The NeighbourNet (NNet) package is currently under development.

The data and reproducible R scripts used to generate the figures in the manuscript can be found at
.

Note that the functions implemented in the final R package will differ from those used in the analysis presented in our paper.
The main difference lies in how NNet prunes the inferred co-expression networks.
The functions used in our paper’s analysis are provided in this script and can be directly sourced into R’s global environment.

In comparison, the pruning strategy implemented by the R package is more statistically rigorous, although it does not necessarily yield better results in our numerical evaluation.
We conducted an analysis, available here, to compare and understand the differences between the two strategies.

⚡ 5-Minute Quick-Start

This mini-walkthrough reproduces the full vignette in about 25 lines.
The full vignette can be found here (script) and here (package).

1 · Install / load core packages

remotes::install_github("https://github.com/meiosis97/NeighbourNet")

2 · Fetch demo dataset (can be found in the data folder)

# Load the package
require(NeighbourNet)
require(ggplot2)
require(Seurat)

# Seurat object with ~4k LUAD cells available under the misc folder of the repository
load("misc/luad.rda")    # loads `obj`

3 · One-liner preprocessing

genes  <- select.gene(obj, min.cells = 10) # QC → TF / target lists

obj <- obj |>
  prepare.seurat(genes = genes$genes) |>   # scale + PCA
  prepare.graph() |>                       # 30-NN graph
  select.cell() |>                         # subsample 
  prepare.reg(predictors = genes$tfs,      # local variance scaffolding
              responses  = genes$targets)

4 · Run NeighbourNet and build meta-networks

top10 <- head(genes$targets, 10)           # demo: first 10 targets
obj   <- run.nn.reg(obj, responses = top10, return.p.val = TRUE) |>
         build.meta.network()

obj@misc$mod now contains:

slot	description
`effect`	(response × predictor × cell) co-expression tensor
`p.val`	matching significance tensor
`meta.network`	(response × predictor × meta-cell) ensemble

5.0 Prepare Visualisation

ctr.genes <- select.central.genes(obj)  
obj <- prepare.visualise(obj, central.genes = ctr.genes)

5.1 · Snapshot plot (Cell #1)

visualise.network(obj, 1, 
                  radius = c(.4,.7,.85,1), pie.radius = .04,
                  text.size = 5)

5.2 · Snapshot plot (meta-network #1)

cut <- mean(apply(obj@misc$NNet.mod$meta.network$p.val[,,1], 1, max))
visualise.network(obj, 1, meta.network = TRUE, cutoff = cut,
                  radius = c(.4,.7,.85,1), pie.radius = .04,
                  text.size = 5)

6 · (Option) Receptor activity per cell

act  <- receptor.activity(obj)             # matrix: receptor × cell
lrp6 <- act$receptor.act["LRP6", ]

Plot on PCA:

ggplot(Embeddings(obj, "pca"), aes(PC_1, PC_2)) +
  geom_point(alpha = .2, size = .6, colour = "grey80") +
  geom_point(data = Embeddings(obj, "pca")[names(lrp6), ],
             aes(col = as.numeric(lrp6)), size = 1)

✨ That’s it

You now have a full NeighbourNet analysis in under two minutes of run-time. Tweak gene lists, increase responses, or switch visualisation parameters exactly as in the full vignette.