sashagusevposts.bsky.social
Statistical geneticist. Associate Prof at Dana-Farber / Harvard Medical School.
www.gusevlab.org
646 posts
6,090 followers
662 following
Regular Contributor
Active Commenter
comment in response to
post
So selection on traits will induce LD. However, it will often be -ve LD between like-effect variants and so goes in the opposite direction as +ve AM (and so biases estimates of AM down). Carl & I discuss this some here, and give some numbers for plausible params:
journals.plos.org/plosbiology/...
comment in response to
post
The actual GRM they construct is also specifically restricted to between-chromosome effects, which adds a wrinkle.
comment in response to
post
Yeah I think GxG should show this pattern regardless of LD. Not sure what happens over generations or after ascertainment though.
comment in response to
post
Yeah I wonder what would happen if you ran this in your MOBA PGI analysis and used the DRGRM as a random effect for bias correction.
comment in response to
post
Yeah, agreed. The extra GRM they use looks very similar to an epistasis GRM. Pop strat will also induce correlations (as they show and investigate).
comment in response to
post
What does predict faster learning rate? You already know.
comment in response to
post
An interesting twist on this study would be to look at whether high cognitive function is also not predictive of faster cognitive gain in early age (i.e. "learning rate"). This too has been noted in observational analyses of schooling and work performance.
gusevlab.org/projects/hsq...
comment in response to
post
Prior work (pubmed.ncbi.nlm.nih.gov/20973608/): "Higher intelligence from early life was apparently protective of intelligence in old age due to the stability of cognitive function across the lifespan, rather than because it slowed the decline experienced in later life."
comment in response to
post
Potential implications:
-- Evidence against high "cognitive reserve" as a protective factor against age-related decline
-- Genetic influences on baseline function are distinct from influences on decline
-- Decline is largely driven by environmental factors (or rare genetics)
comment in response to
post
Seems to run counter to this 2014 prediction by Deary and Plomin regarding the genetics of intelligence and assortative mating:
nature.com/articles/mp2...
comment in response to
post
Main finding: Sib regression heritability (h²) analysis also established substantial family effects (c²):
1️⃣ high h² for IQ, BMI, height
2️⃣ decent h² for risk taking and well-being
3️⃣ c² > h² for education, labor market, substance abuse
⚠️ wide standard errors!
/2
comment in response to
post
Really nice work, curious if you guys have tried @paturley.bsky.social 's IBS based approach, which can disentangle A from AxC in some scenarios.
comment in response to
post
Here's a paper with some other nice visualizations:
elifesciences.org/articles/60107
comment in response to
post
Data from the SGDP:
comment in response to
post
It's obv. super complicated- but the GWAS for most cancers are just based on being diagnosed with the cancer. The genetics of metastasis are likely very different- and cancer dependent. Ian Tomlinson and others have described a 'just enough model'. Example are POLE mutant endometrial-
comment in response to
post
Yeah that makes a ton of sense, especially given metastases often seem to be seeded well before diagnosis but remain dormant (presumably due to immune/microenvironment interactions)
comment in response to
post
Yes, I'm just surprised the saturation is essentially immediate and for pretty much all traits that were tested.
comment in response to
post
For cancer, I would have expected some of the genetic mechanisms to be *waves hands* DNA damage repair, and for this to be relevant both for transformation to cancer and for aggressiveness of cancer.
comment in response to
post
Thanks for reading and discussing!
comment in response to
post
Prior twin h2 estimates for heart failure are hard to find, but death from heart disease was estimated to be 40-60% heritable in twins (with all of the usual caveats about twin studies). 🤔
comment in response to
post
There's the lavaanPlot package for fitted models (lavaanplot.alexlishinski.com/articles/int...) or do you need it to be directly in the AI app?
comment in response to
post
Have you seen this:
maartengr.github.io/BERTopic/ind...
comment in response to
post
These points are mostly glossed over in the article, which focuses more on fantastical GATTACA scenarios and social media controversies. I get that those are more fun to cover and read about, but I think it is worth first resolving the question: does this tech even work at all? /fin
comment in response to
post
Given that companies like Genomic Prediction have reportedly already delivered ineffective predictors to their customers, the continued lack of transparency on the source and accuracy of these predictors -- even after criticism from genetic groups like the PGC -- should be a scandal.
comment in response to
post
Moreover, the whitepaper does not even report the predictive accuracy within-families -- the only metric that is relevant to the expected gains from IVF screening (a "within family" process). See details here: theinfinitesimal.substack.com/p/some-notes...
comment in response to
post
However, the Orchid whitepaper on schizophrenia cites a single genome-wide association study (which is where the data for genetic risk scores must come from). That study was conducted by ... the Psychiatric Genomics Consortium.
comment in response to
post
Orchid also continues to obfuscate about the source of their predictive models. When the Psychiatric Genomics Consortium criticized the use of their data for embryo selection, an Orchid spokesperson claimed that their data was coming from elsewhere. But would not say where!
comment in response to
post
For an underlying continuous trait, small reductions (e.g. BMI 41 to BMI 40, a reduction of ~7lbs) can be marketed as large relative "risk" reductions (e.g. 50%) because people move from just above the "obese" threshold to just below it. More here: theinfinitesimal.substack.com/p/science-fi...
comment in response to
post
The company Orchid is reportedly advertising substantial risk reduction for conditions like obesity. Obesity is an underlying continuous trait, which is then dichotomized to define the case population (e.g. BMI>40). Relative risk reduction can thus be highly misleading.
comment in response to
post
But it is of course possible that the predictor was actually WORSE than noise, either because it was actually picking up a correlated condition (such as autism) or because customers were favoring the (erroneous) prediction over other, actually effective, markers of embryo health.
comment in response to
post
First, the company Genomic Prediction was reportedly providing consumers with an IQ predictor that they later realized does not work, and removed. This is a big deal. Most likely the predictor was just noise, as I've written about before: theinfinitesimal.substack.com/p/genomic-pr....