A hierarchical transformer-based framework that orchestrates multiple genomic language models to detect regulatory elements with both genome-wide context and nucleotide-level precision
Regulatory elements (enhancers and promoters) pose unique detection challenges:
- Distance: Can be up to 1 million base pairs away from target genes
- Size variability: Enhancers range from 100-2000bp with no fixed length
- Fuzzy boundaries: Unlike genes, regulatory elements have gradual transitions
- Context-dependent: Same sequence can be active in one cell type but not another
- Evolutionary flexibility: Regulatory sequences evolve rapidly
Traditional methods miss complex patterns. Single deep learning models face a fundamental tradeoff between long-range context and nucleotide precision.
Key Insight: Instead of building one model that does everything adequately, BRIDGE orchestrates multiple specialized models, each excelling at what it does best, in a biologically-informed pipeline.
Input DNA (up to 450kb)
↓
┌─────────────────────────────────────┐
│ Stage 1: Long-Range Scanning │
│ HyenaDNA (450kb context) │
│ → Coarse predictions │
│ → Candidate regions │
└─────────────────────────────────────┘
↓
┌─────────────────────────────────────┐
│ Stage 2: Precise Boundary Detection │
│ DNABERT-2 (512bp windows) │
│ → Token-level analysis │
│ → Attention gradients │
│ → Exact boundaries │
└─────────────────────────────────────┘
↓
┌─────────────────────────────────────┐
│ Stage 3: Multi-Signal Integration │
│ • LucaOne validation │
│ • DNA shape features │
│ • Conservation scores │
│ • Motif analysis │
└─────────────────────────────────────┘
↓
Validated Regulatory Elements
- 🎯 Multi-scale analysis: Combines long-range (450kb) and short-range (512bp) models
- 🌍 Cross-species support: Works across human, mouse, zebrafish, and other species using mean pooling
- ⚡ Production-ready: vLLM optimization, REST API, batch processing, and continuous serving
- 🧬 Multi-modal validation: DNA, RNA, and protein sequence integration
- 📊 Expression prediction: Links regulatory elements to gene expression
- 🔧 Modular design: Easy to extend with new models or analysis types
- 🎨 Precise boundaries: Multi-signal integration for accurate element delineation
# Clone repository
git clone https://github.com/chazhyseni/BRIDGE-LM.git
cd BRIDGE-LM
# Create conda environment
conda env create -f environment.yml
conda activate dna-models
# Install BRIDGE
pip install -e .
# Download pre-trained models (50GB)
python scripts/download_models.py
python scripts/download_data.py --genomes hg38 mm10from bridge import WorkflowOrchestrator
import asyncio
# Initialize BRIDGE
orchestrator = WorkflowOrchestrator()
await orchestrator.initialize()
# Analyze a sequence
from bridge.core.models import DNASequence
seq = DNASequence(
sequence="ATCGATCG" * 1000, # 8kb sequence
chromosome="chr12",
start=7940000,
end=7948000
)
# Get results
results = await orchestrator.components['model_manager'].predict_regulatory_elements(seq)
# Print detected elements
for element in results:
print(f"{element.element_type}: {element.chr}:{element.start}-{element.end}")
print(f" Score: {element.score:.3f}")
print(f" Core region: {element.core_start}-{element.core_end}")
print(f" TF motifs: {', '.join([m['tf'] for m in element.tf_motifs[:3]])}")# Analyze single sequence
bridge analyze -s ATCGATCG... --output results.json
# Batch analysis
bridge batch -i sequences.fasta -o results.bed --format bed
# Gene-centric analysis
bridge gene -g NANOG --upstream 1000000 --downstream 100000
# Start API server
bridge serve --host 0.0.0.0 --port 8000# Compare regulatory landscapes across species
results = await orchestrator.cross_species_analysis(
reference_species="human",
target_species=["mouse", "zebrafish"],
chromosome="chr2",
start=176000000,
end=176100000
)
# Find ultra-conserved elements
for element in results['ultra_conserved']:
print(f"Conserved {element['type']}: {element['conservation']:.1%}")# Validate DNA→RNA→Protein relationships
validation = await orchestrator.multimodal_analysis(
dna_sequence="ATCG...",
rna_sequence="AUGC...",
protein_sequence="MKVL..."
)
print(f"Central dogma valid: {validation['central_dogma']['all_valid']}")
print(f"Protein-DNA binding score: {validation['protein_dna_interaction']['score']}")DNABERT-2's attention patterns show sharp transitions at regulatory boundaries. We compute attention gradients and use peaks to identify precise start/end positions.
Different signals are weighted based on element type:
- Enhancers: Higher weight on motif density and nucleosome depletion
- Promoters: Higher weight on attention gradients and conservation
HyenaDNA's mean pooling creates species-agnostic representations, enabling analysis of any species without retraining.
- GPU: NVIDIA GPU with 16GB+ VRAM (V100/A100 recommended)
- RAM: 32GB minimum
- Storage: 50GB for models and reference data
- OS: Linux (Ubuntu 20.04+) or macOS
- Python: 3.9-3.10
- Throughput: 1000+ sequences/second (with GPU)
- Memory efficiency: Handles 450kb sequences
- Accuracy: Validated on VISTA enhancers and ENCODE cCREs
- Scalability: Distributed processing support
- Fine-tuning: LoRA-based adaptation for cell types
- Custom pipelines: YAML-based workflow definition
- Variant impact: Predict effects of mutations
- CRISPR design: Target regulatory cores
- Expression prediction: Link elements to genes
We welcome contributions! See CONTRIBUTING.md for guidelines.
BRIDGE-LM builds upon excellent work by:
- Email: chaz.hyseni@gmail.com
- Issues: GitHub Issues
- Discussions: GitHub Discussions
BRIDGE-LM - Advancing genomic understanding through hierarchical model orchestration