And of course, a huge thanks to everyone supervising this work – my awesome PhD advisor @jkpritch.bsky.social, a key lab mentor in @jeffspence.github.io (now at UCSF!), and our fantastic collaborator @gs2747.bsky.social!

Thanks for reading and please let us know what you think of our work!

(12/12)

There’s a lot more on the modeling in our preprint — its narrative arc starts with the expression of a single gene, and builds intuition from motifs up to entire GRNs. Feel free to give it a read!

doi.org/10.1101/2025...

(11/)

All told, this gives us cause for optimism about variance-based aggregation tests for mapping genetic effects, both in the context of expression (i.e., for trans-eQTLs) and complex traits (e.g., gene programs).

(10/)

To belabor this point, the typical gene in our best matched GRNs has ~tens of upstream regulators, and >95% of its h2 explained within two hops (i.e., by a peripheral [master] regulator or its direct regulators).

(9/)

Taken together, modularity and degree dispersion mean that the genetic architecture of gene expression is likely more sparse and pleiotropic than would be expected under a totally random GRN model (matched for median cis-h2%).

(8/)

Master regulators add motifs too, but they also make the GRN “shallower” by shortening path lengths between genes. This tends to decrease the % of trans-h2.

(7/)

It’s nice to know that about GRNs, but why do these properties matter? I’m glad you asked.

Modularity adds local structural motifs (i.e., triangles and diamonds), which can increase or decrease the effects of trans-eQTLs (depending on how many genes are activators).

(6/)

Long story short, some of our simulated GRNs look like real data.

These networks tend to have modular groups, master regulators, a high proportion of activators, and a fixed ratio of sparsity and regulatory strength.

(5/)

Inferring GRNs is tough, but it’s easy enough to simulate them.

Here, we use a linear model of expression on randomly generated DAGs, with the goal of finding GRNs that match the observed distribution of cis-h2 fraction.

(4/)

We’d probably be better at discovering trans-eQTLs if we knew the underlying regulatory biology, but GRNs are also hard to map. So what is there to do? (3/)

Expression QTLs are a nice way to map genetic effects onto genes (e.g., SNP → E → Trait). But a lot of expression variance (~70% of h2) is spread across the genome, ergo hard to discover statistically.

Data from pubmed.ncbi.nlm.nih.gov/31558840/

(2/)