News / Video /

Watch: Brian Schilder Presents CuEvo and hg-horizon at CSHL Biological Data Science

“Exploring the genome at every scale,” recorded November 6 at Cold Spring Harbor Laboratory

On November 6, Brian Schilder, AI Scientist at Standard Model Biomedicine, presented at the seventh Biological Data Science meeting at Cold Spring Harbor Laboratory, in the session on tools, visualization and infrastructure. His talk asks what happens when computing every pairwise relationship in the genome stops being expensive, and what turns up once you can actually look.

Why every pair

Positions in the genome are not independent. They fold together in 3D, are inherited together, evolve together, and share effects across traits. Each of those relationships is a matrix over pairs of positions, and pairs grow quadratically: the 20 million variants in the 1000 Genomes Project make about 200 trillion of them. In practice, most analyses only look within about a megabase of each variant, which is under 0.1% of all pairs. That limit is set by compute budgets, not by biology.

CuEvo: days to minutes

CuEvo is our GPU toolkit for pairwise genomics: linkage disequilibrium (LD) within a species, and coevolution across species.

  • 70 trillion pairs in 14 minutes. Every variant pair across chromosomes 1 and 2 on a single GPU, against a projected 65 hours for PLINK2 on 64 CPU cores. That is about $0.42 of compute instead of roughly $420.
  • Exact, not approximate. Every output is bit-identical to an exact reference.
  • Built for biobank scale. One million people across one million variants runs end to end in 16 minutes on one GPU.
  • Coevolution across 241 mammals. Roughly 380 trillion pairs genome-wide in about six hours.

A map of the results

Trillions of results only help if you can see them. hg-horizon is a browser that works like a map: everything is precomputed at every scale, so you can zoom from a single gene out to the whole genome, with LD, coevolution, chromatin contacts and GWAS tracks linked together.

What looking at everything showed

  • Most long-range LD is not what it seems. Pool ancestries and the whole genome lights up. Remove ancestry, and most of what remains sits on duplicated sequence.
  • Some of it is real, and it changes results. In a published height locus near the chromosome 11 centromere, fine-mapping both chromosome arms together moved the credible sets to the other arm, millions of bases outside the window standard pipelines use.
  • The same holds across species. The strongest coevolution signals within chromosomes are mostly duplicated sequence, a confound that any AI model trained on alignments will learn unless it is removed.

This work is how we are scaling the genomics side of the Standard Model to millions of patients and billions of variants. To use CuEvo or work with us, get in touch.