9. Jobs, AlphaGenome, System

Pages: Jobs · AlphaGenome · System · Goal: produce new data (predictions, imports, training runs) and know what this machine can do.

Predict new tracks

Overview → AlphaGenome predictions (or the same button on Jobs) opens the prediction form:

The prediction form. Choose outputs, tissues and cell types from the AlphaGenome catalog (704 terms, 5,563 tracks), samples, windows and haplotypes. It estimates tracks and disk use before you start.

• Outputs: every AlphaGenome output, with how many tracks already exist on disk. Splice junctions feed the Gene products splicing card.
• Tissues / cell types: searchable by name, CURIE, TF or histone mark, and filterable by biosample type. Only terms with tracks for the chosen outputs are listed.
• Samples: all, the Samples-page cohort, or the pinned individuals. Haplotypes include ref, the reference genome baseline.
• Windows: existing windows, any GENCODE gene, custom NAME=chr:start-end regions, curated regulatory elements (pigmentation enhancers near OCA2, KITLG, IRF4; classic loci such as LCT, FTO/IRX3, 9p21), or the genes of a Gene Ontology term or HGNC gene group. New windows are first built for every sample from the dataset's source VCF.

Windows that already have the outputs are skipped unless overwrite is set. The CLI equivalent is genomics alphagenome predict-dataset.

Cost

One AlphaGenome call per sample × haplotype × window: about 4 s on the hosted API and about 0.8 s on a local GPU for a 524 kb window. A whole 1000 Genomes window set (3,202 × 2 × 44 calls) is days of compute, so predict the cohort you need, not the whole dataset. See Requirements.

Follow jobs

Jobs: imports, AlphaGenome predictions, training and evaluation.

Jobs run as separate processes, so they survive closing the tab and stopping the app. A restarted app lists them again. Each job shows progress, an ETA and a live log, and can be cancelled. Retry re-runs a finished job with the same recipe. Each job folder (results/visualizer/jobs/<id>/) keeps task.json with the exact commands, so any job can also be re-run from a shell. Training and evaluation share a gpu queue, and predictions an alphagenome queue. Each queue runs one job at a time.

The top bar shows running jobs from every page. Notify me when a job finishes adds a desktop notification for when the tab is in the background.

Choose the AlphaGenome backend

The AlphaGenome page: where prediction jobs and the Perturbation Lab send their calls.

Mode	Use when	Needs
Hosted API	You have no GPU, or want to start now	ALPHAGENOME_API_KEY in the environment or ~/.env
Remote server	Another machine serves the model	Its address, e.g. grpc://10.0.0.5:50051 (and a CA certificate for self-signed TLS)
This machine	You have an NVIDIA GPU and want privacy and no quotas	genomics alphagenome server setup once. Then Start server here, which shows loading → ready

Test connection checks the service. Test prediction runs a 16 kb prediction, which proves the model runs (on a fresh local server the first call includes JIT compilation). A job keeps the backend it started with. Details: Visualizer reference.

Check the machine

System: the same report as genomics doctor, feature by feature, plus hardware and free disk space wherever the app writes.

When something is greyed out elsewhere in the app, the System page says what is missing and the command that adds it.

Next: bring your own cohort