Post

It's been 30 years (!!!) since the conceptualization of GWAS. GWAS has driven tons of discoveries, but with recent experimental developments (CRISPR, single-cell omics), we thought it was time to think about how to drive the next 30 years of human genetics. doi.org/10.5281/zeno...

We argue that the goal of human genetics should be to build *Trait Causal Models*. TCMs are quantitative, predictive models that trace the biological processes from genes --> traits.

TCMs should be: - Quantitative: we should move beyond just cataloging associations and instead we should estimate _by how much_ and _in which direction_ a variant or a perturbation would affect the trait

TCMs should be: - Mechanistic: a model capturing the mechanisms by which genes affect a trait should be able to convert knowledge about how a perturbation or variant affects an intermediate phenotype into predictions about how it would affect the trait.

To build TCMs will require 3 distinct "data pillars", each with complementary advantages 1. Genetic association data (e.g., GWAS) 2. Perturbations in model systems (e.g., perturb-seq in cell lines) 3. Human subjects omics profiling (e.g., scRNA-seq from cases and controls).

There have been incredible advances across all 3 of these data pillars, but we discuss what more needs to be done to move from these pillars to building TCMs.

Key challenges include: - Extracting directional, quantitative effects of _genes_ on traits from GWAS data - Choosing cell types and conditions and types of perturbations in screens - Choosing the right level of abstraction for modeling the path from variant --> trait

We discuss all of this and more in the piece. It was a lot of fun thinking through how to use the data that we have and to try to crystallize what the goal of human genetics should be. We're hoping to spark a conversation so please let us know you're thoughts. doi.org/10.5281/zeno...

This piece grew out of many conversations about where the field is going, sparked by work from @jengreitz.bsky.social's group, and from @jkpritch.bsky.social's group in collaboration with @marsonlab.bsky.social (led by @minetoota.bsky.social). Thank you all for all of your insights!

And a huge thank you to @emmamarydann.bsky.social for co-leading the effort and to @jkpritch.bsky.social for the vision!

As an addendum, AI slop has made it difficult to post perspective pieces on preprint servers, so we're sharing ours on Zenodo. We did use LLMs to catch typos etc..., but rest assured that human minds conceived the work, human fingers typed it up, and human eyeballs proofread it.

Any particular reason you didn’t just use the decades old spell check in Word or something instead of an LLM?

LLMs are pretty good at flagging those kinds of duplications. LLMs can also flag awkward phrasing, etc… So I would really view LLMs and something like spellcheck as filling separate niches.