---name: nicheformer-spatial-agent description: Foundation model-powered spatial transcriptomics analysis leveraging 53M+ spatially resolved cells for cellular architecture modeling and tissue niche discovery. license: MIT metadata: author: AI Group version: "1.0.0" created: "2026-01-20" compatibility:

• system: Python 3.10+ allowed-tools:
• run_shell_command
• read_file
• write_file

keywords:

• nicheformer-spatial-agent
• automation
• biomedical measurable_outcome: execute task with >95% success rate. ---"

Nicheformer Spatial Agent

The Nicheformer Spatial Agent leverages the Nicheformer foundation model, trained on over 53 million spatially resolved cells, to model cellular architecture and tissue microenvironments with unprecedented accuracy. It enables spatial context-aware cell type annotation, niche discovery, and tissue organization analysis.

When to Use This Skill

• When analyzing spatial transcriptomics requiring deep cellular context understanding.
• For identifying tissue niches and cellular neighborhoods.
• To predict cell-cell interactions based on spatial proximity.
• When transferring annotations from atlases to new spatial data.
• For studying tissue architecture and organization patterns.

Core Capabilities

1. Spatial Context Embeddings: Generate embeddings that capture both gene expression and spatial context.

2. Niche Discovery: Identify recurrent cellular neighborhoods across tissues.

3. Zero-Shot Cell Type Annotation: Transfer cell type labels without retraining.

4. Spatial Perturbation Prediction: Predict effects of removing cell types from niches.

5. Cross-Tissue Transfer: Apply models trained on one tissue to another.

6. Tissue Architecture Analysis: Quantify spatial organization patterns.

Model Architecture

Supported Spatial Technologies

Workflow

1. Input: Spatial transcriptomics data with coordinates.

2. Preprocessing: Normalize, filter, construct spatial graphs.

3. Embedding Generation: Compute Nicheformer embeddings per cell/spot.

4. Niche Clustering: Identify spatial domains and niches.

5. Annotation Transfer: Map cell types from reference atlases.

6. Interaction Analysis: Predict cell-cell communication in niches.

7. Output: Annotated spatial data, niche assignments, interaction networks.

Example Usage

User: "Use Nicheformer to identify cellular niches in this tumor spatial transcriptomics dataset."

Agent Action:

python3 Skills/Genomics/Nicheformer_Spatial_Agent/nicheformer_analysis.py \
    --spatial_data xenium_tumor.h5ad \
    --model_weights nicheformer_pretrained.pt \
    --k_neighbors 15 \
    --niche_resolution 0.5 \
    --reference_atlas tabula_sapiens.h5ad \
    --output tumor_niches_analysis/

Niche Analysis Outputs

Niche Types Detected

AI/ML Components

Foundation Model:

• Transformer backbone with spatial attention
• Pretrained on 53M cells across tissues
• Self-supervised contrastive learning

Spatial Graph Construction:

• Delaunay triangulation
• k-NN with distance threshold
• Hierarchical multi-scale graphs

Transfer Learning:

• Zero-shot annotation via embedding similarity
• Few-shot fine-tuning for novel cell types
• Domain adaptation for new tissues

Performance Benchmarks

Prerequisites

• Python 3.10+
• PyTorch 2.0+, PyTorch Geometric
• Scanpy, Squidpy
• Nicheformer pretrained weights
• GPU with 16GB+ VRAM recommended

Related Skills

• SIMO_Multiomics_Integration_Agent - For multi-omics spatial integration
• scGPT_Agent - For single-cell foundation models
• Cell_Cell_Communication - For ligand-receptor analysis
• Spatial_Epigenomics_Agent - For spatial epigenomics

Spatial Architecture Metrics

Special Considerations

1. Gene Panel Overlap: Ensure sufficient overlap with training data genes
2. Tissue Context: Model performance varies by tissue type
3. Resolution Effects: Aggregate for spot-based technologies
4. GPU Memory: Batch processing for large datasets
5. Validation: Compare with known tissue architecture

Applications

Author

AI Group - Biomedical AI Platform