Spatial Methods Workbench
Flagship research software · Open source · DOI archived
A statistical methods platform for structure in biomedical data.
The Spatial Methods Workbench transforms specialized methodology into a guided, reproducible analysis environment. Researchers upload a structured dataset, map its scientific variables, choose among compatible pipelines, run the analysis, and receive citation-ready outputs without needing to reconstruct the underlying workflow from separate scripts.
Complete video tutorial · 20 minutes
See the full research workflow before you run it.
The narrated walkthrough covers analysis-ready data, all 12 pipelines, the synthetic Moran’s I demonstration, advanced public engines, citation-ready exports, local installation, and the boundary between computation and scientific interpretation.
Live application
Run the Workbench on this page.
The embedded service may take several seconds to wake after inactivity. Use the full-screen link for large tables, long sessions, or smaller displays.
Purpose
One disciplined entry point to a broad methods program.
Spatial and imaging studies rarely fail because investigators lack another plotting function. The harder problem is choosing an analysis whose assumptions match the available coordinates, regions, identifiers, outcomes, and features—and preserving enough information for the result to be reproduced. The Workbench addresses that gap directly. It evaluates the uploaded data, exposes only compatible analysis paths, and packages the numerical output with the settings, runtime metadata, software object, figure, and required citations.
Before using real data
Read the complete installation, input, restriction, and deployment guide.
The guide covers the public Shiny service, the complete v0.3.0 Zenodo/GitHub source bundle, one-command local installation, all advanced network engines, Docker, real-data preparation, pipeline contracts, privacy, computational limits, citation, and collaboration. Open the latest v0.3.0 installation and usage guide →
For exact controls, calls, return objects, failure modes, and citations, read the complete function-by-function pipeline reference ↗.
Executable documentation · Five PDF guides
Reproduce the workflows with small simulated data.
Each vignette was rendered from executable R code with fixed seeds and compact synthetic datasets. Download a PDF for a self-contained worked example, or inspect the matching source file in the public repository.
Application walkthrough
Create all four demonstration datasets, validate a mapping, run the app-equivalent spatial workflow, and write a reproducibility bundle.
72 spatial rows · 48 image subjects · 6 contours · 72 volumetric rowsSpatial diagnostics & networks
Reproduce spatial autocorrelation, region-aware conditional networks, and the deployment checks for the exact ISPat and GP–GHS engines.
72 observations · 5 features · 19 Moran permutationsPrediction, clustering & mediation
Run held-out machine learning, spatial phenotype clustering, shallow-neural prediction, and bootstrap mediation from one compact table.
72 observations · fixed train/test seed · public workflow citationsImaging, shape & 3D
Exercise scalar-on-image screening, latent image-to-image regression, tumor contour PCA, and the ISPat 3D deployment contract.
48 image subjects · 8 features · 6 aligned contours · 2 volumesReproducibility bundles
Inspect the manifest, write the archive, audit its required artifacts, reopen the R result object, and verify embedded software and method citations.
60 observations · 19 permutations · manifest, result, figure & citationsThe PDFs demonstrate data contracts, function calls, numerical outputs, plots, and reproducibility artifacts. They do not provide biological or scientific interpretation. For study-specific interpretation or collaboration, contact Sagnik Bhadury.
Architecture
A guided analysis sequence
Every path begins with the same explicit data contract.
- 01UploadStart with a CSV or synthetic demonstration dataset.
- 02MapIdentify coordinates, regions, IDs, outcomes, exposures, mediators, and features.
- 03ValidateCheck structure and display precisely what each pipeline still requires.
- 04AnalyzeConfigure and run a compatible built-in or advanced public engine.
- 05ExportDownload results, figures, parameters, R objects, provenance, and citations.
Pipeline catalog
Twelve analysis paths, organized by scientific task
Characterize how measured features and conditional relationships vary across tissue.
Spatial exploration & autocorrelation
Map a numeric feature and test spatial patterning with permutation-based Moran’s I.
Input: x/y + feature · Output: map, Moran’s I, permutation p-valueRegion-aware conditional networks
Estimate ridge-regularized partial-correlation networks overall and within mapped tissue regions.
Input: 3+ numeric markers, optional region · Output: edge tables + heat mapsISPat 2D Bayesian network
Estimate shared and region-specific conditional-association networks after spatial adjustment.
Exact public engine · Input: x/y, zone, 3+ markers · Output: networks + full fitGP–GHS network
Estimate spatially varying edges using Gaussian-process structure with group-horseshoe shrinkage.
Exact public engine · Input: x/y + 3+ features · Output: selected adjacency + full fitMove beyond a single section to serial-section and three-dimensional tumor imaging.
ISPat 3D volumetric networks
Estimate shared and zone-specific conditional networks across registered volumetric or serial-section imaging data.
Exact public engine · Input: x/y/z, zone, volume ID, 3+ markers · Output: 3D networks + fitsAdapt public AI, ML, and deep-learning workflows to spatial biomedical prediction and discovery.
Spatial machine-learning prediction
Predict a numeric outcome from measured features and optional coordinates with held-out validation.
Input: outcome + 3+ predictors · Output: predictions, test error, variable importanceSpatial phenotype clustering
Discover multivariate phenotypes while explicitly controlling how strongly tissue coordinates influence clusters.
Input: x/y + 3+ features · Output: assignments, centers, diagnostics, spatial mapShallow-neural prediction
Fit a standardized single-hidden-layer model with weight decay and held-out evaluation.
Input: outcome + 3+ predictors · Output: held-out predictions + performance metricsAnalyze aligned contours, pixel/voxel predictors, and paired predictor–outcome images.
Wide-image scalar regression
Screen aligned pixel, voxel, or image-derived features against a numeric subject-level outcome.
Fast ridge screen—not the full SV-NN/ST-CAR posterior · Output: predictions, error, coefficientsLatent image-to-image regression
Connect aligned predictor and outcome images through low-dimensional factors and held-out reconstruction.
Fast latent screen—not the full SBLF posterior · Input: input__* + output__* columnsTumor contour & shape PCA
Resample and align ordered boundaries, then summarize tumor morphology with principal components.
Input: contour ID + ordered x/y landmarks · Output: aligned shapes, scores, varianceEstimate a prespecified indirect-effect model while keeping causal requirements explicit.
Mediation analysis
Estimate model-based direct, indirect, and total effects with bootstrap uncertainty and explicit causal requirements.
Input: exposure + mediator + outcome · Output: effect estimates + bootstrap intervalsPublic workflow ecosystem
Broader than a single package—and explicit about attribution.
AI & ML curricula
The AI for Beginners and ML for Beginners repositories inform executable spatial clustering and predictive workflows.
Deep-learning foundations
A Deep Understanding of Deep Learning informs the tested shallow-neural extension, adapted to coordinates and biomedical features.
SV-NN direction
The image-regression tier cites Wu, Wu & Kang’s Bayesian spatially varying neural-network model while labeling the interactive ridge implementation as a screening workflow.
ST-CAR direction
Scalar-on-image and image-on-scalar extensions are connected to the soft-thresholded CAR image-regression publication; full posterior engines remain deployment-gated.
SBLF direction
The paired-image workflow follows the latent-factor problem structure established by Guo, Kang & Johnson’s spatial Bayesian image-on-image model, with exact method identity retained in every label.
Data-isolation boundary
Uploaded values and results remain inside the active analysis service and are not sent to a third-party AI interpretation system. The platform does not automate scientific interpretation.
Built for credible research use
Compatibility before computation. Provenance after it.
Explicit compatibility gates
The interface tells researchers what is ready, what is missing, and whether an advanced engine exists on the deployment.
Exact engine identity
An unavailable Bayesian or 3D method is never replaced by a simpler algorithm under the same name.
Reproducibility bundles
Each download preserves results, a figure, parameters, runtime metadata, the R result object, and software citations.
Citation by design
Users acknowledge the citation condition before analysis. Citation metadata are embedded in every export and the software has a permanent Zenodo DOI.
Research-data caution
The public workflow defaults to synthetic data and explicitly prohibits protected health information on an unsecured deployment.
Tested application behavior
Automated analysis, export, citation, and Shiny server tests protect the public workflow as the platform evolves.
Scientific interpretation
The platform computes results. Scientific conclusions require collaboration.
The Workbench deliberately does not narrate results, assign biological mechanisms, or generate automated explanations. Interpretation depends on study design, tissue context, preprocessing, measurement technology, and the scientific question.
For study-specific interpretation, methodological consultation, or a research collaboration, contact Sagnik Bhadury directly.
Discuss a collaborationSoftware record
A citable platform, not an ephemeral web demo.
Version v0.3.0: Bhadury, S. (2026). SpatialMethodsWorkbench: Spatial Methods Workbench (v0.3.0) [Computer software]. Zenodo.
The all-versions record is doi:10.5281/zenodo.21763606. Individual statistical engines retain their applicable publication and software citations, which the Workbench adds to the corresponding result bundle.
Version history
A permanent record of the Workbench's development.
v0.3.0 is the current recommended release. It supersedes the earlier archives with the complete public-source installer, bundled public method engines, executable vignettes, compact simulated data, expanded documentation, and 62 passing tests. Earlier releases remain available for citation and reproducibility of analyses performed with those versions.
| Version | Status | GitHub release | Zenodo record |
|---|---|---|---|
| v0.3.0 | Current complete public release | Download and release notes | doi:10.5281/zenodo.21777139 |
| v0.2.1 | Archived | Release | doi:10.5281/zenodo.21764196 |
| v0.2.0 | Archived | Release | doi:10.5281/zenodo.21763796 |
| v0.1.1 | Archived | Release | doi:10.5281/zenodo.21763607 |
| v0.1.0 | Initial GitHub release | Release | — |
The permanent all-versions DOI is doi:10.5281/zenodo.21763606; it resolves to the newest live Zenodo version. For exact reproducibility, cite the version-specific DOI used in an analysis.
Funding acknowledgement
National Institutes of Health support
Research reported in this software was supported by the National Cancer Institute of the National Institutes of Health under Award Number R37CA214955-01A1. The content is solely the responsibility of the author and does not necessarily represent the official views of the National Institutes of Health.