Decision Guide: Choosing Your Niche Tools¶
Spatial niche analysis tools fall into two stages: recognizing niches (what are the spatial neighborhoods?) and characterizing their function (what biology drives each niche?). This guide helps you pick the right tool for each stage.
Core insight
The niche definition you choose determines what you can find. A composition-based tool finds cell-type mixtures; a communication-based tool finds signaling hubs. Neither is wrong — they answer different questions.
Table 1: Niche Recognition¶
Goal: Identify and label spatial neighborhoods in your tissue.
| Tool | Approach | How It Defines a Niche | Platforms Tested | Key Strength | Key Limitation | Reference |
|---|---|---|---|---|---|---|
| CellCharter | Composition | Cell-type proportions in spatial neighborhoods, learned via GNN at multiple resolutions | CODEX, CosMx, MERSCOPE, Visium, IMC | Works on both RNA and protein platforms; multi-scale | Requires cell-type annotations as input | Varrone et al., Nature Genetics 2024 |
| CytoCommunity | Composition | Supervised graph partitioning using condition labels to find disease-relevant neighborhoods | CODEX, IMC | Finds condition-specific niches (e.g., responder vs. non-responder) | Requires condition labels; supervised | Hu et al., Nature Methods 2024 |
| SpatialLDA | Composition | Topic model: spatial regions are "documents," cell types are "words," niches are latent topics | CODEX, MERFISH | Probabilistic; soft niche assignments | Sensitive to hyperparameters (number of topics) | Chen et al., Genome Biology 2022 |
| BANKSY | Expression | Gene expression augmented with mean and azimuthal gradient of neighbor expression | Visium, MERFISH, STARmap, Slide-seq | Fast; captures expression boundaries; top DLPFC performer | Defines domains, not niches in most configurations | Singhal et al., Nature Genetics 2024 |
| SpiceMix | Expression | NMF with spatial priors: metagenes with spatially varying loadings | seqFISH, MERFISH, STARmap | Learns spatial gene programs directly | Computationally intensive; limited platform testing | Chidester et al., Nature Genetics 2023 |
| ENVI/COVET | Covariance | Gene-gene covariance tensor across spatial neighbors; ENVI imputes via conditional VAE | Slide-seq, MERFISH, seqFISH | Captures niche effects beyond cell-type composition; imputes whole transcriptome | Conceptually complex; requires matched scRNA-seq for ENVI | Haviv et al., Nature Biotechnology 2024 |
| NicheCompass | Communication | GNN with ligand-receptor prior on spatial graph; niches defined by signaling patterns | Xenium, CosMx | Niches are mechanistically interpretable (L-R programs); cross-sample atlas | RNA-only (no protein); small panels lose L-R programs | Birk et al., Nature Genetics 2025 |
| Nicheformer | Foundation model | 49M-param transformer pre-trained on dissociated + spatial data | Visium, MERFISH, Slide-seq | Transfer learning; flexible downstream tasks | Black box; needs fine-tuning per tissue | Fischer et al., Nature Methods 2025 |
| Novae | Foundation model | Graph-based foundation model with zero-shot niche inference across technologies | Visium, MERFISH, Xenium, Slide-seq | Zero-shot transfer; no retraining needed | Very new (2025); limited independent validation | Music et al., Nature Methods 2025 |
Domain detection ≠ Niche identification
The tools below find spatially contiguous regions (domains), not distributed cellular neighborhoods (niches). They are frequently confused with niche tools but answer a different question. See Niche vs Domain.
| Tool | What It Finds | Why It's Not a Niche Tool | When to Use It Instead | Reference |
|---|---|---|---|---|
| GraphST | Contiguous spatial domains via self-supervised contrastive learning | Enforces spatial smoothness — scattered niches are merged into surrounding tissue | Tissue region segmentation (e.g., cortical layers, tumor vs. stroma) | Long et al., Nature Communications 2023 |
| STAGATE | Spatial domains via graph attention autoencoder | Adaptive graph construction biases toward contiguous regions | When spatial domains align with anatomical structures | Dong & Zhang, Nature Communications 2022 |
| BayesSpace | Spatial clusters via Bayesian model with neighbor priors | Explicit spatial smoothness prior penalizes non-contiguous clusters | Low-resolution data (Visium) where domains are large | Zhao et al., Nature Biotechnology 2021 |
Table 2: Functional Characterization¶
Goal: Understand the biology driving each niche — signaling, gene programs, morphology.
Cell-Cell Communication¶
| Tool | Approach | Spatial-Aware? | Platforms Tested | Key Strength | Key Limitation | Reference |
|---|---|---|---|---|---|---|
| NicheCompass | GNN with L-R prior on spatial graph | Yes — interactions constrained to spatial neighbors | Xenium, CosMx | Joint niche + communication model; mechanistically interpretable | RNA-only; small panels lose L-R programs | Birk et al., Nature Genetics 2025 |
| COMMOT | Optimal transport with spatial distance constraint | Yes — rejects implausible long-range signals | Visium, MERFISH, seqFISH, STARmap, Slide-seq | Principled spatial filtering; scalable | Xenium/CosMx untested; CCC only (no niche output) | Cang et al., Nature Methods 2023 |
| CellChat v2 | Statistical L-R inference + spatial mode | Partial — spatial mode is an add-on to non-spatial core | Visium, seqFISH, MERFISH, Slide-seq, STARmap | Largest L-R database; most widely used; rich visualization | Originally non-spatial; spatial mode less validated | Jin et al., Nature Communications 2024 |
| SpaTalk | Graph network modeling of spatially resolved L-R interactions | Yes — models interactions at single-cell resolution | SpaTalk-tested on ST, Slide-seq, MERFISH | Single-cell resolution L-R; cell-type deconvolution built in | Smaller L-R database than CellChat | Shao et al., Nature Communications 2022 |
| SpatialDM | Bivariate Moran's I for spatially co-expressed L-R pairs | Yes — tests spatial co-expression statistically | Visium, Slide-seq, seqFISH | Simple, statistically rigorous | Tests co-expression, not interaction; spot-level | Li et al., Nature Communications 2023 |
Niche Gene Programs¶
| Tool | What It Tests | Spatial-Aware? | Platforms Tested | Key Strength | Key Limitation | Reference |
|---|---|---|---|---|---|---|
| Niche-DE | Whether gene expression depends on niche context, not cell type alone | Yes — conditions on niche assignment | Visium, Slide-seq, CosMx, Xenium | Separates niche effect from cell-intrinsic expression; niche-LR extension for mechanistic insight | niche-LR fails on small gene panels (~300 genes) | Bai et al., Genome Biology 2024 |
| NCEM | How neighborhood composition influences gene expression via GNN | Yes — models neighborhood as graph context | Visium, MERFISH | Learns niche→expression mapping; interpretable node-centric model | Requires cell-type annotations | Fischer et al., Nature Biotechnology 2022 |
| MISTy | Intra- and intercellular spatial relationships via explainable ML | Yes — multiview model with spatial ranges | Visium, IMC, seqFISH | Explainable (feature importance per spatial range); modular views | R-only; computationally heavy on large datasets | Tanevski et al., Genome Biology 2022 |
Morphology Integration¶
| Tool | What It Does | Spatial-Aware? | Platforms Tested | Key Strength | Key Limitation | Reference |
|---|---|---|---|---|---|---|
| TESLA | Super-resolution annotation combining H&E with spatial expression | Yes — pixel-level integration | Visium (spot-level only) | Subdivides spots below Visium resolution using histology | Visium spot-level only — cannot process Xenium/CosMx | Hu et al., Cell Systems 2023 |
| METI | Maps TME interactions by combining H&E morphology with spatial expression | Yes — morphological feature extraction | Visium (spot-level only) | Integrates tissue morphology into niche characterization | Visium spot-level only; calls sc.read_visium() exclusively |
Hu et al., Nature Communications 2024 |
How to Pick¶
Start with your question:
| Your Question | Stage | Recommended Tool(s) |
|---|---|---|
| What are the spatial neighborhoods in my tissue? | Recognition | CellCharter (composition) or BANKSY (expression) |
| What signaling drives each neighborhood? | Characterization | NicheCompass (spatial L-R) or COMMOT (optimal transport) |
| Which genes are niche-dependent, not just cell-type markers? | Characterization | Niche-DE |
| Do RNA niches agree with protein niches? | Recognition ×2 | CellCharter on each modality independently |
| What are the tissue regions (not niches)? | Domain detection | GraphST or STAGATE (but know these are not niches) |
| Can I get niche analysis without cell-type annotations? | Recognition | BANKSY (expression-based) or Novae (zero-shot) |
| I need to interpret niches mechanistically | Both | NicheCompass (L-R programs define niches) |
See also: What Is a Niche? | Niche vs Domain | Tool Selection for Hackathon