Профиль

Excited to share 2 important milestones in my academic journey- Our work on “Pervasive interactions between exposures & polygenic risk can inform more effective clinical & behavioral interventions,” showing PGSxC interactions & their utility, is out in #NatureGenetics @genomestake.bsky.social

I looked into it but nothing specifically met my needs, which are primarily managing stuff - grant - project alignment, lots of active projects in development/ideation, and have agents automatically looking for new publications/conferences.

Will post this on github once I'm sure there are no bugs, but the LLMs are so good at this kind of stuff right now that I would encourage people to just make their own apps to their unique specifications.

Revealed preference by the AI

Over the weekend GPT built me a lab management app that is more effective than anything I've ever used (a bunch of spreadsheets in Dropbox).

Very exciting new method for inferring latent disease trajectories, led by the tireless Dr. Urbut with Natarajan and Parmigiani labs. If you're considering a postdoc in this area talk to Sarah! www.nature.com/articles/s41...

I was actually surprised at how low these numbers were, I'm guessing it will blow up in 2026. I don't really see any issue with having AI pre-fill summary sections as long as the applicant verifies the text and is responsible for the content.

The interactive analysis and classifier was implemented and coded with GPT-5.6-Sol in about an hour and a half. /x

You can experiment with the classifier and term flagging in a static browser (no data is transmitted), including some sample high-scoring abstracts from FY2000. These scores do not reflect grant fundability or intent and are correlative only.

And, in fact, while the masked classifier and keyword search identified a similar number of abstracts prior to 2025, the reduction was much more severe for keywords being used than for themes after keyword masking.

I was curious to see if these reductions were primarily keyword level or reflected broader changes in theme, so I trained a simple linear classifier of whether an abstract contained a "DEI term" after masking the terms out. A simple TF-IDF linear model worked pretty well:

These proposals are often submitted to diversity supplements and minority access training programs, which are broadly popular with voters and were effective at getting participants into biomedical research (statnews.com/2026/06/17/n...).

Terms that decreased in frequency? Unsurprisingly, mentions of "minority" and "gender" are drastically down in FY2025 and even lower in the first half of FY2026 -- well below the average over the prior 25 years.

AI? I ran a quick Pangram analysis on a sample of abstracts and it's climbing to a rate of 3-5% in FY2025. This does not necessarily reflect full AI grant-writing since abstracts are mostly seen as administrative.

More abstracts are also using fancy words like "leveraging" , "crucial" , "ensuring" , "urgent need" , etc.

Terms that that went up in frequency in 2025: "immune", "transcriptomics", "spatial transcriptomics" -- experiencing steady increases over the past decade -- as well assorted phrases related to radiation exposure (presumably a funding call).

How are NIH abstracts changing with the new administration? A short 🧵 on common term trends over time: sashagusev.github.io/nih-abstract...

We're excited to have an incredible lineup of invited speakers at #IGES2026 this fall! Josée Dupuis, @mikeinouye.bsky.social, Sriram Sankararaman, and @sashagusevposts.bsky.social will present on cutting-edge topics in complex disease genetics, from multi-omics integration to biobank-scale methods.

Anyone have a recommendation for tracking and organizing AI agent work on a codebase? Linear.app? Ideally with an emphasis on organization.

Great article, and thank you for going down the rabbit hole (fox hole?) of the Belyaev experiments which are an interesting contrast.