Full Chromosome Prediction

With AlphaGenome we can generate genome-wide predictions by tiling across entire chromosomes and stitching results into BigWig files. This comes in handy for visualising predicted signal tracks in genome browsers.

Command-Line Script

The script scripts/predict_full_chromosome.py wraps the Python API and writes one BigWig file per chromosome/track.

Quick Start

# Predict ATAC track 0 for chr1 at 128bp resolution (default)
python scripts/predict_full_chromosome.py \
    --model model.pth \
    --fasta hg38.fa \
    --output predictions/ \
    --head atac \
    --tracks 0 \
    --chromosomes chr1

# Full genome at 1bp resolution with center cropping
python scripts/predict_full_chromosome.py \
    --model model.pth \
    --fasta hg38.fa \
    --output predictions/ \
    --head atac \
    --resolution 1 \
    --crop-bp 32768 \
    --batch-size 2

CLI Options

Argument	Default	Description
--model	(required)	Path to model weights (.pth file)
--fasta	(required)	Path to reference genome FASTA file
--output	(required)	Output directory for BigWig files
--head	(required)	Prediction head (atac, dnase, cage, rna_seq, chip_tf, chip_histone, procap)
--tracks	all	Comma-separated track indices to output (e.g. 0,1,2)
--track-names	track_0, …	Comma-separated names for output BigWig files
--resolution	128	Output resolution in bp (1 or 128)
--crop-bp	0	Base pairs to crop from each window edge (e.g. 32768 keeps the center ~50%)
--batch-size	4	Number of windows per inference batch
--window-size	131072	Model input window size in bp
--chromosomes	chr1-22, chrX	Comma-separated list of chromosomes to predict
--organism	0	Organism index (0 = human, 1 = mouse)
--device	cuda	PyTorch device
--dtype-policy	full_float32	Dtype policy for model inference. Use full_float32 for maximum compatibility or mixed_precision for lower GPU memory usage on supported hardware
--quiet	off	Suppress progress bars

Python API

The inference extension lives in alphagenome_pytorch.extensions.inference.

Predicting a Single Chromosome

predict_full_chromosome() returns predictions for one chromosome as a NumPy array:

from alphagenome_pytorch import AlphaGenome
from alphagenome_pytorch.extensions.inference import (
    TilingConfig,
    predict_full_chromosome,
)

model = AlphaGenome.from_pretrained("model.pth", device="cuda")

config = TilingConfig(resolution=1, batch_size=8)

preds = predict_full_chromosome(
    model,
    "hg38.fa",
    chrom="chr1",
    head="atac",
    config=config,
)
# preds.shape == (chrom_length // resolution, n_tracks)

Writing BigWig Files

predict_full_chromosomes_to_bigwig() predicts multiple chromosomes and saves each as a BigWig:

from alphagenome_pytorch.extensions.inference import (
    TilingConfig,
    predict_full_chromosomes_to_bigwig,
)

config = TilingConfig(resolution=128, crop_bp=32768)

results = predict_full_chromosomes_to_bigwig(
    model=model,
    fasta_path="hg38.fa",
    output_dir="./predictions",
    head="atac",
    chromosomes=["chr1", "chr2"],
    config=config,
    track_indices=[0, 1],        # optional: subset of tracks
    track_names=["sample_A", "sample_B"],  # optional: BigWig names
)
# results == {'chr1': [Path('predictions/atac_chr1_sample_A.bw'), ...], ...}

Tiling Configuration

TilingConfig controls how the genome is split into overlapping windows:

config = TilingConfig(
    window_size=131_072,  # model input size (default)
    crop_bp=32_768,       # crop edges to reduce artefacts
    resolution=128,       # 128bp bins (faster) or 1 (base-pair)
    batch_size=4,         # windows per batch
)

TilingConfig fields

Field	Default	Description
window_size	131072	Input window size in bp
crop_bp	0	Base pairs to crop from each edge. Setting this enables overlapping windows so only the center of each window is kept, reducing edge artefacts.
resolution	128	1 for base-pair resolution (requires decoder, slower) or 128 for bin-level resolution (faster)
batch_size	4	Number of windows processed per forward pass

Derived properties:

• effective_size — kept region per window: window_size - 2 * crop_bp

• step_size — equals effective_size for seamless tiling

Tip

Setting crop_bp=32768 (25% of the default 131 072 bp window) keeps the central ~50% of each window. This is a good starting point for reducing edge prediction artefacts.

Supported Heads

Head	Tracks	Resolutions
atac	256	1, 128
dnase	384	1, 128
procap	128	1, 128
cage	640	1, 128
rna_seq	768	1, 128
chip_tf	1664	128 only
chip_histone	1152	128 only

Note

chip_tf and chip_histone only support 128bp resolution. Requesting --resolution 1 with these heads will raise an error.

Performance Tips

• Use resolution 128 when 1bp resolution is not needed.

• Use larger batch size (--batch-size 8) if your GPU memory allows.

• For quick tests, limit chromosomes with --chromosomes chr21,chr22.

• Try loading the model with mixed precision (DtypePolicy.mixed_precision()).

API Reference