Thread by @ewanbirney

A group of us (@minouye271, @JenniferRaff, @aylwyn_scally @AdamRutherford and myself) have written a piece on the language we use in genetics; untangling from previous sometimes racist language and being more precise and less harmful. We welcome feedback. arxiv.org/abs/2106.10041

There are some straightforward 'stop using this term' aspects (the use of "Caucasian" for example); there are some complex "what does this term mean" (ethnicity labels, the ethnicity/race duality in US vs just ethnicity in UK / Europe) and then technical stuff on GWAS >>

The technical piece is about how we describe the common place GWAS protocol of subselecting a group people in cohorts for association analysis; a reminder that the standard process has two steps to achieve pseudo-randomisation of non-genetic factors to genetic factors

Step 1. Calculate genetic principle components and select a group of people with lower observed genetic variance in the cohort being studied (aka, the "blob" in PC1/PC2). Step 2. Add many genetic components into the linear models to "model out" any residual non-genetic effects.

This two step procedure works robustly (works means... the statistical test fits the expected null distribution most of the time, aka, "qq plot is good" or "LDscore based genomic inflation is low" and often some interesting hits are found and nearly always replicate) ...but...

We have to describe in our papers both what we are doing here and also the resulting subset of the cohort. We tend to sort of just gloss over the what we are doing bit and then get into all sorts of trouble in coming up with concise ways to talk about the subset of the cohort.

We advocate using *more* technical language to be clear about this procedure, eg "The European-associated PCA
cluster, which aims to minimise variation in non-genetic factors and genetic factors". The "European" label here is a cultural label, *not* a genetic label

We are pretty convinced that the routine use of this two step procedure and its basis in genetic principle components tends to make people think this sub-group is being defined to control genetics, when it is at least as important that it is controlling non-genetic factors

This is because most societies have complex social stratification and groupings, and much of this is related to recent migrations, either forced and horrible (eg, slavery) or voluntary (eg, economic migration). It's a big part of human society.

This "PCA trick" leverages the recent ancestry signals to select a group of people with a more consistent non-genetic environment, mainly due to social factors. This is cute and clever, but we have to be super-careful in not implying these groups are genetic.

I find moving between farm animal/laboratory genetics and human genetics very informative here. In these settings we can bring stronger pseudorandomisation (farm animal) or explicit randomisation (laboratory) assumptions to our genetics studies.

As such, one can deploy more sophisticated statistical models (eg Linear Mixed Models) which are fragile at best in the human setting where recent ancestry correlates with many phenotypes for plenty of reasons which are *not* genetic.

For most non-geneticists, this all gets a bit bamboozling, which is... because it is a bit complex... but we often sort of slouch back into genetic language (eg "genetic background" in human genetics paper is often not about *genetics* ... it is about this social stratification)

In the paper we have forced ourselves to be explicit in recommendations (I am inspired by a conversation here with @genemodeller that he is very "biddable" in this area) knowing that we will definitely get feedback on this - so please do comment!