library(FeatSeekR)
library(DmelSGI)
library(pheatmap)
library(SummarizedExperiment)

1 Introduction

A fundamental step in many analyses of high-dimensional data is dimension reduction. Feature selection is one approach to dimension reduction whose strengths include interpretability, conceptual simplicity, transferability and modularity. Here, we introduce the FeatSeekR algorithm, which selects features based on the consistency of their signal across replicates and their non-redundancy. It takes a 2 dimensional array (features x samples) of replicated measurements and returns a SummarizedExperiment object storing the selected features ranked by reproducibility.

2 Installation

if (!requireNamespace("BiocManager", quietly = TRUE))
    install.packages("BiocManager")

BiocManager::install("FeatSeekR")

3 Feature selection on simulated data

Here we simulate a data set with features generated by orthogonal latent factors. Features derived from the same latent factor are highly redundant and form distinct clusters. The function simulates 10 redundant features per latent factor. Replicates are generated by adding independent Gaussian noise.

set.seed(111)
# simulate data with 500 conditions, 3 replicates and 5 latent factors 
conditions <- 500
latent_factors <- 5
replicates <- 3

# simData generates 10 features per latent_factor, so choosing latent_factors=5
# will generate 50 features.
# we simulate samples from 500 independent conditions per replicate. setting 
# conditions=500 and replicates=3 will generate 1500 samples, leading to 
# final data dimensions of 50 features x 1500 samples
sim <- simData(conditions=conditions, n_latent_factors=latent_factors,
                replicates=replicates)

# show that simulated data dimensions are indeed 50 x 1500 
dim(assay(sim, "data"))
## [1]   50 1500
# calculate the feature correlation for first replicate
data <- t(assay(sim, "data"))
cor <- cor(data, use = "pairwise.complete.obs")

# plot a heatmap of the features and color features according to their 
# generating latent factors
anno <- data.frame(Latent_factor = as.factor(rep(1:5, each=10)))
rownames(anno) <- dimnames(sim)[[1]]
colors        <- c("red", "blue", "darkorange", "darkgreen", "black")
names(colors) <- c("1", "2", "3", "4", "5")
anno_colors <- list(Latent_factor = colors)
range <- max(abs(cor))
pheatmap(cor, treeheight_row = 0 , treeheight_col = 0, 
        show_rownames = FALSE, show_colnames = FALSE,
        breaks = seq(-range, range, length.out = 100), cellwidth = 6, 
        cellheight = 6, annotation_col = anno, annotation_colors = anno_colors, 
        fontsize = 8)

We first plot the correlation matrix of the data to visualize feature redundancy. As intended by the simulation, the features derived from the same latent factor cluster together. This suggests that the true dimension is indeed lower than the number of features. We now run FeatSeekR to rank the features based on their uniqueness and reproducibility.

# select the top 5 features
res <- FeatSeek(sim, max_features=5)
## Input data has: 
## 1500 samples 
## 500 conditions 
## 50 features 
## 3 replicates
## Starting feature ranking!
## Iteration: 1 selected = Latent_factor2_Feature3, replicate F-statistic = 10873.0023316711
## Iteration: 2 selected = Latent_factor5_Feature6, replicate F-statistic = 9364.10884704126
## Iteration: 3 selected = Latent_factor3_Feature6, replicate F-statistic = 6048.5339756004
## Iteration: 4 selected = Latent_factor4_Feature1, replicate F-statistic = 1777.09431418772
## Iteration: 5 selected = Latent_factor1_Feature5, replicate F-statistic = 60.3857000568657
## Finished feature ranking procedure!
## Calculating explained variance of selected feature sets!
# plot a heatmap of the top 5 selected features 
plotSelectedFeatures(res)

We again visualize the selected features by plotting their correlation matrix. As expected, the top 5 selected features are each from a different latent factor and low correlated. This suggests that we were able to obtain a compressed version of the data, while keeping most of the contained information.

4 Selecting image features from the DmelSGI package

Here we use FeatSeekR to rapidly identify unique features with reproducible signal between measurements in an image dataset from the DmelSGI package. The authors of DmelSGI performed combinatorial gene knock-outs using siRNA, followed by imaging of the cells. The resulting images were segmented and features were extracted using the EBImage package. Here, conditions refer to different gene knock-outs, features to the extracted image features and replicates to repeated measurements of the individual conditions.

# load data from DmelSGI package
data("subSampleForStabilitySelection", package="DmelSGI")
data <- subSampleForStabilitySelection$D
# dimensions are conditions, features, replicates
data <- aperm(data, c(1, 3, 2))
# set feature names
dimnames(data)[[2]] <- subSampleForStabilitySelection$phenotype
# bind samples and create condition factor
conds <- rep(seq_len(dim(data)[1]), 2)
data<- rbind(data[, , 1], data[, , 2])
# show final data dimensions
dim(data)
## [1] 6000  162

The input data has 3000 samples, 162 features and 2 replicates. Again, we plot the correlation matrix of the data to explore the structure of the features.

# calculate correlation matrix of the first 50 features of one of the replicates
cor_mat <- cor(data[, 1:50, drop=FALSE])

# plot correlation matrix, omitting featurenames
pheatmap(cor_mat, show_rownames=FALSE, show_colnames=FALSE,
    treeheight_row=0, treeheight_col=0)

Analogous to the idealized simulated example, the extracted features formed groups of high correlation within and lower correlation between. This supports the idea that the effective dimension of the data matrix is substantially lower than the number of features and that feature selection is a plausible approach to these data. We apply FeatSeek to identify unique features with high replicate consistency.

# run FeatSeekR and rank up to 20 features based on their replicate 
# reproducibility and uniqueness
max_features <- 30
res <- FeatSeek(t(data), 
        conditions=conds, 
        max_features=max_features,
        verbose=TRUE)
## Input data has: 
## 6000 samples 
## 3000 conditions 
## 162 features 
## 2 replicates
## Starting feature ranking!
## Iteration: 1 selected = 4x.isMitoticAll, replicate F-statistic = 39.9436313137099
## Iteration: 2 selected = 4x.areaNucAll, replicate F-statistic = 30.0263582106047
## Iteration: 3 selected = 4x.ratioMitoticAll, replicate F-statistic = 11.5398616989651
## Iteration: 4 selected = 4x.LCD4, replicate F-statistic = 10.8825393255875
## Iteration: 5 selected = 4x.areaNucH1, replicate F-statistic = 7.95956181567868
## Iteration: 6 selected = 4x.areaNucH10, replicate F-statistic = 7.47519394820442
## Iteration: 7 selected = 4x.LCDratio.LCD4, replicate F-statistic = 5.8357355692092
## Iteration: 8 selected = 4x.areaNucH3, replicate F-statistic = 5.55948447657088
## Iteration: 9 selected = 4x.areapH3All, replicate F-statistic = 5.3771381704978
## Iteration: 10 selected = 4x.LCD3, replicate F-statistic = 4.33930418841254
## Iteration: 11 selected = 4x.intNucH10, replicate F-statistic = 3.72653395075801
## Iteration: 12 selected = 4x.intNucQ0.97, replicate F-statistic = 3.56658985821053
## Iteration: 13 selected = 4x.LCD8, replicate F-statistic = 3.55024557636757
## Iteration: 14 selected = 4x.areaNucH7, replicate F-statistic = 3.35197326379491
## Iteration: 15 selected = 10x.meanNonmitotic.cell.Tub.m.eccentricity, replicate F-statistic = 3.20238723367772
## Iteration: 16 selected = 4x.intNucH5, replicate F-statistic = 3.1658168982331
## Iteration: 17 selected = 4x.intH3pH4, replicate F-statistic = 3.05967610288084
## Iteration: 18 selected = 10x.meanNonmitotic.nucleus.DAPI.m.majoraxis, replicate F-statistic = 2.83259945427357
## Iteration: 19 selected = 4x.intNucH1, replicate F-statistic = 2.82822480334084
## Iteration: 20 selected = 4x.intNucQ0.10, replicate F-statistic = 2.755206092354
## Iteration: 21 selected = 4x.areaNucQ0.90, replicate F-statistic = 2.53605767224537
## Iteration: 22 selected = 4x.intNucH8, replicate F-statistic = 2.50090551118911
## Iteration: 23 selected = 4x.LCD2, replicate F-statistic = 2.35829643448643
## Iteration: 24 selected = 4x.intpH3All, replicate F-statistic = 2.34578890414463
## Iteration: 25 selected = 4x.intNucH3, replicate F-statistic = 2.22292998007601
## Iteration: 26 selected = 10x.meanNonmitotic.nucleus.DAPI.m.eccentricity, replicate F-statistic = 2.17920581314792
## Iteration: 27 selected = 4x.areaNucH5, replicate F-statistic = 2.14959536975668
## Iteration: 28 selected = 10x.meanNonmitotic.cell.0.s.perimeter, replicate F-statistic = 2.14253291864597
## Iteration: 29 selected = 10x.meanNonmitotic.cell.Tub.m.majoraxis, replicate F-statistic = 2.34896602854229
## Iteration: 30 selected = 4x.intNucQ0.25, replicate F-statistic = 2.10291760262207
## Finished feature ranking procedure!
## Calculating explained variance of selected feature sets!

In order determine the ideal number of selected features we can have a look at the fraction of explained variance per additionally selected feature.

# plotVarianceExplained plots the fraction of explained variance per 
# additionally selected feature, ranked by FeatSeek.
plotVarianceExplained(res)

The increase in explained variance seems to flatten out at around 70%. We therefore select the number of features, that explain at least 70% of the total variance and plot their correlation matrix.

# get number of features which explain at least 70% of the total variance
n_feat <- min(which(rowData(res)$explained_variance > 0.7))

# plot the top n_feat features based on the ranking by FeatSeek
plotSelectedFeatures(res, n_features=n_feat)

The low correlation between the top selected features confirm their low redundancy. Using FeatSeekR we were able to reduce the dimension of the data to 17 features, while still being able to explain 70% of the variance of the original data.

5 Session Info

sessionInfo()
## R version 4.3.1 (2023-06-16)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 22.04.3 LTS
## 
## Matrix products: default
## BLAS:   /home/biocbuild/bbs-3.18-bioc/R/lib/libRblas.so 
## LAPACK: /usr/lib/x86_64-linux-gnu/lapack/liblapack.so.3.10.0
## 
## locale:
##  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
##  [3] LC_TIME=en_GB              LC_COLLATE=C              
##  [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
##  [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                 
##  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       
## 
## time zone: America/New_York
## tzcode source: system (glibc)
## 
## attached base packages:
## [1] stats4    stats     graphics  grDevices utils     datasets  methods  
## [8] base     
## 
## other attached packages:
##  [1] SummarizedExperiment_1.32.0 Biobase_2.62.0             
##  [3] GenomicRanges_1.54.0        GenomeInfoDb_1.38.0        
##  [5] IRanges_2.36.0              S4Vectors_0.40.0           
##  [7] BiocGenerics_0.48.0         MatrixGenerics_1.14.0      
##  [9] matrixStats_1.0.0           pheatmap_1.0.12            
## [11] DmelSGI_1.33.0              FeatSeekR_1.2.0            
## [13] BiocStyle_2.30.0           
## 
## loaded via a namespace (and not attached):
##  [1] gtable_0.3.4            xfun_0.40               bslib_0.5.1            
##  [4] caTools_1.18.2          rhdf5_2.46.0            lattice_0.22-5         
##  [7] rhdf5filters_1.14.0     tools_4.3.1             bitops_1.0-7           
## [10] pkgconfig_2.0.3         Matrix_1.6-1.1          KernSmooth_2.23-22     
## [13] RColorBrewer_1.1-3      lifecycle_1.0.3         GenomeInfoDbData_1.2.11
## [16] farver_2.1.1            compiler_4.3.1          gplots_3.1.3           
## [19] statmod_1.5.0           munsell_0.5.0           codetools_0.2-19       
## [22] htmltools_0.5.6.1       sass_0.4.7              RCurl_1.98-1.12        
## [25] yaml_2.3.7              pracma_2.4.2            crayon_1.5.2           
## [28] jquerylib_0.1.4         MASS_7.3-60             DelayedArray_0.28.0    
## [31] cachem_1.0.8            limma_3.58.0            magick_2.8.1           
## [34] iterators_1.0.14        TSP_1.2-4               abind_1.4-5            
## [37] foreach_1.5.2           gtools_3.9.4            digest_0.6.33          
## [40] bookdown_0.36           fastmap_1.1.1           grid_4.3.1             
## [43] colorspace_2.1-0        cli_3.6.1               SparseArray_1.2.0      
## [46] magrittr_2.0.3          S4Arrays_1.2.0          scales_1.2.1           
## [49] rmarkdown_2.25          XVector_0.42.0          igraph_1.5.1           
## [52] evaluate_0.22           knitr_1.44              rlang_1.1.1            
## [55] Rcpp_1.0.11             glue_1.6.2              BiocManager_1.30.22    
## [58] jsonlite_1.8.7          R6_2.5.1                Rhdf5lib_1.24.0        
## [61] zlibbioc_1.48.0