MOFA+: downstream analysis in R (original) (raw)

Introduction

In the MOFA2 R package we provide a wide range of downstream analysis to visualise and interpret the model output. Here we provide a brief description of the main functionalities. This vignette is made of simulated data and we do not highlight biologically relevant results. Please see our tutorials for real use cases.

Load libraries

library(ggplot2)
library(MOFA2)

Load trained model

filepath <- system.file("extdata", "model.hdf5", package = "MOFA2")
model <- load_model(filepath)

Overview of data

The function plot_data_overview can be used to obtain an overview of the input data. It shows how many views (rows) and how many groups (columns) exist, what are their corresponding dimensionalities and how many missing information they have (grey bars).

plot_data_overview(model)

Variance decomposition

The first step in the MOFA analysis is to quantify the amount of variance explained (\(R^2\)) by each factor in each data modality.
The variance explained estimates are stored in the hdf5 file and loaded in model@cache[["variance_explained"]]:

# Total variance explained per view
head(get_variance_explained(model)$r2_total[[1]])
##   view_0   view_1 
## 76.20973 76.97777
# Variance explained for every factor in per view
head(get_variance_explained(model)$r2_per_factor[[1]])
##              view_0      view_1
## Factor1 19.20399955 19.41070871
## Factor2 15.47560732 17.94710458
## Factor3 16.47469843 16.48996544
## Factor4 13.42094721 11.09844071
## Factor5 11.76334520 11.82572116
## Factor6  0.03885712  0.07087386

Variance explained estimates can be plotted using plot_variance_explained(model, ...). Options:

In this case we have 5 active factors that explain a large amount of variation in both data modalities.

plot_variance_explained(model, x="view", y="factor")

The model explains ~70% of the variance in both data modalities.

plot_variance_explained(model, x="view", y="factor", plot_total = TRUE)[[2]]

Visualisation of Factors

The MOFA factors capture the global sources of variability in the data. Mathematically, each factor ordinates cells along a one-dimensional axis centered at zero. The value per se is not important, only the relative positioning of samples matters. Samples with different signs manifest opposite “effects” along the inferred axis of variation, with higher absolute value indicating a stronger effect. Note that the interpretation of factors is analogous to the interpretation of the principal components in PCA.

Visualisation of factors one at a time

Factors can be plotted using plot_factor (for beeswarm plots of individual factors) orplot_factors (for scatter plots of factor combinations).

plot_factor(model, 
  factor = 1:3,
  color_by = "age",
  shape_by = "condition"
)

Adding more options

p <- plot_factor(model, 
  factors = c(1,2,3),
  color_by = "condition",
  dot_size = 3,        # change dot size
  dodge = TRUE,           # dodge points with different colors
  legend = FALSE,          # remove legend
  add_violin = TRUE,      # add violin plots,
  violin_alpha = 0.25  # transparency of violin plots
)

# The output of plot_factor is a ggplot2 object that we can edit
p <- p + 
  scale_color_manual(values=c("A"="black", "B"="red")) +
  scale_fill_manual(values=c("A"="black", "B"="red"))

print(p)
## Warning: No shared levels found between `names(values)` of the manual scale and the
## data's colour values.

Visualisation of combinations of factors

Scatter plots

plot_factors(model, 
  factors = 1:3,
  color_by = "condition"
)

Visualisation of feature weights

The weights provide a score for how strong each feature relates to each factor. Features with no association with the factor have values close to zero, while features with strong association with the factor have large absolute values. The sign of the weight indicates the direction of the effect: a positive weight indicates that the feature has higher levels in the cells with positive factor values, and vice versa.

Weights can be plotted using plot_weights or plot_top_weights

plot_weights(model,
  view = "view_0",
  factor = 1,
  nfeatures = 10,     # Number of features to highlight
  scale = TRUE,          # Scale weights from -1 to 1
  abs = FALSE             # Take the absolute value?
)
plot_top_weights(model,
  view = "view_0",
  factor = 1,
  nfeatures = 10
)

Visualisation of covariation patterns in the input data

Instead of looking at weights, it is useful to observe the coordinated heterogeneity that MOFA captures in the original data. This can be done using the plot_data_heatmap and plot_data_scatter function.

Heatmaps

Heatmap of observations. Top features are selected by its weight in the selected factor. By default, samples are ordered according to their corresponding factor value.

plot_data_heatmap(model,
  view = "view_1",         # view of interest
  factor = 1,             # factor of interest
  features = 20,          # number of features to plot (they are selected by weight)
  
  # extra arguments that are passed to the `pheatmap` function
  cluster_rows = TRUE, cluster_cols = FALSE,
  show_rownames = TRUE, show_colnames = FALSE
)

Scatter plots

Scatter plots of observations vs factor values. It is useful to add a linear regression estimate to visualise if the relationship between (top) features and factor values is linear.

plot_data_scatter(model,
  view = "view_1",         # view of interest
  factor = 1,             # factor of interest
  features = 5,           # number of features to plot (they are selected by weight)
  add_lm = TRUE,          # add linear regression
  color_by = "condition"
)

Non-linear dimensionality reduction

The MOFA factors are linear (as in Principal Component analysis), so each one captures limited amount of information, but they can be used as input to other methods that learn compact nonlinear manifolds, e.g. t-SNE or UMAP.

Run UMAP or t-SNE

set.seed(42)
model <- run_umap(model)
model <- run_tsne(model)

Plot (nothing too interesting in this simulated data set)

plot_dimred(model,
  method = "TSNE",  # method can be either "TSNE" or "UMAP"
  color_by = "condition",
  dot_size = 5
)
## Warning: The `<scale>` argument of `guides()` cannot be `FALSE`. Use "none" instead as
## of ggplot2 3.3.4.
## ℹ The deprecated feature was likely used in the MOFA2 package.
##   Please report the issue at <https://github.com/bioFAM/MOFA2>.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.

Other functionalities

Renaming dimensions

The user can rename the dimensions of the model

views_names(model) <- c("Transcriptomics", "Proteomics")
factors_names(model) <- paste("Factor", 1:get_dimensions(model)$K, sep=" ")
views_names(model)
## [1] "Transcriptomics" "Proteomics"