Skip to content

Genomics Workspace

What does a person's genome change about how their genes work? This workspace answers that question one individual, one haplotype and one gene at a time. Each person's two phased copies of a region go through AlphaGenome, which predicts expression, splicing, accessibility and binding. Every prediction is then put next to evidence it can be checked against.

Everything is reachable from one local web app, genomics visualize:

The Tracks page: three individuals' predicted melanocyte RNA-seq over TYRP1, the reference-genome prediction (dashed) and the observed ENCODE signal under each track, with an agreement badge
The Tracks page on the 1000 Genomes dataset. It shows three individuals' predicted melanocyte RNA-seq over TYRP1 and AlphaGenome's prediction for the reference genome (dashed). Under each track is the ENCODE experiment the track was trained on, with how well the two agree.

Where to start

  • Try it in ten minutes

    Install, check the machine with genomics doctor, open the app and import a public cohort.

    Your first session

  • Learn the app, page by page

    A guided tour along one question: what does HG00096's genome do to their pigmentation genes, and can the model be trusted on it?

    Visualizer tutorial

  • Understand the approach

    Why the workspace predicts each haplotype on its own sequence, and how it uses baselines, observed data and negative controls before trusting a result.

    From genome to function

  • Script it

    Dataset builders, AlphaGenome prediction, model training and the other pipelines, from the command line or a notebook.

    Pipelines · CLI reference

Quick start

scripts/env/install.sh       # conda env "genomics" with everything the visualizer can use
conda activate genomics
genomics doctor              # what this machine can run, and how to enable the rest
genomics visualize --open

What is in the box

Area What it does Where
Visualizer Cohorts, AlphaGenome tracks against the reference and observed data, haplotype sequences, variant effects, transcripts and proteins per haplotype, model training and in-silico perturbation, all in one app Tutorial · Reference
AlphaGenome Predictions through the hosted API or a model server on your own NVIDIA GPU AlphaGenome
Dataset builders Canonical per-sample, per-window datasets from 1000 Genomes or any phased VCF Dataset builders
Predictors CNNs over predicted tracks, a variant transformer, and SNP ancestry Genotype predictor
Genome processing FASTQ/BAM/CRAM → VCF (genomes-analyzer), VCF → 23andMe Genomes analyzer

All of it runs through one command, genomics (genomics --help), and one package, src/genomics/.