Parameter selection (VeraFISH Mouse Hippocampus)
7 October 2025
Contents
Here, we demonstrate a grid search of clustering parameters with a mouse
hippocampus VeraFISH dataset. BANKSY currently provides four algorithms for
clustering the BANKSY matrix with clusterBanksy: Leiden (default), Louvain,
k-means, and model-based clustering. In this vignette, we run only Leiden
clustering. See ?clusterBanksy for more details on the parameters for
different clustering methods.
1 Loading the data
The dataset comprises gene expression for 10,944 cells and 120 genes in 2
spatial dimensions. See ?Banksy::hippocampus for more details.
# Load libs
library(Banksy)
library(SummarizedExperiment)
library(SpatialExperiment)
library(scuttle)
library(scater)
library(cowplot)
library(ggplot2)
# Load data
data(hippocampus)
gcm <- hippocampus$expression
locs <- as.matrix(hippocampus$locations)Here, gcm is a gene by cell matrix, and locs is a matrix specifying the
coordinates of the centroid for each cell.
head(gcm[,1:5])
#> cell_1276 cell_8890 cell_691 cell_396 cell_9818
#> Sparcl1 45 0 11 22 0
#> Slc1a2 17 0 6 5 0
#> Map 10 0 12 16 0
#> Sqstm1 26 0 0 2 0
#> Atp1a2 0 0 4 3 0
#> Tnc 0 0 0 0 0
head(locs)
#> sdimx sdimy
#> cell_1276 -13372.899 15776.37
#> cell_8890 8941.101 15866.37
#> cell_691 -14882.899 15896.37
#> cell_396 -15492.899 15835.37
#> cell_9818 11308.101 15846.37
#> cell_11310 14894.101 15810.37Initialize a SpatialExperiment object and perform basic quality control. We keep cells with total transcript count within the 5th and 98th percentile:
se <- SpatialExperiment(assay = list(counts = gcm), spatialCoords = locs)
colData(se) <- cbind(colData(se), spatialCoords(se))
# QC based on total counts
qcstats <- perCellQCMetrics(se)
thres <- quantile(qcstats$total, c(0.05, 0.98))
keep <- (qcstats$total > thres[1]) & (qcstats$total < thres[2])
se <- se[, keep]Next, perform normalization of the data.
# Normalization to mean library size
se <- computeLibraryFactors(se)
aname <- "normcounts"
assay(se, aname) <- normalizeCounts(se, log = FALSE)2 Parameters
BANKSY has a few key parameters. We describe these below.
2.1 AGF usage
For characterising neighborhoods, BANKSY computes the weighted neighborhood
mean (H_0) and the azimuthal Gabor filter (H_1), which estimates gene
expression gradients. Setting compute_agf=TRUE computes both H_0 and H_1.
2.2 k-geometric
k_geom specifies the number of neighbors used to compute each H_m for
m=0,1. If a single value is specified, the same k_geom will be used
for each feature matrix. Alternatively, multiple values of k_geom can be
provided for each feature matrix. Here, we use k_geom[1]=15 and
k_geom[2]=30 for H_0 and H_1 respectively. More neighbors are used to
compute gradients.
For datasets generated using Visium v1/v2, use
k_geom=18(ork_geom <- c(18, 18)ifcompute_agf = TRUE), since that corresponds to taking as neighbourhood two concentric rings of spots around each spot.
We compute the neighborhood feature matrices using normalized expression
(normcounts in the se object).
k_geom <- c(15, 30)
se <- computeBanksy(se, assay_name = aname, compute_agf = TRUE, k_geom = k_geom)
#> Computing neighbors...
#> Spatial mode is kNN_median
#> Parameters: k_geom=15
#> Done
#> Computing neighbors...
#> Spatial mode is kNN_median
#> Parameters: k_geom=30
#> Done
#> Done
#> Centering
#> DonecomputeBanksy populates the assays slot with H_0 and H_1 in this
instance:
se
#> class: SpatialExperiment
#> dim: 120 10205
#> metadata(1): BANKSY_params
#> assays(4): counts normcounts H0 H1
#> rownames(120): Sparcl1 Slc1a2 ... Notch3 Egfr
#> rowData names(0):
#> colnames(10205): cell_1276 cell_691 ... cell_11635 cell_10849
#> colData names(4): sample_id sdimx sdimy sizeFactor
#> reducedDimNames(0):
#> mainExpName: NULL
#> altExpNames(0):
#> spatialCoords names(2) : sdimx sdimy
#> imgData names(1): sample_id2.3 lambda
The lambda parameter is a mixing parameter in [0,1] which
determines how much spatial information is incorporated for downstream analysis.
With smaller values of lambda, BANKY operates in cell-typing mode, while at
higher levels of lambda, BANKSY operates in domain-finding mode. As a
starting point, we recommend lambda=0.2 for cell-typing and lambda=0.8 for
zone-finding, except for datasets generated using the Visium v1/v2 technology, for which
we recommend
lambda=0.2 for domain finding. See the note in the tutorial on the
main page for more info.
Here, we run lambda=0 which corresponds to non-spatial
clustering, and lambda=0.2 for spatially-informed cell-typing. We compute PCs
with and without the AGF (H_1).
lambda <- c(0, 0.2)
se <- runBanksyPCA(se, use_agf = c(FALSE, TRUE), lambda = lambda, seed = 1000)
#> Using seed=1000
#> Using seed=1000
#> Using seed=1000
#> Using seed=1000runBanksyPCA populates the reducedDims slot, with each combination of
use_agf and lambda provided.
reducedDimNames(se)
#> [1] "PCA_M0_lam0" "PCA_M0_lam0.2" "PCA_M1_lam0" "PCA_M1_lam0.2"2.4 Clustering parameters
Next, we cluster the BANKSY embedding with Leiden graph-based clustering. This
admits two parameters: k_neighbors and resolution. k_neighbors determines
the number of k nearest neighbors used to construct the shared nearest
neighbors graph. Leiden clustering is then performed on the resultant graph
with resolution resolution. For reproducibiltiy we set a seed for each
parameter combination.
k <- 50
res <- 1
se <- clusterBanksy(se, use_agf = c(FALSE, TRUE), lambda = lambda, k_neighbors = k, resolution = res, seed = 1000)
#> Using seed=1000
#> Using seed=1000
#> Using seed=1000
#> Using seed=1000clusterBanksy populates colData(se) with cluster labels:
colnames(colData(se))
#> [1] "sample_id" "sdimx"
#> [3] "sdimy" "sizeFactor"
#> [5] "clust_M0_lam0_k50_res1" "clust_M0_lam0.2_k50_res1"
#> [7] "clust_M1_lam0_k50_res1" "clust_M1_lam0.2_k50_res1"3 Comparing cluster results
To compare clustering runs visually, different runs can be relabeled to
minimise their differences with connectClusters:
se <- connectClusters(se)
#> clust_M1_lam0_k50_res1 --> clust_M0_lam0_k50_res1
#> clust_M0_lam0.2_k50_res1 --> clust_M1_lam0_k50_res1
#> clust_M1_lam0.2_k50_res1 --> clust_M0_lam0.2_k50_res1Visualise spatial coordinates with cluster labels.
cnames <- colnames(colData(se))
cnames <- cnames[grep("^clust", cnames)]
cplots <- lapply(cnames, function(cnm) {
plotColData(se, x = "sdimx", y = "sdimy", point_size = 0.1, colour_by = cnm) +
coord_equal() +
labs(title = cnm) +
theme(legend.title = element_blank()) +
guides(colour = guide_legend(override.aes = list(size = 2)))
})
plot_grid(plotlist = cplots, ncol = 2)
Compare all cluster outputs with compareClusters. This function computes
pairwise cluster comparison metrics between the clusters in colData(se) based
on adjusted Rand index (ARI):
compareClusters(se, func = "ARI")
#> clust_M0_lam0_k50_res1 clust_M0_lam0.2_k50_res1
#> clust_M0_lam0_k50_res1 1.000 0.67
#> clust_M0_lam0.2_k50_res1 0.670 1.00
#> clust_M1_lam0_k50_res1 1.000 0.67
#> clust_M1_lam0.2_k50_res1 0.747 0.87
#> clust_M1_lam0_k50_res1 clust_M1_lam0.2_k50_res1
#> clust_M0_lam0_k50_res1 1.000 0.747
#> clust_M0_lam0.2_k50_res1 0.670 0.870
#> clust_M1_lam0_k50_res1 1.000 0.747
#> clust_M1_lam0.2_k50_res1 0.747 1.000or normalized mutual information (NMI):
compareClusters(se, func = "NMI")
#> clust_M0_lam0_k50_res1 clust_M0_lam0.2_k50_res1
#> clust_M0_lam0_k50_res1 1.000 0.741
#> clust_M0_lam0.2_k50_res1 0.741 1.000
#> clust_M1_lam0_k50_res1 1.000 0.741
#> clust_M1_lam0.2_k50_res1 0.782 0.915
#> clust_M1_lam0_k50_res1 clust_M1_lam0.2_k50_res1
#> clust_M0_lam0_k50_res1 1.000 0.782
#> clust_M0_lam0.2_k50_res1 0.741 0.915
#> clust_M1_lam0_k50_res1 1.000 0.782
#> clust_M1_lam0.2_k50_res1 0.782 1.000See ?compareClusters for the full list of comparison measures.
4 Session information
Vignette runtime:
#> Time difference of 52.86112 secssessionInfo()
#> R version 4.5.1 Patched (2025-08-23 r88802)
#> Platform: x86_64-pc-linux-gnu
#> Running under: Ubuntu 24.04.3 LTS
#>
#> Matrix products: default
#> BLAS: /home/biocbuild/bbs-3.22-bioc/R/lib/libRblas.so
#> LAPACK: /usr/lib/x86_64-linux-gnu/lapack/liblapack.so.3.12.0 LAPACK version 3.12.0
#>
#> locale:
#> [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
#> [3] LC_TIME=en_GB LC_COLLATE=C
#> [5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
#> [7] LC_PAPER=en_US.UTF-8 LC_NAME=C
#> [9] LC_ADDRESS=C LC_TELEPHONE=C
#> [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
#>
#> time zone: America/New_York
#> tzcode source: system (glibc)
#>
#> attached base packages:
#> [1] stats4 stats graphics grDevices utils datasets methods
#> [8] base
#>
#> other attached packages:
#> [1] ExperimentHub_2.99.5 AnnotationHub_3.99.6
#> [3] BiocFileCache_2.99.6 dbplyr_2.5.1
#> [5] spatialLIBD_1.21.6 cowplot_1.2.0
#> [7] scater_1.37.0 ggplot2_4.0.0
#> [9] harmony_1.2.3 Rcpp_1.1.0
#> [11] data.table_1.17.8 scran_1.37.0
#> [13] scuttle_1.19.0 Seurat_5.3.0
#> [15] SeuratObject_5.2.0 sp_2.2-0
#> [17] SpatialExperiment_1.19.1 SingleCellExperiment_1.31.1
#> [19] SummarizedExperiment_1.39.2 Biobase_2.69.1
#> [21] GenomicRanges_1.61.5 Seqinfo_0.99.2
#> [23] IRanges_2.43.5 S4Vectors_0.47.4
#> [25] BiocGenerics_0.55.1 generics_0.1.4
#> [27] MatrixGenerics_1.21.0 matrixStats_1.5.0
#> [29] Banksy_1.5.8 BiocStyle_2.37.1
#>
#> loaded via a namespace (and not attached):
#> [1] bitops_1.0-9 spatstat.sparse_3.1-0 httr_1.4.7
#> [4] RColorBrewer_1.1-3 doParallel_1.0.17 tools_4.5.1
#> [7] sctransform_0.4.2 DT_0.34.0 R6_2.6.1
#> [10] lazyeval_0.2.2 uwot_0.2.3 GetoptLong_1.0.5
#> [13] withr_3.0.2 gridExtra_2.3 progressr_0.16.0
#> [16] cli_3.6.5 spatstat.explore_3.5-3 fastDummies_1.7.5
#> [19] labeling_0.4.3 sass_0.4.10 S7_0.2.0
#> [22] spatstat.data_3.1-8 ggridges_0.5.7 pbapply_1.7-4
#> [25] Rsamtools_2.25.3 dbscan_1.2.3 aricode_1.0.3
#> [28] dichromat_2.0-0.1 sessioninfo_1.2.3 parallelly_1.45.1
#> [31] attempt_0.3.1 maps_3.4.3 limma_3.65.5
#> [34] pals_1.10 RSQLite_2.4.3 BiocIO_1.19.0
#> [37] shape_1.4.6.1 ica_1.0-3 spatstat.random_3.4-2
#> [40] dplyr_1.1.4 Matrix_1.7-4 ggbeeswarm_0.7.2
#> [43] abind_1.4-8 lifecycle_1.0.4 yaml_2.3.10
#> [46] edgeR_4.7.5 SparseArray_1.9.1 Rtsne_0.17
#> [49] paletteer_1.6.0 grid_4.5.1 blob_1.2.4
#> [52] promises_1.3.3 dqrng_0.4.1 crayon_1.5.3
#> [55] miniUI_0.1.2 lattice_0.22-7 beachmat_2.25.5
#> [58] mapproj_1.2.12 KEGGREST_1.49.1 magick_2.9.0
#> [61] pillar_1.11.1 knitr_1.50 ComplexHeatmap_2.25.2
#> [64] metapod_1.17.0 rjson_0.2.23 future.apply_1.20.0
#> [67] codetools_0.2-20 glue_1.8.0 spatstat.univar_3.1-4
#> [70] vctrs_0.6.5 png_0.1-8 spam_2.11-1
#> [73] gtable_0.3.6 rematch2_2.1.2 cachem_1.1.0
#> [76] xfun_0.53 S4Arrays_1.9.1 mime_0.13
#> [79] survival_3.8-3 RcppHungarian_0.3 iterators_1.0.14
#> [82] tinytex_0.57 statmod_1.5.0 bluster_1.19.0
#> [85] fitdistrplus_1.2-4 ROCR_1.0-11 nlme_3.1-168
#> [88] bit64_4.6.0-1 filelock_1.0.3 RcppAnnoy_0.0.22
#> [91] bslib_0.9.0 irlba_2.3.5.1 vipor_0.4.7
#> [94] KernSmooth_2.23-26 colorspace_2.1-2 DBI_1.2.3
#> [97] tidyselect_1.2.1 bit_4.6.0 compiler_4.5.1
#> [100] curl_7.0.0 httr2_1.2.1 BiocNeighbors_2.3.1
#> [103] DelayedArray_0.35.3 plotly_4.11.0 rtracklayer_1.69.1
#> [106] bookdown_0.45 scales_1.4.0 lmtest_0.9-40
#> [109] rappdirs_0.3.3 stringr_1.5.2 digest_0.6.37
#> [112] goftest_1.2-3 spatstat.utils_3.2-0 rmarkdown_2.30
#> [115] benchmarkmeData_1.0.4 RhpcBLASctl_0.23-42 XVector_0.49.1
#> [118] htmltools_0.5.8.1 pkgconfig_2.0.3 fastmap_1.2.0
#> [121] GlobalOptions_0.1.2 rlang_1.1.6 htmlwidgets_1.6.4
#> [124] shiny_1.11.1 farver_2.1.2 jquerylib_0.1.4
#> [127] zoo_1.8-14 jsonlite_2.0.0 BiocParallel_1.43.4
#> [130] mclust_6.1.1 config_0.3.2 RCurl_1.98-1.17
#> [133] BiocSingular_1.25.0 magrittr_2.0.4 dotCall64_1.2
#> [136] patchwork_1.3.2 viridis_0.6.5 reticulate_1.43.0
#> [139] leidenAlg_1.1.5 stringi_1.8.7 MASS_7.3-65
#> [142] plyr_1.8.9 parallel_4.5.1 listenv_0.9.1
#> [145] ggrepel_0.9.6 deldir_2.0-4 Biostrings_2.77.2
#> [148] sccore_1.0.6 splines_4.5.1 tensor_1.5.1
#> [151] circlize_0.4.16 locfit_1.5-9.12 igraph_2.1.4
#> [154] spatstat.geom_3.6-0 RcppHNSW_0.6.0 reshape2_1.4.4
#> [157] ScaledMatrix_1.17.0 XML_3.99-0.19 BiocVersion_3.22.0
#> [160] evaluate_1.0.5 golem_0.5.1 BiocManager_1.30.26
#> [163] foreach_1.5.2 httpuv_1.6.16 RANN_2.6.2
#> [166] tidyr_1.3.1 purrr_1.1.0 polyclip_1.10-7
#> [169] benchmarkme_1.0.8 clue_0.3-66 future_1.67.0
#> [172] scattermore_1.2 rsvd_1.0.5 xtable_1.8-4
#> [175] restfulr_0.0.16 RSpectra_0.16-2 later_1.4.4
#> [178] viridisLite_0.4.2 tibble_3.3.0 GenomicAlignments_1.45.4
#> [181] memoise_2.0.1 beeswarm_0.4.0 AnnotationDbi_1.71.1
#> [184] cluster_2.1.8.1 shinyWidgets_0.9.0 globals_0.18.0