Thread by @ryanlcollins13

At long last, our WGS association study in autism quartets is live @NatureGenet :
nature.com/articles/s4158…

Several points worth emphasizing from this paper.

Twitter explainer below 👇, also great N&V piece by Wray & Gratten here:
nature.com/articles/s4158…

1/15

First and most importantly: huge multi-year collaboration b/t @TalkowskiLab & many others.

Particular congrats to @DonnaWerling, Harrison, @joonomics, @m_sto, and Lingxue, who spearheaded the analyses from start to finish!

Now, on to the science...

2/15

First point: WGS is a *lot* of data. You are dealing with ~6 BILLION nt per diploid genome.

As such, WGS association studies must be careful about multiple testing burden and adjust p-values accordingly.

3/15

Easy to get nominal p<0.05 to support cherry-picked pet hypotheses (look no further than replication rates of early candidate gene studies vs rigorous p<10^-8 GWAS associations for proof of this).

4/15

For example: we find equally compelling trends for de novo SNVs as nominally significant in cases and controls.

Most signif. in cases: DNVs in conserved bp within postsynaptic density genes

Most signif. in controls: noncoding DNVs in genic enhancers near constrained genes

5/15

A priori, good luck picking which one was more likely to be signif in controls vs cases!

6/15

But it's not all doom and gloom. Conservative WGS association is possible: correlation structure b/t different genome annotations permits estimation of # independent tests performed.

Meet the category-wide association study (CWAS) -- great work from Devlin & Roeder labs

7/15

Here, we estimate 4,213 independent hypotheses, and estimate that the multiple testing burden asymptotes at ~10k tests as you add more annotations. From this, can adjust p-values accordingly.

8/15

This study is also in part a plea to the field to be realistic with power calculations, particularly for common and complex diseases. Must weigh expected effect size with incidence of relevant mutation.

9/15

We clearly see our study of >2k genomes (519 cases) was underpowered to detect even the strongest known *coding* signals in ASD, let alone noncoding signals (where prior on effect sizes is << coding space)

10/15

So why WGS? Well, one key advantage = capturing structural variation (SV).

We developed an SV pipeline that yields close to 6k SV (>50bp) per genome at FDR < 1%.

>5k high-quality SV can be routinely captured from garden-variety 30-40X Illumina WGS.

11/15

The _vast_ majority of SV are small & rare. Analyses restricted to large or common SV get the tip of a much bigger iceberg (but still better than nothing!)

12/15

Further, we find a couple hundred de novo & mosaic SV in children one would want to consider in WGS studies. Molecular confirmation rate > 97%

Some are even likely relevant to autism, like a de novo balanced translocation of GRIN2B or a de novo 4-exon CHD2 deletion.

13/15

In summary: WGS is an exciting/tantalizing data type that is on track to revolutionize aspects of human genetics.

That said... please, please treat significant noncoding associations from small studies with skepticism until multiply replicated in larger cohorts.

14/15

/my first Twitter explainer

Was awesome to work on this over last ~3 years. Great group of collaborators! Looking forward to much more WGS soon.

~Fin~

15/15

unroll (cc @chromosomeOMICS). Hope it works this time!