetic clock

This 2018 presentation by the founder of the epigenetic clock method described the state of the art up through July 2018. The webinar was given on the release day of The epigenetic clock now includes skin study.

Segments before the half-hour mark provide an introduction to the method and several details about the concurrently-released study. The Q&A section starts a little before the hour mark.

Reversing epigenetic changes with CRISPR/Cas9

This 2018 Chinese review highlighted areas in which CRISPR/Cas9 technology has, is, and could be applied to rewrite epigenetic changes:

“CRISPR/Cas9-mediated epigenome editing holds a great promise for epigenetic studies and therapeutics.

It could be used to selectively modify epigenetic marks at a given locus to explore mechanisms of how targeted epigenetic alterations would affect transcription regulation and cause subsequent phenotype changes. For example, inducing histone methylation or acetylation at the Fosb locus in the mice brain reward region, nucleus accumbens, could affect relevant transcription network and thus control behavioral responses evoked by drug and stress.

Epigenome editing has the potential for epigenetic treatment, especially for the disorders with abnormal gene imprinting or epigenetic marks. Targeted epigenetic silencing or reactivation of the mutant allele could be a potential therapeutic approach for diseases such as Rett syndrome and Huntington’s disease.

Noncoding RNA plays important roles in gene imprinting and chromatin remodeling. CRISPR/Cas9 has been shown to be potential for manipulating noncoding RNA expression, including microRNA, long noncoding RNA, and miRNA families and clusters.

In vivo overexpression of the Yamanaka factors have proven to be able to fully or partially help somatic cells to regain pluripotency in situ. These rejuvenated cells would subsequently differentiate again to replace the lost cell types.”

The last paragraph was described in The epigenetic clock theory of aging as a promising technique:

“To date, the most effective in vitro intervention against epigenetic ageing is achieved through expression of Yamanaka factors, which convert somatic cells into pluripotent stem cells, thereby completely resetting the epigenetic clock.”

The reviewers cited three references for in vivo studies of this technique. Overall, I didn’t see that any of the review’s references were in vivo human studies. “Novel Epigenetic Techniques Provided by the CRISPR/Cas9 System”

The epigenetic clock now includes skin

The originator of the 2013 epigenetic clock improved its coverage with this 2018 UCLA human study:

“We present a new DNA methylation-based biomarker (based on 391 CpGs) that was developed to accurately measure the age of human fibroblasts, keratinocytes, buccal cells, endothelial cells, skin and blood samples. We also observe strong age correlations in sorted neurons, glia, brain, liver, and bone samples.

The skin & blood clock outperforms widely used existing biomarkers when it comes to accurately measuring the age of an individual based on DNA extracted from skin, dermis, epidermis, blood, saliva, buccal swabs, and endothelial cells. Thus, the biomarker can also be used for forensic and biomedical applications involving human specimens.

The biomarker applies to the entire age span starting from newborns, e.g. DNAm of cord blood samples correlates with gestational week.

Furthermore, the skin & blood clock confirms the effect of lifestyle and demographic variables on epigenetic aging. Essentially it highlights a significant trend of accelerated epigenetic aging with sub-clinical indicators of poor health.

Conversely, reduced aging rate is correlated with known health-improving features such as physical exercise, fish consumption, high carotenoid levels. As with the other age predictors, the skin & blood clock is also able to predict time to death.

Collectively, these features show that while the skin & blood clock is clearly superior in its performance on skin cells, it crucially retained all the other features that are common to other existing age estimators.” “Epigenetic clock for skin and blood cells applied to Hutchinson Gilford Progeria Syndrome and ex vivo studies”

An introduction to the study highlighted several items:

“Although the skin-blood clock was derived from significantly less samples (~900) than Horvath’s clock (~8000 samples), it was found to more accurately predict chronological age, not only across fibroblasts and skin, but also across blood, buccal and saliva tissue. A potential factor driving this improved accuracy in blood could be related to the approximate 18-fold increase in genomic coverage afforded by using Illumina 450k/850k beadarrays.

It serves as a roadmap for future clock studies, pointing towards the importance of constructing tissue or cell-type specific epigenetic clocks, to more accurately measure biological aging in the given tissue/cell-type, and therefore with the potential to be more informative of disease-risk or the success of disease interventions in the tissue or cell-type of interest.” “Epigenetic clocks galore: a new improved clock predicts age-acceleration in Hutchinson Gilford Progeria Syndrome patients”

Hijacking the epigenetic clock paradigm

This 2018 German human study’s last sentence was:

“Additionally we found an association between DNAm [DNA methylation] age acceleration and rLTL [relative leukocyte telomere length], suggesting that this epigenetic clock, at least partially and possibly better than other epigenetic clocks, reflects biological age.”

Statements in the study that contradicted, qualified, and limited the concluding sentence included:

“The epigenetic clock seems to be mostly independent from the mitotic clock as measured by the rLTL.

It could be possible that associations are confounded due to short age ranges or non-continuous age distribution, as displayed in the BASE-II cohort (no participants between the age of 38 and 59 years). [see the below graphic]

The BASE-II is a convenience sample and participants have been shown to be positively selected with respect to education, health and cognition.

Samples in which DNAm age and chronological age differed more than three standard deviations from the mean were excluded (N=19).

While the original publication employed eight CpG sites for DNAm age estimation, we found that one of these sites did not significantly improve chronological age prediction in BASE-II. Thus, we reduced the number of sites considered to seven in the present study and adapted the algorithm to calculate DNAm age.

  • Horvath described a subset of 353 methylation sites predicting an individual’s chronological age with high accuracy..
  • Even though the available methods using more CpG sites to estimate DNAm age predict chronological age with higher accuracy..
  • It is not clear how much of the deviation between chronological age and DNAm age reflects measurement error/low number of methylation sites and which proportion can be attributed to biological age.

Due to the statistical method employed, we encountered a systematic deviation of DNAm age in our dataset.”

Findings that aren’t warranted by the data is an all-too-common problem with published research. This study illustrated how researcher hypothesis-seeking behavior – that disregarded what they knew or should have known – can combine with a statistics package to produce almost any finding.

It reminded me of A skin study that could have benefited from preregistration that made a similar methodological blunder:

The barbell shape of the subjects’ age distribution wouldn’t make sense if the researchers knew they were going to later use the epigenetic clock method.

The researchers did so, although the method’s instructive study noted:

“The standard deviation of age has a strong relationship with age correlation”

and provided further details in “The age correlation in a data set is determined by the standard deviation of age” section.

Didn’t the researchers, their organizations, and their sponsors realize that this study’s problematic design and performance could misdirect readers away from the valid epigenetic clock evidence they referenced? What purposes did it serve for them to publish this study? “Epigenetic clock and relative telomere length represent largely different aspects of aging in the Berlin Aging Study II (BASE-II)” (not freely available)

Epigenetic effects of breast cancer treatments

This 2018 UC San Diego review subject was the interplay between breast cancer treatments and their effects on aging:

“Although current breast cancer treatments are largely successful in producing cancer remission and extending lifespan, there is concern that these treatments may have long lasting detrimental effects on cancer survivors, in part, through their impact on non-tumor cells. It is unclear whether breast cancer and/or its treatments are associated with an accelerated aging phenotype.

In this review, we have highlighted five of nine previously described cellular hallmarks of aging that have been described in the context of cytotoxic breast cancer treatments:

  1. Telomere attrition;
  2. Mitochondrial dysfunction;
  3. Genomic instability;
  4. Epigenetic alterations; and
  5. Cellular senescence.”

The review was full of caveats weakening the above graphic’s associations:

  1. “Telomere attrition – Blood TL [telomere length] was not associated with chemotherapy in three out of four studies;
  2. Mitochondrial dysfunction – How cancer therapies affect cellular energetics as they relate to rate of aging is unclear;
  3. Genomic instability – Potentially contributing to accelerated aging;
  4. Epigenetic alterations – Although some of the key regulators of these processes have begun to be identified, including DNA and histone methylases and demethylases, histone acetylases and de-acetylases and chromatin remodelers, how they regulate the changes in aging through alteration of global transcriptional programs, remains to be elucidated; and
  5. Cellular senescence – Dysregulated pathways can be targeted by cytotoxic chemotherapies, resulting in preferential cell death of tumor cells, but how these treatments also affect normal cells with intact pathways is unclear.”

To their credit, the reviewers at least presented some of the contrary evidence, and didn’t continue on with a directed narrative as many reviewers are prone to do. “Breast cancer treatment and its effects on aging” (not freely available)

The originator of the epigenetic clock methodology was a coauthor of the review. Only one of his works was cited in the Epigenetic alterations subsection: “DNA methylation age is elevated in breast tissue of healthy women”

This freely-available 2017 study quoted below highlighted that epigenetic clock measurements as originally designed were tissue-specific:

“To our knowledge, this is the first study to demonstrate that breast tissue epigenetic age exceeds that of blood tissue in healthy female donors. In addition to validating our earlier finding of age elevation in breast tissue, we further demonstrate that the magnitude of the difference between epigenetic age of breast and blood is highest in the youngest women in our study (age 20–30 years) and gradually diminishes with advancing age. As women approach the age of the menopausal transition, we found that the epigenetic of age of blood approaches that of the breast.”

Additional caution was justified in both interpreting age measurements and extending them into “cellular hallmarks” when the tissue contained varying cell types:

“Our studies were performed on whole breast tissue. Diverse types of cells make up whole breast tissue, with the majority of cells being adipocytes. Other types of cells include epithelial cells, cuboidal cells, myoepithelial cells, fibroblasts, inflammatory cells, vascular endothelial cells, preadipocytes, and adipose tissue macrophages.

This raises the possibility that the magnitude of the effects we observe, of breast tissue DNAm age being greater than other tissues, might be an underestimation, since it is possible that not all of the cells of the heterogenous sample have experienced this effect. Since it is difficult to extract DNA from adipose tissue, we suspect that the majority of DNA extracted from our whole breast tissues was from epithelial and myoepithelial cells.”