geniml

# Geniml: Genomic Interval Machine Learning

## Overview

Geniml is a Python library that bridges genomic interval biology and machine learning. It provides region2vec for learning dense vector representations of genomic regions from BED files, BEDSpace for nearest-neighbor search in embedding space, dataset classes for ML-ready genomic interval loading, and evaluation utilities for embedding quality. Geniml is designed for researchers who want to apply modern ML techniques to chromatin accessibility, histone modification, or other region-based genomic data.

## When to Use

- Learn dense embeddings of genomic regions from a collection of BED files to enable ML-based analysis (region2vec)
- Cluster chromatin accessibility peaks or histone modification sites by embedding similarity
- Search for genomic regions similar to a query region using approximate nearest-neighbor search (BEDSpace)
- Build training datasets for ML models from BED-format genomic intervals with a PyTorch-compatible interface
- Compare embedding quality across training runs or datasets using quantitative metrics
- Integrate genomic region representations into custom neural network architectures
- For basic BED file parsing and set operations without ML, use `gtars` or `pysam-genomic-files` instead

## Prerequisites

- **Python packages**: `geniml`, `torch`, `numpy`, `pandas`, `anndata`
- **Data requirements**: BED files (minimum 3 columns: chr, start, end); optionally a pre-built universe file
- **Environment**: Python 3.8+; GPU optional but recommended for region2vec training on large datasets

```bash
pip install geniml torch numpy pandas anndata
```

## Quick Start

```python
from geniml.region2vec import Region2VecExModel
from geniml.io import RegionSet

# Load a collection of BED files and train region2vec embeddings
region_sets = [RegionSet("sample1.bed"), RegionSet("sample2.bed"), RegionSet("sample3.bed")]
model = Region2VecExModel("path/to/universe.bed")
model.train(region_sets, epochs=10, batch_size=32)

# Get embedding for a specific region
embedding = model.encode("chr1", 1000000, 1500000)
print(f"Embedding shape: {embedding.shape}")
# Embedding shape: (100,)
```

## Core API

### Module 1: RegionSet — Genomic Interval I/O

Load BED files into geniml's primary data structure for downstream operations.

```python
from geniml.io import RegionSet

# Load a BED file
rs = RegionSet("peaks.bed")
print(f"Loaded {len(rs)} regions")
print(f"First region: {rs[0]}")          # Region object: chr, start, end
print(f"Chromosomes: {set(r.chr for r in rs)}")

# Convert to list of Region objects
regions = list(rs)
for r in regions[:3]:
    print(f"  {r.chr}:{r.start}-{r.end}")
```

```python
from geniml.io import RegionSet

# Create RegionSet from a list of (chr, start, end) tuples
regions_data = [
    ("chr1", 100000, 101000),
    ("chr1", 200000, 201500),
    ("chr2",  50000,  51200),
]
rs = RegionSet(regions_data)
print(f"RegionSet with {len(rs)} regions from list")

# Access by index
r = rs[0]
print(f"chr={r.chr}, start={r.start}, end={r.end}, width={r.end - r.start}")
```

### Module 2: Universe Building

A universe defines the set of consensus regions used as the vocabulary for region2vec. Build it from a collection of BED files.

```python
from geniml.universe import UniverseBuilder
from geniml.io import RegionSet

# Collect BED files representing diverse samples
bed_files = ["sample1.bed", "sample2.bed", "sample3.bed", "sample4.bed"]
region_sets = [RegionSet(f) for f in bed_files]

# Build universe (consensus non-overlapping regions)
builder = UniverseBuilder()
universe = builder.build(region_sets)

# Save universe to BED file
universe.to_bed("universe.bed")
print(f"Universe size: {len(universe)} consensus regions")
```

```python
from geniml.universe import UniverseBuilder
from geniml.io import RegionSet

# Build universe with custom parameters
builder = UniverseBuilder(
    fraction=0.5,    # Region must appear in >= 50% of samples to be included
    merge_dist=0,    # Merge adjacent regions within this distance (bp)
)
bed_files = [f"sample_{i}.bed" for i in range(1, 11)]
region_sets = [RegionSet(f) for f in bed_files]
universe = builder.build(region_sets)
print(f"Filtered universe: {len(universe)} regions (fraction >= 0.5)")
```

### Module 3: Region2Vec — Training Embeddings

Train word2vec-style embeddings on genomic regions, treating each BED file as a "document" and each region as a "word."

```python
from geniml.region2vec import Region2VecExModel
from geniml.io import RegionSet

# Initialize model with a pre-built universe
model = Region2VecExModel(universe="universe.bed", embedding_dim=100)

# Load training data (collection of BED files = corpus)
bed_files = [f"atac_{i}.bed" for i in range(1, 51)]
region_sets = [RegionSet(f) for f in bed_files]

# Train
model.train(
    region_sets,
    epochs=20,
    batch_size=64,
    window_size=5,
    min_count=1,
)
print("Training complete")

# Save trained model
model.save("region2vec_model/")
print("Model saved to region2vec_model/")
```

```python
from geniml.region2vec import Region2VecExModel

# Load a pre-trained model
model = Region2VecExModel.load("region2vec_model/")

# Encode a single genomic region
embedding = model.encode("chr1", 1_000_000, 1_500_000)
print(f"Single region embedding shape: {embedding.shape}")
# Single region embedding shape: (100,)

# Encode an entire BED file → matrix of region embeddings
from geniml.io import RegionSet
rs = RegionSet("query_peaks.bed")
embeddings = model.encode_region_set(rs)
print(f"BED file embeddings shape: {embeddings.shape}")
# BED file embeddings shape: (N_regions, 100)
```

### Module 4: BEDSpace — Embedding Nearest-Neighbor Search

Index a corpus of BED file embeddings for fast similarity search.

```python
from geniml.bedspace import BEDSpace
from geniml.region2vec import Region2VecExModel
from geniml.io import RegionSet

# Build a BEDSpace index from a set of BED files
model = Region2VecExModel.load("region2ve