GEO RNA-seq Cancer Meta-analysis


This app integrates differential expression (DE) results generated from a uniform
processing pipeline into an interactive meta-analysis framework. The app allows
for clustering of DE results, dimensionality reduction, exploration of volcano
and MA plots, gene set enrichment, and over-representation analysis.


Meta-analysis


The meta-analysis tab enables the exploration of DE results across many
experiments. Each of it’s sub-tabs allows for different levels of across
aggregated differential expression results.


1. Data Selection


The data selection tab provides an interface for selecting contrasts
(i.e. Treatment vs Control comparisons) that will be included in any page that
falls under the “meta-analysis” tab. Users can use the dropdowns to dynamically
select subsets of contrasts by their BioProject ID, drug class, tissue type, etc.
By default, every contrast in the database is used.


2. PCA


PCA is performed on data from the selected contrasts. The user can opt to
perform PCA on either the genes alone, TEs alone, or both. The input data values
for PCA are one of the values computed from the differential expression analysis
for a particular feature, e.g. logFC, P-value, or z-statistic.
Using these values for dimensionality reduction allows for the exploration of
patterns of differential expression across disparate treatments, cell lines,
experiments, etc.


The available options for PCA are:


    

  • Center data?: Whether or not the input data should be mean-centered prior to performing PCA
    • 

  • Scale data?: Whether or not the input data should be scaled prior to performing PCA
    • 

  • Use complete cases?: Should only features present across all contrasts be used in
the PCA computation? If complete cases is not selected then default values
are imputed for each data type (logFC=0, P.Value=1, z-statistic=0, t-statistic=0).
    • 

  • Number of components: How many components should be calculated when performing PCA
    • 

  • Remove proportion low variance features: Removes this proportion of features
in the selected data by unsupervised variance across all contrasts. If left blank
then only features with constant variance are removed prior to downstream
PCA computation.
    • 

  • PCA algorithm: A BiocSingularParam used to specify the algorithm to use for
SVD.
    • 



After performing PCA, a biplot of the results is displayed to the user. Clicking
the gear icon at the top left allows the user to select which components to
use on each axis as well as options for coloring data points by metadata
information. Users can use the lasso selector on the plot to select specific
data points. Metadata pertaining to the selected points is displayed below the
plot.


3. UMAP


Like PCA, UMAP is performed on data from the selected contrasts. The user can opt
to perform UMAP on either the genes alone, TEs alone, or both. The input data values
for UMAP are one of the values computed from the differential expression analysis
for a particular feature, e.g. logFC, t-statistic, P-value, or z-statistic.


The available options for UMAP are:


    

  • Center data?: Whether or not the input data should be mean-centered prior to performing UMAP
    • 

  • Scale data?: Whether or not the input data should be scaled prior to performing UMAP
    • 

  • Use complete cases?: Should only features present across all contrasts be used in
the UMAP computation? If complete cases is not selected then default values
are imputed for each data type (logFC=0, P.Value=1, z-statistic=0, t-statistic=0).
    • 

  • Remove proportion low variance features: Removes this proportion of features
in the selected data by unsupervised variance across all contrasts. If left blank
then only features with constant variance are removed prior to downstream
UMAP computation.
    • 

  • N neighbors: Number of nearest neighbors
    • 

  • Minimum distance: Determines how close points are in the final layout
    • 

  • Iterations: the number of iterations used during layout optimization
    • 

  • Metric: Determines how distances between points are computed
    • 



After performing UMAP, a scatter plot of the computed embeddings is displayed
to the user. Clicking the gear icon at the top left allows the user to select
options for coloring data points by metadata information. Users can use the
lasso selector on the plot to select specific data points. Metadata pertaining
to the selected points is displayed below the plot.


4. Meta-Combine


The meta-combine tab allows for the exploration of global patterns of
dysregulation across all selected contrasts by implementing p-value combination
methods on the differential expression results via the metapod R package
(see package vignette for more details). Users can select one of the following p-value combination techniques to be performed on the raw p-values and logFC values from each contrast:


The following yield a significant result where any individual tests are significant


    

  • Fisher: Most sensitive to smallest p-value
    • 

  • Pearson: Most sensitive to largest p-value
    • 

  • Stouffer: Compromise between Fisher’s and Pearson’s methods
    • 

  • Simes: More conservative than the above but functional in the presence of dependencies
    • 



The following yield significant results where some of the individual tests are
significant (a minimum number of tests)


    

  • Wilkinson: Requires independence. User most supply minimum proportion of
tests or minimum number.
    • 

  • Holm: Does not require independence. User most supply minimum proportion of
tests or minimum number.
    • 

  • Berger: Will yield a significant result for groups where all of the
individual tests are significant
    • 



After calculating the meta-differential expression results a meta-volcano plot is
displayed. The meta-volcano can be used to examine the representative logFC vs
the combined p-values for all features of the selected contrasts. Users can use the
lasso selector on the plot to select specific data points to be displayed in the
results table below.


5. Ranked Expression


The ranked expression tab simply allows users to examine a table of all selected
contrasts ranked by the number of differentially expressed features (genes, TEs,
or both) per contrast. This table is useful for quickly finding and comparing
those contrasts where expression is most dysregulated.


Differential Expression


The differential expression tab enables the exploration of individual
experiment-level DE results as volcano or MA plots. The user can select features
(genes, TEs, or both) and create volcano plots of the differential expression
results for any contrast. Users can adjust the FDR cutoff and logFC thresholds
used in the plots. A table of differential expression results for the selected
contrast is displayed below the plots.


GSEA


The GSEA tab enables interactive gene-set enrichment analysis for any selected
contrast. All MSigDB gene sets are available for analysis. GSEA is performed
using the fgsea::fgsea() function. After performing GSEA, an enrichment plot
is displayed for the most significant result by default. Users can use the gear
icon in the top left of the plot to create an enrichment plot for any available
contrast. A results table is displayed below the plot.


NOTE: GSEA is performed only on the genes from the differential expression
analysis. The ranking stat used for GSEA is the z-statistic which is computed
from the moderated t-statistic of the differential expression results.


Over-representation


The over-representation tab enables interactive over-representation analysis
between sets of selected features using the clusterProfiler::enrichGO()
function for the selected contrast. Users can select the ontology set to
perform GO analysis on and an FDR cutoff for defining significant genes in the
study. After performing GO analysis, a dotplot and network graph of the
significantly enriched terms is displayed for each gene set
(up or down-regulated).


NOTE: GO is performed only on the genes from the differential expression
analysis. the background gene set used is comprised of all genes observed in
that particular study.