Roshni Patel

Thank you Alex! Excited to see our paper published in @nature.com ! Huge thanks to @jeffspence.github.io , @tkyzeng.bsky.social , @emmamarydann.bsky.social, @nikhilmilind.dev, @marsonlab.bsky.social, @jkpritch.bsky.social, and all the members of the Pritchard and Marson labs for your enormous help!

After time in the Bay Area, I’ve started a new role as Lecturer in the Department of Allergy and Rheumatology at the University of Tokyo. We’re the group of clinicians who see patients with autoimmune diseases, while researching new treatments and patient stratification. (continued)

Within family prediction of psychopathology Honey it's fine. The ratio with my shoe size is actually pretty good.

“What I do think is that it has become normalized in modern behavioral genomics to do what you have to do in order to make every result, no matter how small, look like a win for team genetics”

ericturkheimer.substack.com/p/within-fam...

I taught Genetics again this year. We need to include discussion of the problematic history of our field, especially as the claims of eugenics are once again centered in our political discourse. Last year I wrote this piece, explaining my reasoning and approach 🧪 1/n
www.cell.com/trends/genet...

Intracellular interactions shape antiviral resistance outcomes in poliovirus via eco-evolutionary feedback - Nature Ecology & Evolution A model of intrahost poliovirus replication shows that, after several rounds of replication, pocapavir, a poliovirus capsid inhibitor, collapses viral density, preventing intracellular interactions th...

My first lead author paper is out with Ben Kerr and @alisonfeder.bsky.social! We found that making an antiviral too strong can sometimes make resistance easier to evolve. This has implications for how we design drugs, choose doses, and think about viral evolution in the face of treatment. (1/n)

Text from article advocating for transparent recognition of racist past in our disciplines to encourage, not discourage, more diverse engagement.

And this. My experience also supports that being *more* open about racist, sexist, and ableist histories of our fields is more engaging, not less, for students (and faculty) from minoritized backgrounds. Transparency can only enhance rigor.

12 former commissioners of the FDA came together to write a Perspectives piece for the New England Journal of Medicine; raising our concerns about recent changes to vaccine approval policy at the FDA and its implications for patients and public health.

yes don't worry we'll write the rules such that DR EVIL passes with flying colors

I half agree but also I think papers should stop inventing acronyms for niche biological phenomena. tools are fine! but why are we making the reader do the mental legwork of mapping a string of letters to a phrase, and a phrase to a concept, all for something that nobody in the field will reuse??

There's a bit more in this thread about how the GRM relates to existing methods -- happy to chat more as well, I'd of course be curious to know what you think

Re: your Q about types of confounding - the GRM helps control for pleiotropy too! Using the GRM, we separate trait correlation mediated through genetics from that independent of genetics (via bivariate GREML, essentially), and doing so allows us to control for pleiotropy-induced confounding

...and since the environment of Black British individuals in UKB isn't identical to that of individuals in AoU, I'm not sure I would expect to estimate the same effect. But it's possible!

Thanks Gen! Definitely curious about analyzing in All of Us. I do think exposure effects should be interpreted as cohort-specific rather than population-specific, since I imagine what's happening is social and environmental context modifying the effect of e.g. loneliness on WBC

It was a total pleasure to work with @roshnipatel.bsky.social on this, who really led the charge in all respects. Anyone interested in learning about the intersection of population genetics and statistical genetics should check out her new lab in Oregon!

here's hoping! and thanks for the kind words!

my delusional pipe dream this entire time has been "maybe this paper means behavioral genetics will stop estimating polygenic scores"

Unifying approaches from statistical genetics and phylogenetics for mapping phenotypes in structured populations Statistical genetics and phylogenetics have different methods of finding associations with phenotypes while controlling for ancestry. This paper shows that the standard approaches in these fields are ...

Big thanks again to my co-authors @jgschraiber.bsky.social @docedge.bsky.social and Matt Pennell for their work on this, as well as their excellent PLoS Bio paper (doi.org/10.1371/jour...) that motivated so many of the ideas and approaches here (10/10)

This paper was a fun one to write, in large part because it highlighted so many connections across disparate lines of work in epidemiology, stat gen, and phylogenetics -- it was a good reminder that talking across our silos is hard but there's a lot to be learned by doing so (9/10)

Phylogenetically conservative trait correlation: Quantification and interpretation The phylogenetically conservative share of a trait correlation ought not to be excluded from consideration as potentially ecologically functional. The conservative trait correlation (CTC) concept tre...

In fact (!!) the model we propose is analogous to one recently described by Westoby et al. in the context of controlling phylogenetic confounding: doi.org/10.1111/1365... (8/10)

Finally, our approach was inspired in large part by similar problems in phylogenetics: researchers often use interspecific data to test for an association between two traits, and must account for confounding that results from species’ shared ancestry (7/10)

The differences between these methods become apparent in *real world settings*, where some traits might be difficult to measure at biobank scale, and some populations are systematically underrepresented across databases (6/10)

The correspondence between LDSC- and GRM-based estimates of heritability is well-known in statgen, and here it means PENGUIN- and GRM-corrected estimates of exposure effect are equivalent in idealized settings where large sample sizes are available for all traits and populations (5/10)

More recently, a method named PENGUIN sought to control genetic confounding by using LDSC to estimate the genetic correlation between risk factor and outcome, and subtracting that off from the total phenotypic correlation (4/10)

In other words: the reason that PGS-based methods fail to control genetic confounding is not because of a flawed model -- the true PGS could successfully control genetic confounding -- but because of a methodological limitation (we can't estimate the true PGS, and we'll never be able to!) (3/10)

A lot of prior work attempted to control for genetic confounding by including the PGS for the exposure/risk factor as a regression covariate. We show that the model assumed by PGS-based approaches is equivalent to the model implemented in our GRM-based method (2/10)

As promised, a longer thread on what I consider to be some of the most interesting and important contributions of this paper (1/10)

But because the thing we're looking at here is the genetic covariance of both traits *relative to* the genetic variance of one trait, we wind up being a bit insulated from situations where both traits are stratified in similar ways

Totally -- we did some simulations here but definitely didn't exhaust the entire search space. I think the tldr is that if there's a huge difference in the way that pop strat impacts the two traits you're studying, you're going to run into problems.

This was a super fun project to do my small part on. @roshnipatel.bsky.social really helped me understand what we're doing when we're trying to "control for genetics" when assessing associations between interventions and outcomes, and how a lot of the standard ways to do so are simply inadequate.

Using our method, we analyze the effect of loneliness on complex traits in three subsets of the UK Biobank -- the kind of population-specific analysis that would be difficult do with existing methods. Lots of interesting empirical work lies ahead, and we'd love to hear questions/comments! (6/6)