genebeddings

Unified interface for extracting embeddings from genomic foundation models.

Overview

genebeddings provides:

• Standardized wrappers for 16 genomic foundation models (transformers, CNNs, state-space models, track predictors)
• Geometric analysis tools for single-variant and epistasis embeddings
• Embedding storage via SQLite key-value store
• Benchmarking utilities for pathogenicity prediction

Installation

pip install -e .

Install with model-specific dependencies:

pip install -e ".[nt]"          # Nucleotide Transformer
pip install -e ".[borzoi]"      # Borzoi
pip install -e ".[alphagenome]" # AlphaGenome (requires JAX + GPU)
pip install -e ".[all]"         # Common models

Quick Start

Embeddings

from genebeddings.wrappers import NTWrapper

model = NTWrapper()
embedding = model.embed("ACGTACGT" * 100, pool="mean")  # (hidden_dim,) numpy array

Nucleotide Predictions

probs = model.predict_nucleotides("ACGTNACGT", positions=[4])
# [{'A': 0.1, 'C': 0.2, 'G': 0.3, 'T': 0.4}]

Track Predictions

from genebeddings.wrappers import BorzoiWrapper

borzoi = BorzoiWrapper()
tracks = borzoi.predict_tracks("ACGT" * 131_072)  # (num_tracks, length) numpy array

Variant Geometry

from genebeddings import SingleVariantGeometry

geom = SingleVariantGeometry(wt_embedding, mut_embedding)
print(geom.cosine_distance, geom.euclidean_distance)

Supported Models

See wrappers/summary.md for full details.

Testing

python quick_test.py