Data Integration using Unsupervised Multiple Kernel Learning

Introduction

The TARA Oceans expedition facilitated the study of plankton communities by providing ocean metagenomic data combined with environmental measures to the scientific community. This study focuses on 139 prokaryotic-enriched samples collected from 68 stations and spread across three depth layers: the surface (SRF), the deep chlorophyll maximum (DCM) layer and the mesopelagic (MES) zones. Samples were located in 8 different oceans or seas: Indian Ocean (IO), Mediterranean Sea (MS), North Atlantic Ocean (NAO), North Pacific Ocean (NPO), Red Sea (RS), South Atlantic Ocean (SAO), South Pacific Ocean (SPO) and South Ocean (SO).

In this vignette, we consider a subset of the original data analyzed in the article (Mariette & Villa-Vialaneix, 2018) and illustrate the usefulness of mixKernel to 1/ perform an integrative exploratory analysis as in (Mariette & Villa-Vialaneix, 2018) and to 2/ select relevant variables for unsupervised analysis.

The data include 1% of the 35,650 prokaryotic OTUs and of the 39,246 bacterial genes that were randomly selected. The aim is to integrate prokaryotic abundances and functional processes to environmental measure with an unsupervised method.

Install and load the mixOmics and mixKernel packages:

## required python modules: autograd, numpy, scipy, sklearn
## To properly install packages, run:
# install.packages("BiocManager")
# BiocManager::install("mixOmics")
# BiocManager::install("phyloseq")
# install.packages("mixKernel")
library(mixOmics)
library(mixKernel)

Loading TARA Ocean datasets

The (previously normalized) datasets are provided as matrices with matching sample names (rownames):

data(TARAoceans)
# more details with: ?TARAOceans
# we check the dimension of the data:
lapply(list("phychem" = TARAoceans$phychem, "pro.phylo" = TARAoceans$pro.phylo, 
            "pro.NOGs" = TARAoceans$pro.NOGs), dim)

## $phychem
## [1] 139  22
## 
## $pro.phylo
## [1] 139 356
## 
## $pro.NOGs
## [1] 139 638

Multiple kernel computation

Individual kernel computation

For each input dataset, a kernel is computed using the function compute.kernel with the choice of linear, phylogenic or abundance kernels. A user defined function can also be provided as input(argument kernel.func, see ?compute.kernel).

The results are lists with a ‘kernel’ entry that stores the kernel matrix. The resulting kernels are symmetric matrices with a size equal to the number of observations (rows) in the input datasets.

phychem.kernel <- compute.kernel(TARAoceans$phychem, kernel.func = "linear")
pro.phylo.kernel <- compute.kernel(TARAoceans$pro.phylo, kernel.func = "abundance")
pro.NOGs.kernel <- compute.kernel(TARAoceans$pro.NOGs, kernel.func = "abundance")

# check dimensions
dim(pro.NOGs.kernel$kernel)

## [1] 139 139

A general overview of the correlation structure between datasets is obtained as described in Mariette and Villa-Vialaneix (2018) and displayed using the function cim.kernel:

cim.kernel(phychem = phychem.kernel,
           pro.phylo = pro.phylo.kernel,
           pro.NOGs = pro.NOGs.kernel, 
           method = "square")

The figure shows that pro.phylo and pro.NOGs is the most correlated pair of kernels. This result is expected as both kernels provide a summary of prokaryotic communities.

Combined kernel computation

The function combine.kernels implements 3 different methods for combining kernels: STATIS-UMKL, sparse-UMKL and full-UMKL (see more details in Mariette and Villa-Vialaneix, 2018). It returns a meta-kernel that can be used as an input for the function kernel.pca (kernel PCA). The three methods bring complementary information and must be chosen according to the research question.

The STATIS-UMKL approach gives an overview on the common information between the different datasets. The full-UMKL computes a kernel that minimizes the distortion between all input kernels. The sparse-UMKL is a sparse variant of full-UMKL that selects the most relevant kernels in addition to distortion minimization.

meta.kernel <- combine.kernels(phychem = phychem.kernel,
                               pro.phylo = pro.phylo.kernel,
                               pro.NOGs = pro.NOGs.kernel, 
                               method = "full-UMKL")

Exploratory analysis: Kernel Principal Component Analysis (KPCA)

Perform KPCA

A kernel PCA can be performed from the combined kernel with the function kernel.pca``. The argumentncomp` allows to choose how many components to extract from KPCA.

kernel.pca.result <- kernel.pca(meta.kernel, ncomp = 10)

Sample plots using the plotIndiv function from mixOmics:

all.depths <- levels(factor(TARAoceans$sample$depth))
depth.pch <- c(20, 17, 4, 3)[match(TARAoceans$sample$depth, all.depths)]
plotIndiv(kernel.pca.result,
          comp = c(1, 2),
          ind.names = FALSE,
          legend = TRUE,
          group = as.vector(TARAoceans$sample$ocean),
          col.per.group = c("#f99943", "#44a7c4", "#05b052", "#2f6395", 
                            "#bb5352", "#87c242", "#07080a", "#92bbdb"),
          pch = depth.pch,
          pch.levels = TARAoceans$sample$depth,
          legend.title = "Ocean / Sea",
          title = "Projection of TARA Oceans stations",
          size.title = 10,
          legend.title.pch = "Depth")

The explained variance supported by each axis of KPCA is displayed with the plot function, and can help choosing the number of components in KPCA.

plot(kernel.pca.result)

The first axis summarizes ~ 20% of the total variance.

Assessing important variables

Here we focus on the information summarized on the first component. Variable values are randomly permuted with the function permute.kernel.pca.

In the following example, physical variable are permuted at the variable level (kernel phychem), OTU abundances from pro.phylo kernel are permuted at the phylum level (OTU phyla are stored in the second column, named Phylum, of the taxonomy annotation provided in TARAoceans object in the entry taxonomy) and gene abundances from pro.NOGs are permuted at the GO level (GO are provided in the entry GO of the dataset):

head(TARAoceans$taxonomy[ ,"Phylum"], 10)

##  [1] Proteobacteria Proteobacteria Proteobacteria Proteobacteria Proteobacteria
##  [6] Cyanobacteria  Proteobacteria Proteobacteria Chloroflexi    Proteobacteria
## 56 Levels: Acidobacteria Actinobacteria aquifer1 Aquificae ... WCHB1-60

head(TARAoceans$GO, 10)

##  [1] NA  NA  "K" NA  NA  "S" "S" "S" NA  "S"

# here we set a seed for reproducible results with this tutorial
set.seed(17051753)
kernel.pca.result <- kernel.pca.permute(kernel.pca.result, ncomp = 1,
                                        phychem = colnames(TARAoceans$phychem),
                                        pro.phylo = TARAoceans$taxonomy[, "Phylum"],
                                        pro.NOGs = TARAoceans$GO)

Results are displayed with the function plotVar.kernel.pca. The argument ndisplay indicates the number of variables to display for each kernel:

plotVar.kernel.pca(kernel.pca.result, ndisplay = 10, ncol = 3)

Proteobacteria is the most important variable for the pro.phylo kernel.

The relative abundance of `Proteobacteria`` is then extracted in each of our 139 samples, and each sample is colored according to the value of this variable in the KPCA projection plot:

selected <- which(TARAoceans$taxonomy[, "Phylum"] == "Proteobacteria")
proteobacteria.per.sample <- apply(TARAoceans$pro.phylo[, selected], 1, sum) /
  apply(TARAoceans$pro.phylo, 1, sum)
colfunc <- colorRampPalette(c("royalblue", "red"))
col.proteo <- colfunc(length(proteobacteria.per.sample))
col.proteo <- col.proteo[rank(proteobacteria.per.sample, ties = "first")]
plotIndiv(kernel.pca.result,
          comp = c(1, 2),
          ind.names = FALSE,
          legend = FALSE,
          col = col.proteo,
          pch = depth.pch,
          pch.levels = TARAoceans$sample$depth,
          legend.title = "Ocean / Sea",
          title = "Representation of Proteobacteria abundance",
          legend.title.pch = "Depth")

Similarly, the temperature is the most important variable for the phychem kernel. The temperature values can be displayed on the kernel PCA projection as follows:

col.temp <- colfunc(length(TARAoceans$phychem[, 4]))
col.temp <- col.temp[rank(TARAoceans$phychem[, 4], ties = "first")]
plotIndiv(kernel.pca.result,
          comp = c(1, 2),
          ind.names = FALSE,
          legend = FALSE,
          col = col.temp,
          pch = depth.pch,
          pch.levels = TARAoceans$sample$depth,
          legend.title = "Ocean / Sea",
          title = "Representation of mean temperature",
          legend.title.pch = "Depth")

Selecting relevant variables

Here, we use a feature selection approach that does not rely on any assumption but explicitly takes advantage of the kernel structure in an unsupervised fashion. The idea is to preserve at best the similarity structure between samples. These examples requires the installation of the python modules autograd, scipy, numpy, and sklearn. See detailed instructions in the installation vignette or on mixKernel website : http://mixkernel.clementine.wf

have_depend <- reticulate::py_module_available("autograd") &
  reticulate::py_module_available("scipy") &
  reticulate::py_module_available("numpy") &
  reticulate::py_module_available("sklearn") 

if (have_depend) {
  ukfs.res <- select.features(TARAoceans$pro.phylo, kx.func = "bray", lambda = 1, 
                              keepX = 5, nstep = 1)
  selected <- sort(ukfs.res, decreasing = TRUE, index.return = TRUE)$ix[1:5]
  TARAoceans$taxonomy[selected, ]
}

The select.features function allows to add a structure constraint to the variable selection. The adjacency matrix of the graph representing relations between OTUs can be obtained by computing the Pearson correlation matrix as follows:

library("MASS")
library("igraph")
library("correlationtree")

pro.phylo.alist <- data.frame("names" = colnames(TARAoceans$pro.phylo), 
                              t(TARAoceans$pro.phylo))
L <- mat2list(df2mat(pro.phylo.alist, 1))
corr.mat <- as.matrix(cross_cor(L, remove = TRUE))
pro.phylo.graph <- graph_from_adjacency_matrix(corr.mat, 
                                               mode = "undirected",
                                               weighted = TRUE)
Lg <- laplacian_matrix(pro.phylo.graph, sparse=TRUE)

if (have_depend) {
  load(file = file.path(system.file(package = "mixKernel"), "loaddata", "Lg.rda"))
  ukfsg.res <- select.features(TARAoceans$pro.phylo, kx.func = "bray", 
                               lambda = 1, method = "graph", Lg = Lg, keepX = 5,
                               nstep = 1)
  
  selected <- sort(ukfsg.res, decreasing = TRUE, index.return = TRUE)$ix[1:5]
  TARAoceans$taxonomy[selected, ]
}

References

1. Mariette, J. and Villa-Vialaneix, N. (2018). Unsupervised multiple kernel learning for heterogeneous data integration. Bioinformatics, 34(6), 1009-1015.

2. Zhuang, J., Wang, J., Hoi, S., and Lan, X. (2011). Unsupervised multiple kernel clustering. Journal of Machine Learning Research (Workshop and Conference Proceedings), 20, 129–144.

3. Lavit, C., Escoufier, Y., Sabatier, R., and Traissac, P. (1994). The act (statis method). Computational Statistics & Data Analysis, 18(1), 97–119.

Session information

sessionInfo()

## R version 4.3.2 (2023-10-31)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 22.04.3 LTS
## 
## Matrix products: default
## BLAS:   /usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3 
## LAPACK: /usr/lib/x86_64-linux-gnu/openblas-pthread/libopenblasp-r0.3.20.so;  LAPACK version 3.10.0
## 
## locale:
##  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
##  [3] LC_TIME=fr_FR.UTF-8        LC_COLLATE=C              
##  [5] LC_MONETARY=fr_FR.UTF-8    LC_MESSAGES=en_US.UTF-8   
##  [7] LC_PAPER=fr_FR.UTF-8       LC_NAME=C                 
##  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
## [11] LC_MEASUREMENT=fr_FR.UTF-8 LC_IDENTIFICATION=C       
## 
## time zone: Europe/Paris
## tzcode source: system (glibc)
## 
## attached base packages:
## [1] stats     graphics  grDevices utils     datasets  methods   base     
## 
## other attached packages:
## [1] mixKernel_0.9-1   reticulate_1.34.0 mixOmics_6.26.0   ggplot2_3.4.4    
## [5] lattice_0.22-5    MASS_7.3-60.0.1  
## 
## loaded via a namespace (and not attached):
##   [1] mnormt_2.1.1            bitops_1.0-7            gridExtra_2.3          
##   [4] permute_0.9-7           rlang_1.1.3             magrittr_2.0.3         
##   [7] ade4_1.7-22             matrixStats_1.2.0       compiler_4.3.2         
##  [10] mgcv_1.9-1              png_0.1-8               vctrs_0.6.5            
##  [13] reshape2_1.4.4          quadprog_1.5-8          stringr_1.5.1          
##  [16] pkgconfig_2.0.3         crayon_1.5.2            fastmap_1.1.1          
##  [19] XVector_0.42.0          labeling_0.4.3          utf8_1.2.4             
##  [22] rmarkdown_2.25          markdown_1.12           purrr_1.0.2            
##  [25] xfun_0.41               zlibbioc_1.48.0         cachem_1.0.8           
##  [28] GenomeInfoDb_1.38.5     jsonlite_1.8.8          biomformat_1.30.0      
##  [31] highr_0.10              rhdf5filters_1.14.1     Rhdf5lib_1.24.1        
##  [34] BiocParallel_1.36.0     psych_2.4.1             parallel_4.3.2         
##  [37] cluster_2.1.6           R6_2.5.1                bslib_0.6.1            
##  [40] stringi_1.8.3           RColorBrewer_1.1-3      jquerylib_0.1.4        
##  [43] Rcpp_1.0.12             iterators_1.0.14        knitr_1.45             
##  [46] IRanges_2.36.0          Matrix_1.6-5            splines_4.3.2          
##  [49] igraph_1.6.0            tidyselect_1.2.0        rstudioapi_0.15.0      
##  [52] yaml_2.3.8              vegan_2.6-4             codetools_0.2-19       
##  [55] tibble_3.2.1            plyr_1.8.9              Biobase_2.62.0         
##  [58] withr_3.0.0             rARPACK_0.11-0          evaluate_0.23          
##  [61] survival_3.5-7          Biostrings_2.70.1       pillar_1.9.0           
##  [64] phyloseq_1.46.0         corrplot_0.92           foreach_1.5.2          
##  [67] stats4_4.3.2            ellipse_0.5.0           generics_0.1.3         
##  [70] rprojroot_2.0.4         RCurl_1.98-1.14         S4Vectors_0.40.2       
##  [73] munsell_0.5.0           scales_1.3.0            glue_1.7.0             
##  [76] tools_4.3.2             data.table_1.14.10      RSpectra_0.16-1        
##  [79] rhdf5_2.46.1            grid_4.3.2              tidyr_1.3.0            
##  [82] ape_5.7-1               colorspace_2.1-0        nlme_3.1-164           
##  [85] GenomeInfoDbData_1.2.11 cli_3.6.2               rappdirs_0.3.3         
##  [88] fansi_1.0.6             dplyr_1.1.4             corpcor_1.6.10         
##  [91] gtable_0.3.4            sass_0.4.8              digest_0.6.34          
##  [94] BiocGenerics_0.48.1     ggrepel_0.9.5           LDRTools_0.2-2         
##  [97] farver_2.1.1            htmltools_0.5.7         multtest_2.58.0        
## [100] lifecycle_1.0.4         here_1.0.1