spatial-cluster-compare

Automated clustering algorithm comparison pipeline for spatial proteomics, cytometry data, or any numerical feature table.

Originally an analysis notebook, restructured into a tested, reusable Python package.

Table of contents

• Why this package
• Part of MACSima_Spatial-Omics-Pipeline
• What the pipeline does
• Installation
• Quick usage
• Advanced usage (step by step)
• Expected data format
• Roadmap

Why this package

Choosing a clustering algorithm is often an arbitrary decision: KMeans gets picked by default, parameters get tuned until the result "looks right," and the choice is rarely backed by an objective comparison. For biological data in particular where clusters are expected to reflect real cell populations this matters: the wrong algorithm or the wrong number of clusters can produce plausible-looking but biologically meaningless groups.

spatial-cluster-compare runs several clustering algorithms on the same data, scores each one on multiple complementary metrics (including bootstrap stability, not just a single internal index), and picks the best method through a transparent, reproducible composite score — instead of relying on a single, unverified default choice.

Part of MACSima_Spatial-Omics-Pipeline

spatial-cluster-compare is one of the packages of MACSima_Spatial-Omics-Pipeline, which centralizes the protocols (detailed, step-by-step documentation) and notebooks (original, standalone analyses) for the whole pipeline.

What the pipeline does

1. Clusterability test (Hopkins statistic) before/after normalization
2. Conditional normalization: StandardScaler applied only if it improves cluster structure
3. Dimensionality reduction (PCA, configurable explained variance)
4. Automatic selection of k via a composite score (Silhouette, Davies-Bouldin, Calinski-Harabasz)
5. Multi-algorithm benchmark: KMeans, Agglomerative, Spectral, GMM, DBSCAN (eps auto-estimated via the elbow method)
6. Bootstrap stability (mean ARI) for each method
7. Automatic selection of the best method based on a normalized composite score
8. Visualizations: comparative bar chart of metrics, mean-expression heatmap per cluster

Installation

• Python 3.9+
• Dependencies (numpy, pandas, scikit-learn, scipy, matplotlib, seaborn) are installed automatically.

pip install spatial-cluster-compare

For a local/development install:

pip install -e .

Quick usage

The fastest way to get a result: load your data, run the full comparison, plot it.

import pandas as pd
from spatial_cluster_compare import compare_clusters, plot_comparison_bars, plot_cluster_heatmap

data = pd.read_csv("Cluster_ImmuneCell.csv")
result = compare_clusters(data)

print(result.results_df)
print(f"Best method: {result.best_method}")

plot_comparison_bars(result.results_df)
plot_cluster_heatmap(result.X_original, result.best_labels, method_name=result.best_method)

Advanced usage (step by step)

For control over each stage of the pipeline (scaling, PCA, choice of k, benchmark) instead of the single compare_clusters call above:

from spatial_cluster_compare import (
    auto_scale, find_best_k, run_benchmark, select_best_method
)
from sklearn.decomposition import PCA

X = data.select_dtypes(include=["float64", "int64"])
X_best, report = auto_scale(X)

pca = PCA(n_components=0.9)
X_reduced = pca.fit_transform(X_best)

optimal_k, k_results = find_best_k(X_reduced, range(2, 10))

results_df, labels_per_method = run_benchmark(X_reduced, optimal_k=optimal_k)
best_method, best_labels = select_best_method(results_df, labels_per_method)

Expected data format

• One row per cell, one column per marker/descriptor. Numeric columns are used automatically as clustering features; non-numeric columns (ID, ROI name...) are ignored automatically.
• Drop non-biological numeric columns before calling compare_clusters (e.g. Cell_ID, centroid coordinates), or they'll be wrongly included:

  data = data.drop(columns=["Cell_ID", "Centroid X", "Centroid Y"])
  

• Missing values (NaN) aren't handled automatically — clean (data.dropna()) or impute beforehand.
• Dataset size: plan for at least a few hundred cells for a reliable Hopkins statistic and ARI bootstrap. Spectral Clustering can get slow beyond ~10-20k rows.

Full details (MACSiQView export settings, column naming pitfalls, biological rationale for each step) are in the protocol.

Roadmap

• Automatic PDF/figure export of the comparison report
• Support for multi-batch datasets (clustering comparison across runs)
• CLI (spatial-cluster-compare run data.csv)