sequencing my genome and using agents to help me explore the results taught me a ton about genomics in general and obviously a lot about my own genetics, so I love that this is becoming more popular (h/t @patrickc and @sethshowes) exploring my genome has also helped me understand the important limits of our current understanding of the human genome and genetic variation genomic variants 101 DNA is >99% similar across humans so if you get your whole genome sequenced (WGS) the interesting results are where you differ from a "reference genome." the locations where you differ are "variants" these variants can be single nucleotide swaps (eg the reference genome has a T at a certain location but you have a G), missing or extra letters (aka insertion/deletion or "indels") or entire long chunks that are different there are large databases like clinvar (ncbi.nlm.nih.gov/clinvar/) where researchers can submit findings about variants to catalog what we know about them, sort of like wikipedia for variants example: suppose your genome sequencing results say you have a variant in chromosome 17 at position 43051118. you have an A there and in the reference genome there's a C. clinvar tells you what researchers have found about this specific variant in the BRCA1 gene -- unfortunately it means you are at increased risk of hereditary breast and ovarian cancer in practice however most people's genomes are full of rare variants that are poorly studied. that BRCA1 variant is pretty famous but most clinically relevant variants are either absent from clinvar or are labeled as "variants of unknown significance" (VUS) when I cross referenced my WGS results with clinvar I had 310 variants of unknown significance. aka 310 instances where researchers have looked into a variant and yet we still can't say whether it's harmful or harmless or irrelevant using DNA foundation models to understand variants there are several ways to determine whether a variant is harmful -- you can run large-scale population studies and look for correlations between variants and diseases, you can try to find families where some people have the variant and others don't as a pseudo controlled trial, you can run actual wet lab experiments to see if introducing a variant into a gene disrupts its ability to produce viable proteins, etc. the newest generation of foundation models trained on huge quantities of DNA offers a new computational approach evo2 is a model trained on trillions of nucleotides across organisms of all kinds including bacteria & plants & animals. of all the genomes it was trained on, only one was a human genome. the training wasn't meant to teach it about human genetics but instead to develop a broader understanding of DNA as the language of all biology, to learn the general grammar of what "good" DNA looks like based on the broad arc of evolutionary history with models like evo, we can feed in DNA sequences that have specific variants and see how the model "reacts" to each one. sort of like taking a sentence and making a change to one word, then using a language model to see whether that change was a typo or just a different but valid phrase. or swapping a line of code and using the knowledge inside a coding model to gauge whether that introduced a bug there are simple ways to "gauge the model's reaction" (eg using the model's loss) that work pretty well, but last week we @GoodfireAI released a novel method for using evo's internal activations to achieve state of the art performance on variant effect prediction, in collaboration with @MayoClinic. this was a super exciting breakthrough for the team and we hope it's a big moment for this field as a whole we then did the natural thing -- take all 4.2 million variants on clinvar, get a whole bunch of gpus running with parallel copies of evo, and use our new approach to predict the pathogenicity of every single one of 'em. we’ve released the results to the public (link below!) in addition to these raw predictions, the team made a major breakthrough in actually understanding and communicating how/why evo considers a variant to be harmful or not. using additional classifiers to understand key variables inside the model's activations, we can generate a specific human-readable hypothesis for each variant. these explanations are also all made available on our public interactive explorer: evee.goodfire.ai/ - - - i mentioned that my genome has 310 variants of unknown significance. i built an MCP to so agents could query our EVEE tool to understand each of them, and it's public in case others want to do the same as part of their personal genome analysis: github.com/goodfire-ai/evee-… fortunately most of my variants that evo predicts to be pathogenic with high probability are in genes linked to recessive diseases and I’m heterozygous. (this means only one of my two copies of the gene has the variant, and even having a single functional copy of the gene is enough for me to be healthy.) it's nevertheless useful for me to know about these variants because it means i'm a carrier for a few rare diseases like reticular dysgenesis stepping back, it's a truly insane tech stack that makes all of this possible. modern sequencing tech means I can get my genome sequenced for a few hundred bucks; frontier LLM agents let me process and analyze my genome using all sorts of tools and databases despite having no formal bioinformatics training; frontier scientific models like evo can understand modalities like DNA as fluently as we understand english giving them an ability to intuit which variants are harmful; interpretability techniques allow us to extract knowledge from these superhuman systems. right now personal genome analysis using AI might sound like an SF biohacker thing but every technology I just mentioned is getting cheaper and more broadly accessible -- the future is already here, it’s just not evenly distributed