Nicheformer

Nicheformer

Nicheformer is a transformer-based model designed for understanding and predicting cellular niches and their interactions. The model uses masked language modeling to learn representations of cellular contexts and their relationships.

Model Description

Nicheformer is built on a transformer architecture with the following key features:

• Architecture: Transformer encoder with customizable number of layers and attention heads
• Pre-training: Masked Language Modeling (MLM) objective with dynamic masking
• Input Processing: Handles cell type, assay, and modality information
• Positional Encoding: Supports both learnable and fixed positional embeddings
• Masking Strategy:
  • 80% of selected tokens are replaced with [MASK]
  • 10% are replaced with random tokens
  • 10% remain unchanged

Model Architecture

• Transformer encoder layers: 12
• Hidden dimension: 512
• Attention heads: 16
• Feedforward dimension: 1024
• Maximum sequence length: 1500
• Vocabulary size: 25000
• Masking probability: 15%

Usage

from transformers import AutoModelForMaskedLM, AutoTokenizer
import anndata as ad

# Load model and tokenizer
model = AutoModelForMaskedLM.from_pretrained("aletlvl/Nicheformer", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("aletlvl/Nicheformer", trust_remote_code=True)

# Set technology mean for HF tokenizer
technology_mean_path = 'technology_mean.npy'
technology_mean = np.load(technology_mean_path)
tokenizer._load_technology_mean(technology_mean)

# Load your single-cell data
adata = ad.read_h5ad("your_data.h5ad")

# Tokenize the data
inputs = tokenizer(adata)

# Get embeddings
embeddings = model.get_embeddings(
              input_ids=inputs["input_ids"],
              attention_mask=inputs["attention_mask"],
              layer=-1,
              with_context=False
            )

Training Data

The model was trained on single-cell gene expression data from various tissues and organisms. It supports:

• Modalities: spatial and dissociated
• Species: human and mouse
• Technologies:
  • MERFISH
  • CosMx
  • Xenium
  • 10x Genomics (various versions)
  • CITE-seq
  • Smart-seq v4

Limitations

• The model is specifically designed for gene expression data and may not generalize to other types of biological data
• Performance may vary depending on the quality and type of input data
• The model works best with data from supported species and technologies

License

This model is released under the MIT License. See the LICENSE file for more details.

Contact

For questions and issues, please open an issue on the GitHub repository or contact the maintainers.

nicheformer

This is the official repository for Nicheformer: a foundation model for single-cell and spatial omics

 

Citation

If you use our tool or build upon our concepts in your own work, please cite it as

Schaar, A.C., Tejada-Lapuerta, A., et al. Nicheformer: a foundation model for single-cell and spatial omics. bioRxiv (2024). doi: https://doi.org/10.1101/2024.04.15.589472

Contact

For questions and help requests, you can reach out on GitHub or email to the corresponding author (alejandro.tejadalapuerta@helmholtz-munich.de).