SEI (Sequence-to-Effect Inference) is a deep learning framework for predicting chromatin profiles and sequence classes from DNA sequences. This framework allows you to extract embeddings from DNA sequences using the pre-trained SEI model.

The SEI framework provides tools to:

• Extract embeddings from FASTA sequences using the pre-trained SEI model
• Predict chromatin profiles and sequence classes from DNA sequences
• Process sequences in batches for efficient inference

• Python 3.9 or higher
• uv package manager (install from https://github.com/astral-sh/uv)
• CUDA-capable GPU (optional, but recommended for faster inference)

# Or using pip
pip install uv

Run the environment setup cmd:

uv sync
source .venv/bin/activate

The trained SEI model needs to be downloaded from Zenodo. You can use the provided download script:

bash download_data.sh

This will download:

• The trained SEI model (sei_model.tar.gz) from Zenodo
• SEI framework resources (FASTA files) from Zenodo

Note: The model files should be extracted to the model/ directory. The main model file should be at model/sei.pth.

Alternatively, you can manually download from:

• Model: https://zenodo.org/record/4906996 (DOI: 10.5281/zenodo.4906996)
• Extract the model files to the model/ directory

Or just run

bash env_setup.sh

The main script for extracting embeddings is retrieve_embeddings/retrieve_embeddings.py.

python retrieve_embeddings/retrieve_embeddings.py \
    --input-file retrieve_embeddings/test.fasta \
    --output-file output/embeddings.npz

python retrieve_embeddings/retrieve_embeddings.py \
    --input-file <path-to-input.fasta> \
    --output-file <path-to-output.npz> \
    --model-path model/sei.pth \
    --batch-size 32 \
    --sequence-length 4096 \
    --use-hooks

• --input-file (required): Path to input FASTA file containing DNA sequences
• --output-file (required): Path to output .npz file where embeddings will be saved
• --model-path (optional): Path to SEI model file (default: model/sei.pth)
• --batch-size (optional): Batch size for processing sequences (default: 32)
• --sequence-length (optional): Target sequence length for encoding (default: 4096)
• --use-hooks (default): Use register_hooks method for embedding extraction (recommended)
• --no-use-hooks: Use manual layer-by-layer method instead of hooks

The script outputs a compressed NumPy archive (.npz) file containing:

• ids: Array of sequence IDs from the FASTA file
• embeddings: Array of embeddings with shape (num_sequences, 960, 16)

# Extract embeddings using the default method (hooks)
python retrieve_embeddings/retrieve_embeddings.py \
    --input-file retrieve_embeddings/test.fasta \
    --output-file output/embeddings.npz

# Extract embeddings using manual method
python retrieve_embeddings/retrieve_embeddings.py \
    --input-file retrieve_embeddings/test.fasta \
    --output-file output/embeddings.npz \
    --no-use-hooks

You can load the saved embeddings in Python:

import numpy as np

# Load embeddings
data = np.load('output/embeddings.npz')
sequence_ids = data['ids']
embeddings = data['embeddings']

print(f"Loaded {len(sequence_ids)} sequences")
print(f"Embeddings shape: {embeddings.shape}")  # (num_sequences, 960, 16)

sei-framework/
├── model/                # SEI model files
│   ├── sei.py            # Model architecture
│   ├── sei.pth           # Trained model weights (download required)
│   └── *.names           # Target and sequence class names
├── retrieve_embeddings/  # Embedding extraction scripts
│   ├── retrieve_embeddings.py  # Main embedding extraction script
│   ├── util.py           # Utility functions for embeddings
│   └── test.fasta        # Example input file
├── encode/               # Sequence encoding utilities
├── pca/                  # PCA analysis tools
├── tests/                # Unit tests
├── env_setup.sh          # Environment setup script
├── download_data.sh      # Model download script
└── pyproject.toml        # Project dependencies

Run the test suite to verify your installation:

pytest tests/

If you encounter FileNotFoundError: Model file not found: model/sei.pth, make sure you have:

1. Downloaded the model using bash download_data.sh
2. Extracted the model files to the model/ directory
3. Verified that model/sei.pth exists

The framework will automatically use CPU if CUDA is not available.