borzoi

1# Borzoi — DNA → Functional Track Prediction
2 
3## Prerequisites
4 
5| Requirement | Minimum | Recommended |
6| ----------- | ------- | ----------- |
7| Python      | 3.10+   | 3.11        |
8| CUDA        | 12.1+   | 12.4+       |
9| GPU VRAM    | 16 GB   | 24 GB+      |
10 
11## How to run
12 
13```python
14from borzoi_pytorch import Borzoi
15 
16model = Borzoi.from_pretrained("johahi/borzoi-replicate-0").cuda().eval()
17# input: (batch, 4, 524288) one-hot DNA  → output: (batch, tracks, 6144) bins
18```
19 
20Borzoi consumes ~524 kb one-hot windows and emits binned predictions across
217,611 human tracks (the separate 2,608-track mouse head is off by default;
22enable via `enable_mouse_head=True` and select with
23`forward(..., is_human=False)`). For variant scoring, run ref/alt windows
24centred on the variant and compare per-track output.
25 
26## Output format
27 
28`(B, T, L)` tensor — `T` tracks × `L` 32-bp bins. Track metadata (assay,
29biosample) 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.
30 
31## Remote compute
32 
33Needs ≥24 GB VRAM and either pre-cached HF weights or egress to
34`huggingface.co`. Read `compute_details({provider, mode:'read'})` for an
35environment with `borzoi-pytorch`, then:
36 
37```python
38c = host.compute.create(provider)
39job = c.submitJob(
40    intent="Borzoi track prediction for 1 locus — 1×GPU, ~2 min",
41    inputs=[{"src": "borzoi_run.py", "dstFilename": "borzoi_run.py"}],
42…