Skip to content

Voyager

Exploratory spatial data analysis (ESDA) on Bioconductor spatial-omics objects with geospatial statistics — a Bioconductor package by Lambda Moses (Caltech) that applies tools from geographic-information-systems and spatial statistics (Moran's I, local indicators of spatial association, variograms, Geary's C) to SpatialFeatureExperiment (SFE) data so spatial transcriptomics analysts can quantify spatial autocorrelation at the gene and cell level.

Talk at GBCC2025

  • Speakers: Lambda Moses, Bianca Dumitrascu (Columbia)
  • Day / session: Day 2 (Tue Jun 24, 2025) — Oral Session 2B, Grace Auditorium (chair: Vincent Carey)
  • Talk title: "Multiscale analysis of lung adenocarcinoma spatial transcriptomics data"
  • Slides: TBD
  • Video: GBCC2025 playlist
  • Abstract / schedule: GBCC2025 program

What it does

Voyager treats a spatial-transcriptomics sample as a geographic dataset: gene expression at each cell or spot is a value at a coordinate, and the natural questions are "where is it autocorrelated", "how do its patterns relate to a neighborhood graph", "what's the variogram". The package wraps spdep and related GIS-stats packages so that those tests can be run on SpatialFeatureExperiment objects (which carry geometries as Simple Features alongside the count matrix). The GBCC2025 talk extends this with multiscale analysis — patterns at fine-grained cellular scales often co-exist with coarser tissue-level patterns, and the LUAD case study demonstrates the package's ability to characterize both simultaneously and track how the scale–response profile shifts during tumor progression.

How it works

Core idea. Voyager imports the geographic-information-systems exploratory-spatial-data-analysis (ESDA) toolkit — neighborhood graphs, spatial autocorrelation statistics, variograms — and runs it directly on a SpatialFeatureExperiment (SFE), which stores cell/spot geometries as sf Simple Features alongside the count matrix.

Inputs / outputs. Input is an SFE object (counts + coordinates + sf geometries). Output is a set of per-gene or per-cell spatial statistics — global autocorrelation indices, local indicators of spatial association (LISA), variogram models — written back into the SFE object's metadata slots and rendered through companion plotSpatialFeature / ESDA plotting functions.

Key innovation. A geospatial-stats vocabulary applied to spatial-omics: rather than reimplementing autocorrelation tests, Voyager wraps spdep and related R-spatial libraries against an SE-family object so the same neighborhood-graph + statistic pipeline that GIS analysts use on census tracts runs on Visium spots or Xenium cells.

Parameters / API surface worth knowing. - Univariate: Moran's I (with permutation testing) and correlograms for per-gene autocorrelation. - Bivariate: Lee's L statistic and cross-variograms for gene–gene spatial dependence. - Multivariate: MULTISPATI PCA and multivariate local Geary's C for joint analysis across many features. - Local indicators: LISA-family functions to map where autocorrelation is anomalously high or low.

Canonical example. From the package docs: build an SFE from Visium output, construct a k-nearest-neighbor or contiguity weights graph, then call the univariate Moran's I on a gene set and visualize correlograms — yielding both a global "is this gene spatially patterned" answer and a map of where the local clustering is concentrated.

Where it fits in the corpus

  • AACR 2026: axis = single-cell & spatial omics; LUAD-specific spatial-progression dossier
  • OSTA (eurobioc entry) — Voyager is one of the canonical exploratory tools in the OSTA book
  • spatialLIBD (entry) — sister spatial-omics package, both target SpatialExperiment-family objects
  • SpatialData (eurobioc entry) — Python-side counterpart; cross-language interop is an active direction
  • sosta (entry) — same Day-3 session, also targets spatial-omics structure analysis
  • Nextflow Summit 2026: not mentioned

Notes

Voyager is the operational counterpart to SFE: SFE is the data structure, Voyager is the analysis layer. The multiscale angle in this GBCC talk is the methodological news — applying it to LUAD progression connects the geospatial-stats framing to a concrete cancer-biology question.