Visualizer Tutorial¶

genomics visualize is one local web app over a cohort's genomes and their AlphaGenome predictions. This tutorial goes through its pages in the order a question naturally takes you through them. Each step is short, shows what you should see, and says how to read it.

genomics visualize --open      # http://localhost:8780/

The running example¶

All screenshots come from the 1000 Genomes high-coverage dataset (1kg_high_coverage): 3,202 genomes from 26 populations, 44 gene windows of 524 kb, and AlphaGenome RNA-seq for three skin cell types. The 11 windows listed in the dataset's metadata are pigmentation genes (OCA2, HERC2, TYR, TYRP1, MC1R, SLC24A5, SLC45A2, …). The other 33 are control genes.

We follow one individual, HG00096 (British, GBR), and one question:

What does HG00096's genome change about their pigmentation genes, and how far can we trust the prediction?

flowchart TD
    A["1 · Overview<br/>open the dataset"] --> B["2 · Samples<br/>cohort, pins, ancestry"]
    B --> C["3 · Tracks<br/>predictions vs reference and observed data"]
    C --> D["4 · Sequence<br/>which variants?"]
    D --> E["5 · Variant<br/>one site across 3,202 genomes"]
    E --> F["6 · Gene products<br/>transcripts and proteins"]
    C --> G["7 · Experiments<br/>train with negative controls"]
    G --> H["8 · Perturbation Lab<br/>why did the model decide?"]

Step	Page	You will learn to
1. Open a dataset	Overview	read a dataset at a glance, open gene cards, restore saved views
2. Build a cohort	Samples	filter and pin samples, check ancestry with a PCA, define per-sample scalars
3. Read predicted tracks	Tracks	compare individuals, groups and the whole cohort with the reference genome and observed data, and export figures
4. Compare haplotypes	Sequence	see each haplotype's bases and variants against the reference
5. Test a variant	Variant	check a variant's predicted effect against GTEx
6. From DNA to protein	Gene products	follow one individual's two copies of a gene into mRNA, protein and structure
7. Train and evaluate models	Experiments	train from a form, read runs, and add negative controls
8. Ask the model why	Perturbation Lab	edit a genome in silico and see what the model reads
9. Jobs, AlphaGenome, System	Jobs · AlphaGenome · System	run predictions, choose a backend, check the machine
10. Bring your own cohort	Import dataset	build a dataset from any phased VCF
11. From a notebook	–	get the same arrays in Python

You can read the steps in any order. Steps 1–6 need only a dataset. Steps 7–8 need the [genotype] extra (PyTorch), and step 8 also needs an AlphaGenome backend.

The app at a glance¶

The app shell with numbered callouts — The app shell. The numbers match the list below.

1 Dataset switcher. Every open dataset (registered, opened by path, or imported). Switching keeps a separate cohort, pins and locus per dataset.

2 Pages. The upper group explores data: Overview, Samples, Tracks, Sequence, Variant, Gene products, Perturbation Lab. The lower group runs work: Experiments, Jobs, AlphaGenome, System.

3 Cohort chip. How many samples pass the Samples-page filters, and how many individuals are pinned. Click it to edit the cohort. The cohort and the pins are shared by every page.

4 Theme. Light or dark. Figure export can always use a light background.

5 Page actions. Each page's main actions sit under its title.

Habits that help¶

Links are shareable. The URL holds the page and its main parameters, e.g. #/products?gene=MC1R&sample=HG00096&tissue=CL:1000458.
Names are clickable. A gene name opens a gene card (HGNC, Ensembl, UniProt, GTEx, ClinVar, Gene Ontology…). A track name opens a track card (ontology term, assay, the ENCODE / FANTOM5 experiments behind it). Ontology CURIEs and sample IDs link to OLS and IGSR.
Long work does not block you. Cohort-wide computations show a progress card and keep running if you leave the page. Their results are cached on disk, so the second time is instant.
Plots respond to the keyboard and mouse. Drag to pan, Ctrl/⌘+scroll to zoom, Shift+drag to zoom to a region, double-click to zoom in, and use ← → + − on the focused plot.
The locus box understands genes and rsIDs. Type chr15:28120402-28120541, TYRP1 or rs12913832.

Every control is described in full in the Visualizer reference.

Start: open a dataset