borzoi

Borzoi — DNA → Functional Track Prediction

Prerequisites

Requirement	Minimum	Recommended
Python	3.10+	3.11
CUDA	12.1+	12.4+
GPU VRAM	16 GB	24 GB+

How to run

from borzoi_pytorch import Borzoi

model = Borzoi.from_pretrained("johahi/borzoi-replicate-0").cuda().eval()
# input: (batch, 4, 524288) one-hot DNA  → output: (batch, tracks, 6144) bins

Borzoi consumes ~524 kb one-hot windows and emits binned predictions across 7,611 human tracks (the separate 2,608-track mouse head is off by default; enable via enable_mouse_head=True and select with forward(..., is_human=False)). For variant scoring, run ref/alt windows centred on the variant and compare per-track output.

Output format

(B, T, L) tensor — T tracks × L 32-bp bins. Track metadata (assay, biosample) is in borzoi_pytorch.pytorch_borzoi_model.TRACKS_DF (or model.tracks_df when using the AnnotatedBorzoi subclass) — the base Borzoi model has no targets attribute.

Remote compute

Needs ≥24 GB VRAM and either pre-cached HF weights or egress to huggingface.co. Read compute_details({provider, mode:'read'}) for an environment with borzoi-pytorch, then:

c = host.compute.create(provider)
job = c.submit_job(
    intent="Borzoi track prediction for 1 locus — 1×GPU, ~2 min",
    inputs=[{"src": "borzoi_run.py", "dst_filename": "borzoi_run.py"}],
    command="python3 borzoi_run.py",   # env selection is host-specific — see compute_details for your provider
    outputs=["tracks.npz"],
    timeout_seconds=1800,
)
print(job.job_id)   # cell ends here — kernel never blocks on compute

Then call the wait_for_notification brain-tool. When the compute_done notification arrives, act on its payload:

save_artifacts(payload["featured_files"])   # paths under hpc/<job_id>/

For the full result dict (output_files, remote_workdir, …), re-enter the kernel: c.attach_job(job_id).result() then c.close(). See the remote-compute-ssh / remote-compute-modal skill for the orchestration details.