On this point, here's a plot of total cholesterol from lipid kits measured by a clinical biochem lab in Birmingham as part of lipid EQA. There's 'assay' level variation in lipids that clinicians do not even think about as they use LDL etc for risk assessment (!) journals.sagepub.com/doi/10.1258/... - ThreadSky

mikeinouye.bsky.social • 17 days ago

On this point, here's a plot of total cholesterol from lipid kits measured by a clinical biochem lab in Birmingham as part of lipid EQA. There's 'assay' level variation in lipids that clinicians do not even think about as they use LDL etc for risk assessment (!) https://journals.sagepub.com/doi/10.1258/acb.2007.007120?icid=int.sj-abstract.similar-articles.3

Comments

shaicarmi.bsky.social•16 days ago

Interesting, I thought measuring LDL-C was a solved problem. (Or are things better now?)

Anyway, I think the problem with PRS is that (in contrast to LDL-C) there's no "true" risk that just needs to be measured better. Variability is not only technical. It's also which GWAS/method is used etc.

timfrayling.bsky.social•16 days ago

not sure I get the second point. Sure there is variability and we need to do better in people of non Euro ancestry, but common genetic variants and their association with disease are precisely measured now and becoming more so.

shaicarmi.bsky.social•16 days ago

I'll clarify that I'm not against adding PRS to risk models, despite the variability. Quite the opposite. And I don't think the variability is a statistical problem (I agree with @mikeinouye.bsky.social).
(But it's certainly a serious clinical/practical problem.) 1/3

shaicarmi.bsky.social•16 days ago

Re. your question, let me try to refine the argument. For LDLC, there is an actual number of molecules in the sample. We can (hypothetically) count them and always reach the same answer. A genetic risk is a statistical construct, which depends on assumptions. 2/3

shaicarmi.bsky.social•16 days ago

For example, it depends on the definition of the disease, and on which ages, environments, and sub-populations were used to estimate the risk. We have to make decisions on what we're estimating, even when sample sizes are very large. 3/3

mikeinouye.bsky.social•16 days ago

With PRS we assume there is a true fixed genetic value for each person. We ourselves actually create the noise in estimate ourselves (gwas, methods etc). Other conventional RFs (LDL, BP etc) there's that + natural spatiotemporal variation. Much harder to hit a moving target...

mikeinouye.bsky.social•16 days ago

Its also worth pointing out that people who are shocked by individual variation in PRS perhaps are forgetting this is THE premise of many popular PRS methods (iPGS, meta scoring etc), widely used for at least 7 years now… also its a extension of basic stats (regression dilution bias, ensembling etc)

shoaclarke.bsky.social•17 days ago

Just to add my 2 cents. 1, PRS absolutely SHOULD pass a higher bar than other approaches to risk assessment. 2, LDL-C is fundamentally different than PRS and is probably not the right comparison/benchmark.

mikeinouye.bsky.social•17 days ago

I'd ask why re both as thats the key bit but this isnt the right platform for why (unless you like long threads that properly justify). Is there a link/literature that thoroughly explains your thoughts?

shoaclarke.bsky.social•17 days ago

😂 agree, probably hard to do justice to the discussion here. Point 1 is my perspective based on how clinicians & patients understand/communicate risk. I think genetics is unique compared to other risk tools. But this is not based on data, just my personal experiences.

shoaclarke.bsky.social•17 days ago

Re: point 2. The effect of LDL is causal, modifiable, & generalizable. PRS are fairly abstract. I have hope for value of PRS but the value will be different than traditional RFs where we have well-understood targetable causal mechanisms.

mikeinouye.bsky.social•17 days ago

I can understand those reasonings. For 2, I do think PRS does or soon will tick the causal, modifiable and generalisable boxes. It is to some extent early days especially wrt LDL (a grand dame of molecular risk factors)

mikeinouye.bsky.social•17 days ago

Also LDL is a relatively interpretable biomarker but isnt any more associated to CAD than PRS. The abstraction of PRS comes from a mixture of pathways feeding into CAD, some via arterial lipid accumulation (tagged by LDL) but others not... q is to what extent PRS captures causal & modifiable biology

Comments

Posting Rules

Reply