Skip to contents

metamoRph

A framework for projecting query (new) data onto an reference PCA space. This is easiest as a two step process where the user runs our run_pca function (which wraps prcomp and provides some sensible defaults and enhanced outputs. After run_pca the user can then use metamoRph to normalize the query (new) data in the same manner as the reference data and then project it onto the existing reference PCA.

We also provides two functions, model_build and model_apply to very quickly transfer metadata onto new data using the shared PCA space between the reference and the query.

Quick Start 

remotes::install_github("davemcg/metamoRph")
library(metamoRph)
library(dplyr)
library(ggplot2)
library(dplyr)
library(ggplot2)

genes <- c('RHO','RPE65','ABCA4','PMEL', 'KRT5')
samples <- c('Retina1','Retina2','RPE1','RPE2','Cornea1')

faux_mat <- cbind(c(560,650,5,6,0), # rho
                  c(42,32, 1103,1201,2), #rpe65
                  c(810,903,202,205,45),  #abca4
                  c(100,105,2004,1980,101),# pmel
                  c(3,32,101,202,1567)) |> data.frame()# krt5
colnames(faux_mat) <- genes
row.names(faux_mat) <- samples

new_data <- cbind(c(5,6,3), # rho
                  c(32,23,54), #rpe65
                  c(65,22,10),  #abca4
                  c(122,101,57),# pmel
                  c(2567,1755,2218)) * 100 # krt5
colnames(new_data) <- genes
row.names(new_data) <- c("Cornea2","Cornea3","Cornea4")

mm_pca <- run_pca(t(faux_mat), meta = samples |> data.frame())
projected_pca <- metamoRph(t(new_data), 
                               mm_pca$PCA$rotation, 
                               center_scale = mm_pca$center_scale)

# example plot
# bind_rows(as_tibble(mm_pca$PCA$x, rownames = 'samples'),
#       as_tibble(projected_pca, rownames = 'Cornea')) |>
#   mutate(samples = gsub('\\d','',samples)) |>
#   ggplot(aes(x=PC1,y=PC2,color=samples)) + 
#   geom_point() +
#   cowplot::theme_cowplot()

Label projection 

# continue from code chunk above
## WARNING: Use many more num_PCs (20+) for "real" genomic data
trained_model <- model_build(mm_pca$PCA$x,
                             gsub('\\d+','', mm_pca$meta$samples), #remove trailing digit 
                             model = 'lm', num_PCs = 2)

label_guesses <- model_apply(trained_model,
                             projected_pca,
                             c("Cornea","Cornea","Cornea")
)

# label_guesses

Existing prcomp object projection example (pseudocode-ish)

It is also possible to use a pre-existing prcomp object with metamoRph, skipping the run_pca step. You will have to ensure that you sample scale your new_data in the same manner that you used for your prcomp run (here we use log2 - but replace with whatever scaling your have done. Maybe none?). Note how we have turned off the cpm and log1p scaling in the metamoRph function with normaliztion = FALSE.

library(metamoRph)
projected_pca <- metamoRph(t(log2(new_data)), 
                               your_prcomp_object$PCA$rotation, 
                               center_scale = extract_prcomp_scaling(your_prcomp_object),
                               normalization = FALSE)

Some Related Tools

  • ProjecTILs
    • scRNA/Seurat based tool that uses a similar approach to project data (essentialy multiplying the new against the rotation matrix)
  • Seurat
    • The big kahuna itself. scRNA analysis playground.
  • projectR
    • Can transfer data onto PCA and NMF
    • At least for PCA does not provide any guardrails for properly scaling the query data