Shiron Drusinsky

PhD candidate @UCSF Bioinformatics and @UCBerkeley Computational Precision Health

[Shiron Drusinsky @Shirondrusinsky]

Excited to announce a new preprint that investigates the ability of sequence-to-function models (S2F; e.g., AlphaGenome) to identify causal expression modifying variants. (1\n)
https://t.co/1iktWUhfNI

[Anshul Kundaje @anshulkundaje]

https://t.co/ykNzBcleaC
Practical utility of sequence-to-omics models for improving the reproducibility of genetic fine-mapping 2/

[Shiron Drusinsky @Shirondrusinsky]

@StevenSalzberg1 @TuXinming @zhanxw Shameless plug, but maybe you’ll find this paper helps reconcile your finding/experience with our claim that alphagenome is not slop and can be quite useful (but also often is not. noncoding variant interpretation is very hard)

https://t.co/1iktWUhfNI

[Shiron Drusinsky @Shirondrusinsky]

@StevenSalzberg1 @TuXinming I agree with @TuXinming. The paper may be over sensationalized but it’s far from slop

[Shiron Drusinsky @Shirondrusinsky]

@StevenSalzberg1 @TuXinming Suggesting it’s more likely to get folks to give up on strong leads than pursue spurious ones

[Kevin A. Scott, PhD @Sculpturatus]

at ucsf

[gen0m1cs @gen0m1cs]

Causal variant underestimation is a major overlooked driver of sequence-to-function model underperformance.

https://t.co/chJQGCOUfR

[Anshul Kundaje @anshulkundaje]

Valuable work dissecting some of the key challenges with variant effect prediction on gene expression using supervised sequence models.

[Shiron Drusinsky @Shirondrusinsky]

and their prevalence and causes**

[Shiron Drusinsky @Shirondrusinsky]

This suggests issues for fine-mapping and variant prioritization in most loci. But I also find that the very strongest predictions are highly reliable and supported by the data, suggesting AlphaGenome “hits” can often be trusted. (6/n)

[Shiron Drusinsky @Shirondrusinsky]

I hope this work helps shed light on (esp. for less advanced users) important S2F failure modes -- and their prevalence causes -- and when variant effect predictions can and can’t be trusted (11/11)

[Gladstone Institutes @GladstoneInst]

AI holds huge promise for finding disease-driving genetic mutations, but top models frequently underestimate their impact.

New research exposes this underlying flaw and provides a roadmap to build smarter models for personalized medicine.

https://t.co/9WnSZqtDrD

[Jacob Schreiber @jmschreiber91]

A very important issue, glad to see this out!

[Shiron Drusinsky @Shirondrusinsky]

In these cases AlphaGenome doesn't prominently upweight any other candidate SNV that could explain the eQTL association. So this isn't disagreement with fine-mapping; it underestimates all candidate causal eQTLs relative to nearby non-eQTLs. (5/n)

[Shiron Drusinsky @Shirondrusinsky]

Excited to announce a new preprint that investigates the ability of sequence-to-function models (S2F; e.g., AlphaGenome) to identify causal expression modifying variants. (1\n)
https://t.co/1iktWUhfNI

[Shiron Drusinsky @Shirondrusinsky]

It seems that yes, most causal eQTLs are underestimated by AlphaGenome (and likely other models), not just in absolute magnitude but also relative to nearby ones with no detectable expression association. Often severely enough that the causal eQTL is effectively ignored. (4/n)

[Shiron Drusinsky @Shirondrusinsky]

Are S2F models ignoring most causal SNVs? Or maybe predicted effects are small, but are still stronger than nearby non-causal SNVs? Maybe fine-mapping is wrong and they aren’t causal, and models upweight other variant(s) instead that explain the observed eQTL association?

[Shiron Drusinsky @Shirondrusinsky]

This paper was in part motivated by figures like this, found in most S2F papers. Predicted vs observed effect sizes for putatively causal fine-mapped eQTLs. Overall correlation looks good, but most points sit near a predicted effect of 0. What’s going on there? (2/n)