Thread by @KarlPfleger

Direct-to-consumer aging clocks give very inconsistent answers. I tried TruDiagnostic, Elysium, GlycanAge, & reg blood draw for PhenoAge on same day in Feb & from same venous blood draw at LabCorp, except Elysium's uses saliva which I did right before going to LabCorp. Notes: 🧵

I chose these 4 clocks for the same-day comparison since they are most respected & used clocks avail to consumers currently. Specific reasons for each:

@TruDiagnostic: team (eg @RyanSmithEpiAge, @VarunDw, @EverythingEpi) comes & presents at aging confs regularly, its the basis for the leaderboard Bryan Johnson's team maintains, they licensed DunedinPACE, & they provided an intrinsic vs extrinsic split per the mix change of immune cell types issue @alantomusiak from Eric Verdin's lab identified recently as a problem with most clocks.

Elysium briefly had @DrMorganLevine as advisor & she supposedly helped develop the 100k+ cpg site Index test, which should be robust to the test-retest variability issue that her recent PC clock paper identified as a problem with most clocks. (TruDiagnostic's clock is also supposed to be robust wrt this issue.)

@GlycanAge: team (@GordanLauc, @entreprylexia) also comes & presents at aging confs & its based on a completely different signal.

PhenoAge (also from @DrMorganLevine) is a gen2 clock that's widely used even clinically by notable longevity expert doctors, & it was easy to add with just a basic blood draw.

Not gonna give all details on which # from which clock but top lines results were all over the place. The 4 age point estimates: 20, 40, 50, 55.
This was a bigger spread than I expected.

2 tests give rate of aging & these also had a big range: 0.77 & 0.95.

TruDiagnostic's extrinsic & intrinsic age #s differed by 18yrs but that's not weird, but what I didn't expect was for the overall age estimate they provided to be outside of the range covered by these 2 numbers.

I had copied the results but no longer see these on their website.

TruDiagnostic licensed the systems aging methylation clock f/ Yale (PhD work of @rv_sehgal started under @DrMorganLevine; yes, she's in this thread a lot) & just launched it inc ability to add to earlier tests. Elysium did the same to their test earlier (I noticed in 2023). Elysium doesn't say how theirs was developed. Too bad @WyssCoray lab's recent proteomic systems clocks (being commercially developed by TealOmics) isn't avail to consumers, at least for now.

Elysium's has 9 systems: brain, liver, kidney, heart, blood, hormone, metabolic, & inflammation. TruDx's has those + lung & musculoskeletal.
Both give %iles. Elysium also aging rates.
System clocks potentially more clinically useful: can suggest areas for conventional dx tests.

The 2 systems clocks differed a lot. One's worst system (age 61, younger than only 12% of others) was not bad in the other (age 49, younger than 81% of others). One's best (44) was 3rd worst (54) in the other.
If we look at just whether each was above or below actual age they disagreed in 5 of the 9, with 3 of these 5 disagreeing by 7+ years.

Just as TruDiagnostic's overall age est was not within the range of its extrinsic vs intrinsic #s, Elysium's systems ages were sometimes odd relative to the overall est (on older pre-2024 tests):
Overall age est was outside the range of the min to max systems ages (on 2 earlier tests).

Also, on 1 earlier test Elysium's overall rate-of-aging estimate was also outside the min to max range of the rate-of-aging estimates of all systems.

There was more weirdness in some of TruDiagnostic's other reports. They report some conventional blood biomarkers. I had these tested by LabCorp from the same venous draw but got very diff #s:
Specifically, they provided estimates of HbA1C, creatinine, alk phos, fasting glucose, & albumin which differed from the labcorp measured values by 12-30%.

Ryan explained that epigenetic biomarker proxies (EBP) can differ from the actual marker they proxy, implying (but I'm not totally clear on this) that the reports numbers were EBPs not direct measurements of the markers from the dried blood spot. He further explained that EBPs are in some ways better as they are more stable than direct blood markers in many cases.

I still find the HbA1C discrepancy odd though as A1C is already a very stable approx of 3month glucose levels so any extra stability of an EBP is not an extra benefit there. LabCorp reported 5.2% (normal) but TruDx reported 6.2% (well into prediabetic). Fasting glucose numbers over many years agree with the LabCorp conclusion of not even pre-diabetic.

Summary take-homes:
The opacity of Elysium's test & participation in the community from the TruDiagnostic team plus their openness about whose clocks they are licensing makes me trust them a bit more, but the oddities in all this makes me not trust any particular # very much.

And I'm not sure how to value GlycanAge vs the methylation based clocks. I've heard it claimed that it mostly measures inflammation. Supposedly they will have a way to diagnose menopause with it coming. But otherwise, not sure what to do with GlycanAge estimates when they differ substantially from DNAm ones.

These tests are still prob useful for tracking changes over time longitudinally & it's great they seem to apply new analyses to old test data. I wouldn't trust changes in response to treatments being meaningful yet. Nor IMO are these ready for use in competitions/leaderboards.

I should also make some comments here about the appropriate use of rate-of-aging clocks vs point-estimates of bio age. Everyone should understand that pace estimates & age estimates respond to interventions differently depending on whether the intervention is a rate-of-aging modulator or rejuvenation (jumping some part of molecular state to that of an earlier age, eg infrequent senolytic tx). A pace clock is the better tool to see changes to aging rate shortly after a change. An age-est clock simply cannot see such a thing in a short term since the absolute diff in resulting bio age will be within the noise. It can only see a diff after a long time has passed (much more than 2 years). Eg it was wrong for some people to take GrimAge's lack of ability to distinguish treatment vs control of the CALERIE study after 2 years as indication of something wrong with that clock. It wasn't. See … (scroll up & down). The DunedinPACE estimated aging rate for treatment was 0.931 vs control 0.9535. A ~0.02 year/yr diff in aging rate over 2 years would result in a bio age diff of 0.04yrs. No way GrimAge could reliably detect that. Estimate used in above thread was overly generous and assumed a real diff in aging pace of 10%, but even for such a big pace change one wouldn't expect a point-estimate clock to be able to reliably detect a 0.2 year age diff after only 2 years of intervention.

Similarly, pace clocks can't really deal with rejuvenation. If one ages at a normal rate then jumps back 5 years worth of biological age in a 1-2 week treatment then ages at a normal rate again, a pace clock before & after wouldn't even see the backwards jump. It's no clear that such clocks would register big negative rates during the treatment either. (The short-term during such a treatment might even look like a faster rate of aging temporarily to them as it's really outside the scope of their training data.) So for things like senolytics, epigenetic reprogramming, removal of lipofuscin or AGE crosslinks or misfolded proteins, or a parabiosis like treatment or a regenerative med / stem cell or their secretions treatment or replacement of organ(s) w/ younger ones, etc. point-estimate clocks are more appropriate.

I should also note that most such rejuvenation therapies deal with only 1 kind of molecular change with aging and/or also only 1 tissue type or bodypart/organ. A combination of many such treatments to cover a wide variety of types/tissues/organs will prob be necessary to see meaningful changes to overall biological age, so lack of a single such treatment reversing any given clock is neither a condemnation of the treatment nor the clock, in humans anyway.

The right way to validate point-estimate clocks is to use rodent (or some other short lived species) versions not human ones. This allows rate-of-aging slowing interventions measured over meaningfully long (wrt lifespan of the species in question) timescales. Or one can use rejuvenation type treatments with known lifespan benefits. Eg someone should redo the transgenic p16+ ablation ("senolytic") studies (or use banked samples if there are still some around) and test the mouse version of GrimAge after vs before treatment & vs controls. Ideally, @LEVfoundation's extensive testing of its RMR studies will include mouse aging clocks used in this manner.

https://x.com/KarlPfleger/status/1624811768232181760

Another reply from follow-up discussion that is worth emphasizing here at the end of the main thread:
We aren't yet at the point where most of the DTC clocks can detangle mix changes (eg of immune cells).

All clocks will need to handle this eventually.

https://x.com/KarlPfleger/status/1808171664251015543