Batch Integration with Linked NMF

Zach DeBruine

2022-09-09

Get Started

Use the Seurat ifnb dataset:

library(Seurat)
library(ggplot2)
library(singlet)
library(SeuratData)
library(cowplot)
data(ifnb)
ifnb <- NormalizeData(ifnb, verbose = FALSE)

Determine NMF Rank

First we determine number of factors using cross-validation.

Note the use of split.by = "stim", where we are indicating that the “stim” field in the meta.data slot of our Seurat object is a factor giving discrete groupings of samples (either stimulated or unstimulated). The RunNMF function will weight samples from both groups equally in the NMF objective, regardless of whether the groups are of equal size.

set.seed(123)
ifnb <- RunNMF(ifnb, split.by = "stim")

RankPlot(ifnb) + scale_y_continuous(limits = c(1, 1.1))

## Scale for 'y' is already present. Adding another scale for 'y', which will
## replace the existing scale.

## Warning: Removed 1 rows containing missing values (geom_point).

## Warning: Removed 1 row(s) containing missing values (geom_path).

Visualize contribution of groups to both factors:

MetadataPlot(ifnb, split.by = "stim", reduction = "nmf")

Some factors are almost exclusively explaining signal from one dataset, and not the other.

Run Linked NMF

Linked NMF will uncouple sample groups from factors in which they are only weakly represented. LNMF is initialized with the joint NMF model that we trained before, we just specify a cutoff for the minimum fractional representation of any sample group in any given factor at which it will be uncoupled from the factor.

ifnb <- RunLNMF(
  ifnb, 
  split.by = "stim", 
  reduction.use = "nmf", 
  link.cutoff = 0.7, 
  verbose = FALSE)

LNMF creates a new reduction in the Seurat object, lnmf. Now examine how each group is represented in NMF factors:

MetadataPlot(ifnb, split.by = "stim", reduction = "lnmf")

We can visualize these models on UMAP coordinates using the joint model, the entire linked NMF model, and the linked NMF model using only shared factors:

ifnb <- RunUMAP(ifnb, 
                reduction = "nmf", 
                dims = 1:ncol(ifnb@reductions$nmf), 
                reduction.name = "jnmf_all", 
                verbose = FALSE)

ifnb <- RunUMAP(ifnb, 
                reduction = "lnmf", 
                dims = GetSharedFactors(ifnb, split.by = "stim"), 
                reduction.name = "lnmf_shared", 
                verbose = FALSE)

p_jnmf_umap <- DimPlot(ifnb, reduction = "jnmf_all", group.by = "stim")
p_lnmf_umap <- DimPlot(ifnb, reduction = "lnmf_shared", group.by = "stim")

Visualize

Plot the results:

plot_grid(
  p_jnmf_umap + 
    ggtitle("joint NMF") + 
    theme(legend.position = "none"), 
  p_lnmf_umap + 
    ggtitle("linked NMF") + 
    theme(legend.position = "none"), 
  get_legend(p_jnmf_umap), 
  ncol = 3, 
  rel_widths = c(1, 1, 0.2)
)