geniml

Geniml: Genomic Interval Machine Learning

Overview

Geniml is a Python package for building machine learning models on genomic interval data from BED files. It provides unsupervised methods for learning embeddings of genomic regions, single cells, and metadata labels, enabling similarity searches, clustering, and downstream ML tasks.

Installation

Install geniml using uv:

uv pip install geniml

For ML dependencies (PyTorch, etc.):

uv pip install 'geniml[ml]'

Development version from GitHub:

uv pip install git+https://github.com/databio/geniml.git

Core Capabilities

Geniml provides five primary capabilities, each detailed in dedicated reference files:

1. Region2Vec: Genomic Region Embeddings

Train unsupervised embeddings of genomic regions using word2vec-style learning.

Use for: Dimensionality reduction of BED files, region similarity analysis, feature vectors for downstream ML.

Workflow:

Reference: See references/region2vec.md for detailed workflow, parameters, and examples.

2. BEDspace: Joint Region and Metadata Embeddings

Train shared embeddings for region sets and metadata labels using StarSpace.

Use for: Metadata-aware searches, cross-modal queries (region→label or label→region), joint analysis of genomic content and experimental conditions.

Workflow:

Reference: See references/bedspace.md for detailed workflow, search types, and examples.

3. scEmbed: Single-Cell Chromatin Accessibility Embeddings

Train Region2Vec models on single-cell ATAC-seq data for cell-level embeddings.

Use for: scATAC-seq clustering, cell-type annotation, dimensionality reduction of single cells, integration with scanpy workflows.

Workflow:

Reference: See references/scembed.md for detailed workflow, parameters, and examples.

4. Consensus Peaks: Universe Building

Build reference peak sets (universes) from BED file collections using multiple statistical methods.

Use for: Creating tokenization references, standardizing regions across datasets, defining consensus features with statistical rigor.

Workflow:

Methods:

Reference: See references/consensus_peaks.md for method comparison, parameters, and examples.

5. Utilities: Supporting Tools

Additional tools for caching, randomization, evaluation, and search.

Available utilities:

Reference: See references/utilities.md for detailed usage of each utility.

Common Workflows

Basic Region Embedding Pipeline

from geniml.tokenization import hard_tokenization
from geniml.region2vec import region2vec
from geniml.evaluation import evaluate_embeddings

# Step 1: Tokenize BED files
hard_tokenization(
    src_folder='bed_files/',
    dst_folder='tokens/',
    universe_file='universe.bed',
    p_value_threshold=1e-9
)

# Step 2: Train Region2Vec
region2vec(
    token_folder='tokens/',
    save_dir='model/',
    num_shufflings=1000,
    embedding_dim=100
)

# Step 3: Evaluate
metrics = evaluate_embeddings(
    embeddings_file='model/embeddings.npy',
    labels_file='metadata.csv'
)

scATAC-seq Analysis Pipeline

import scanpy as sc
from geniml.scembed import ScEmbed
from geniml.io import tokenize_cells

# Step 1: Load data
adata = sc.read_h5ad('scatac_data.h5ad')

# Step 2: Tokenize cells
tokenize_cells(
    adata='scatac_data.h5ad',
    universe_file='universe.bed',
    output='tokens.parquet'
)

# Step 3: Train scEmbed
model = ScEmbed(embedding_dim=100)
model.train(dataset='tokens.parquet', epochs=100)

# Step 4: Generate embeddings
embeddings = model.encode(adata)
adata.obsm['scembed_X'] = embeddings

# Step 5: Cluster with scanpy
sc.pp.neighbors(adata, use_rep='scembed_X')
sc.tl.leiden(adata)
sc.tl.umap(adata)

Universe Building and Evaluation

# Generate coverage
cat bed_files/*.bed > combined.bed
uniwig -m 25 combined.bed chrom.sizes coverage/

# Build universe with coverage cutoff
geniml universe build cc \
  --coverage-folder coverage/ \
  --output-file universe.bed \
  --cutoff 5 \
  --merge 100 \
  --filter-size 50

# Evaluate universe quality
geniml universe evaluate \
  --universe universe.bed \
  --coverage-folder coverage/ \
  --bed-folder bed_files/

CLI Reference

Geniml provides command-line interfaces for major operations:

# Region2Vec training
geniml region2vec --token-folder tokens/ --save-dir model/ --num-shuffle 1000

# BEDspace preprocessing
geniml bedspace preprocess --input regions/ --metadata labels.csv --universe universe.bed

# BEDspace training
geniml bedspace train --input preprocessed.txt --output model/ --dim 100

# BEDspace search
geniml bedspace search -t r2l -d distances.pkl -q query.bed -n 10

# Universe building
geniml universe build cc --coverage-folder coverage/ --output universe.bed --cutoff 5

# BEDshift randomization
geniml bedshift --input peaks.bed --genome hg38 --preserve-chrom --iterations 100

When to Use Which Tool

Use Region2Vec when:

Use BEDspace when:

Use scEmbed when:

Use Universe Building when:

Use Utilities when:

Best Practices

General Guidelines

Performance Considerations

Integration Patterns

Related Projects

Geniml is part of the BEDbase ecosystem:

Additional Resources

Troubleshooting

"Tokenization coverage too low":

"Training not converging":

"Out of memory errors":

"StarSpace not found" (BEDspace):

For detailed troubleshooting and method-specific issues, consult the appropriate reference file.