Adam Auton

GitHub - 23andMe/PRSformer Contribute to 23andMe/PRSformer development by creating an account on GitHub.

Preprint: www.biorxiv.org/content/10.1...
Github: github.com/23andMe/PRSf...

Feedback welcome!

Huge congrats to the team: Payam Dibaeinia, Chris German, Suyash Shringarpure, and Aly Khan on getting this out the door. Come by Poster Session 1 on Wed 3 Dec 11 a.m. - 2 p.m PST at #NeurIPS2025 San Diego!

We show that non-linear gains emerge for immune-
mediated diseases at N > 1M. Harnessing phenome-wide (NxGxD) models seems a fruitful direction to borrow information across traits.

Core design: Neighborhood attention (O(L) complexity/layer) models the genome in ~100kb intervals, akin to LD blocks. By stacking layers, the model learns local dependencies first, then integrates long-range info and makes predictions feasible at genome scale.

We trained PRSformer at 23andMe Research Institute on population scale data (N > 1M persons; G > 100k variants; D > 10 autoimmune/inflammatory traits). PRSformer significantly outperforms linear baselines and summary-statistic methods (LDPred2) derived from the same cohort.

PRSformer: Disease Prediction from Million-Scale Individual Genotypes Predicting disease risk from DNA presents an unprecedented emerging challenge as biobanks approach population scale sizes (N>106 individuals) with ultra-high-dimensional features (L>105 genotypes). Cu...

Delighted to see our method, PRSformer, at #NeurIPS2025! PRSformer is AI model for population-scale disease-risk prediction from individual genomes. It lays the groundwork for phenome-wide risk prediction.

www.biorxiv.org/content/10.1...

PRSformer: Disease Prediction from Million-Scale Individual Genotypes https://www.biorxiv.org/content/10.1101/2025.10.26.684578v1

YouTube video by My Robot Teacher How Professors Are Responding to AI: Resistance Is Futile (and Brief) | My Robot Teacher Episode 1

The brilliant Sarah Senk, together with Taiyo Inoue, has just launched a podcast that explores the implications of AI for higher education: My Robot Teacher. Please give it a listen! #MyRobotTeacher #HigherEd www.youtube.com/watch?v=H2Ta...

YOUR 2024/25 CARABAO CUP WINNERS 😍

On the x-axis is every human gene, ranked by number of publications containing mention of the gene name. Lots left to discover...

Postdoc - Autism Research – 23andMe Careers Read about our mission-based culture, look up open positions and check out the perks of working here.

We're hiring a postdoc to help shape our autism research program; please consider applying.

www.23andme.com/careers/jobs...

Opinion | The Science of Blue Zones and Extreme Longevity Is Deeply Flawed Some of the claims behind the longest-lived people are simply improbable.

"the science of extreme longevity continues as an immense joke."

www.nytimes.com/2025/01/20/o...

Scientist / Senior Scientist, Statistical Genetics – 23andMe Careers Read about our mission-based culture, look up open positions and check out the perks of working here.

We're looking for a talented statistical geneticist to come work with us!
www.23andme.com/careers/jobs...

Scientist / Senior Scientist, Statistical Genetics – 23andMe Careers Read about our mission-based culture, look up open positions and check out the perks of working here.

We're looking for a talented statistical geneticist to come work with us!
www.23andme.com/careers/jobs...

Site-saturation mutagenesis of 500 human protein domains - Nature Large-scale experimental analysis of Human Domainome 1, a library containing more than 500,000 missense mutation variants across more than 500 human protein domains, reveals that 60% of pathogenic mis...

Humans tend to inherently believe that context matters. But context doesn’t seem to matter all that much in genetics.

Epistasis between mutations doesn’t seem to influence their stability in this amazing saturation mutagenesis paper.

www.nature.com/articles/s41...

So given that, can you spot the error right at the start of this Wikipedia article on HERC2?

"HERC2 is a giant E3 ubiquitin protein ligase, implicated in DNA repair regulation, pigmentation and neurological disorders."

🙋‍♂️

Good news, Altmetric has now started watching BlueSky for mentions of publications. And by the way, provides an easy comparison between this and the old site for a recent preprint of mine which I posted simultaneousl at both. Numbers speak by themselves !

Yesterday was a hard day at 23andMe, and we said goodbye to a number of tremendously talented colleagues. If people have job openings in the genetics space that they'd like me to share with the impacted folks, please do post here.

Today seems like a good day to share this Scientific American story on how vaccines have saved more lives throughout history than any other intervention

www.scientificamerican.com/article/see-...

Hello #ASHG 2024!

I want to try something new at #ASHG24 this year: I'm going to block some time on Friday afternoon to meet with any trainees who would be interested to chat on any topic.

We demand examples!

Long COVID #GWAS preprint identifies #HLA class II associations | #23andMe www.medrxiv.org/content/10.1...

I made a starter pack full of statistical genetics (adjacent) scientists. it covers all flavours of behaviour, psychiatric, social science and population genetics people. One click follow all of em! let me know if I missed key people!

Thank you! Super useful.

Large language models identify causal genes in complex trait GWAS medRxiv - The Preprint Server for Health Sciences

A super fun project. Congrats to Suyash Shringarpure, Wei Wang, Sotiris Karagounis, Xin Wang, Anna Reisetter, and Aly Khan on getting this out the door. Feedback very welcome!

www.medrxiv.org/content/10.1...

Definitely a fun result that I would not have expected! It's very early days, but it is exciting to think how LLM approaches could be combined with approaches that rely on functional data to incorporate prior knowledge of biology. Could we get the best of both worlds?

Interestingly, you can probe the internal embeddings of these models, and we found that the causal genes tend to be 'proximal' to the phenotypes that they influence in embedding space. So they do seem to be learning some relationship between these two concepts.

Nonetheless, the LLMs also have biases; they tend to favor genes with lots of existing literature, which perhaps isn't surprising given how they're trained. They also struggle to identify causal genes in loci containing large numbers of genes.