BioMM: Biological-informed Multi-stage Machine learning framework for phenotype prediction using omics data
Modified: 17 July 2019. Compiled: 29 Oct 2019
Contents
1 Overview
1.1 Motivation
The identification of reproducible biological patterns from high-dimensional omics data is a key factor in understanding the biology of complex disease or traits. Incorporating prior biological knowledge into machine learning is an important step in advancing such research.
1.2 Deliverables
We have implemented a biologically informed multi-stage machine learing framework termed BioMM [1] specifically for phenotype prediction using omics-scale data based on biological prior information, for example, gene ontological pathways.
Features of BioMM in a nutshell:
- Applicability for various omics data modalities.
- Prioritizing outcome-associated functional patterns.
- End-to-end prediction at the individual level based on biological
pathway stratification.
- Possibility for extension to machine learning models of interest.
- Parallel computing.
2 Getting started
2.1 Installation
- Development version from Github (R 3.5):
install.packages("devtools")
library("devtools")
install_github("transbioZI/BioMM")- Install BioMM from Bioconductor (R 3.6):
if (!requireNamespace("BiocManager", quietly = TRUE))
install.packages("BiocManager")
# The following initializes usage of Bioc devel
BiocManager::install(version='devel')
BiocManager::install("BioMM")Note that, two different parallel computing strategies are implemented: For BioMM package available on Github, the parallel R package is implemented. The BiocParallel R package is employed in Bioconductor BioMM package which allows the use of parallel computation across any system.
- Load required libraries
library(BioMM)
library(BiocParallel)
library(parallel)
library(ranger)
library(rms)
library(glmnet)
library(e1071)
library(pROC)
library(vioplot)
library(variancePartition)
library(CMplot)3 Omics data
A wide range of genome-wide omics data is supported for the use of BioMM
including whole-genome DNA methylation, gene expression and genome-wide SNP
data. Other types of omics data that can map into pathways are also encouraging.
For better understanding of the framework, we used a preprocessed genome-wide
DNA methylation data with 26486 CpGs and 40 samples consisting of 20 controls
and 20 patients. (0: healthy control and 1: patient) for demonstration.
## Get DNA methylation data
studyData <- readRDS(system.file("extdata", "/methylData.rds",
package="BioMM"))
head(studyData[,1:5])## label cg00000292 cg00002426 cg00003994 cg00005847
## GSM951223 1 0.0274 0.0029 -0.0027 -0.0196
## GSM951231 1 0.0644 0.0181 -0.0088 0.0057
## GSM951249 1 -0.0304 -0.0013 -0.0083 -0.0116
## GSM951273 1 0.0252 0.0039 -0.0091 0.0030
## GSM951214 1 0.0289 -0.0011 -0.0129 0.0173
## GSM951270 1 0.0635 0.0329 0.0184 -0.0023dim(studyData)## [1] 40 264874 Feature stratification
Features like CpGs, genes or SNPs can be mapped into pathways based on genomic location and pathway annotation, as implemented in the function omics2pathlist(). The examples of pathway databases are gene ontology (GO), Reactome and KEGG, which are widely used public pathway repositories. Gene ontological pathway is used in this tutorial.
## Load annotation data
featureAnno <- readRDS(system.file("extdata", "cpgAnno.rds", package="BioMM"))
pathlistDB <- readRDS(system.file("extdata", "goDB.rds", package="BioMM"))
head(featureAnno)## ID chr entrezID symbol
## 1 cg00000292 16 487 ATP2A1
## 2 cg00002426 3 7871 SLMAP
## 3 cg00003994 7 4223 MEOX2
## 4 cg00005847 2 3232 HOXD3
## 5 cg00006414 7 57541 ZNF398
## 6 cg00007981 11 24145 PANX1str(pathlistDB[1:3])## List of 3
## $ GO:0000002: Named chr [1:12] "291" "1890" "4205" "4358" ...
## ..- attr(*, "names")= chr [1:12] "TAS" "IMP" "ISS" "IMP" ...
## $ GO:0000012: Named chr [1:11] "3981" "7141" "7515" "23411" ...
## ..- attr(*, "names")= chr [1:11] "IDA" "IDA" "IEA" "IMP" ...
## $ GO:0000027: Named chr [1:31] "4839" "6122" "6123" "6125" ...
## ..- attr(*, "names")= chr [1:31] "IMP" "IBA" "IBA" "IBA" ...## Map to pathways (only 100 pathways in order to reduce the runtime)
pathlistDBsub <- pathlistDB[1:100]
pathlist <- omics2pathlist(data=studyData, pathlistDBsub, featureAnno,
restrictUp=100, restrictDown=20, minPathSize=10) ## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 12.00 38.00 50.50 60.84 74.25 164.005 BioMM framework
5.1 Recapitulation
Briefly, the BioMM framework consists of two learning stages [1]. During the first stage, biological meta-information is used to ‘compress’ the variables of the original dataset into pathway-level ‘latent variables’ (henceforth called stage-2 data) using either supervised or unsupervised learning models. In the second stage, a supervised model is built using the stage-2 data with non-negative outcome-associated features for prediction.
5.2 End-to-end prediction modules
5.2.1 Interface to machine learning models
The end-to-end prediction is performed using BioMM() function. Both
supervised and unsupervised learning are implemented in the BioMM framework,
which are indicated by the argument supervisedStage1=TRUE or
supervisedStage1=FALSE. Commonly used supervised classifiers: generalized
regression models with lasso, ridge or elastic net regularization (GLM) [4],
support vector machine (SVM) [3] and random forest [2] are included. For the
unsupervised method, regular or sparse constrained principal component
analysis (PCA) [5] is used. Generic resampling methods include
cross-validation (CV) and bootstrapping (BS) procedures as the argument
resample1="CV" or resample1="BS". Stage-2 data is reconstructed using
either resampling methods during machine learning prediction or independent
test set prediction if the argument testData is provided.
5.2.1.1 Example
To apply random forest model, we use the argument classifier=randForest
in BioMM() with the classification mode at both
stages. predMode indicate the prediction type, here we use
classification for binary outcome prediction. A set of model hyper-parameters
are supplied by the argument paramlist at both stages.
Pathway-based stratification is carried out in this example. We
focused on the autosomal region to limit the potential influence of sex on
machine learning due to the phenomenon of X chromosome inactivation or the
existence of an additional X chromosome in female samples. Therefore it’s
suggested to exclude sex chromosome in the user-supplied featureAnno input
file.
## Parameters
supervisedStage1=TRUE
classifier <- "randForest"
predMode <- "classification"
paramlist <- list(ntree=300, nthreads=10)
param1 <- MulticoreParam(workers = 1)
param2 <- MulticoreParam(workers = 10)
## If BioMM is installed from Github, please use the following assignment:
## param1 <- 1
## param2 <- 10
studyDataSub <- studyData[,1:2000] ## to reduce the runtime
set.seed(123)
result <- BioMM(trainData=studyDataSub, testData=NULL, pathlistDB, featureAnno,
restrictUp=100, restrictDown=10, minPathSize=10,
supervisedStage1, typePCA="regular",
resample1="BS", resample2="CV", dataMode="allTrain",
repeatA1=50, repeatA2=1, repeatB1=20, repeatB2=1,
nfolds=10, FSmethod1=NULL, FSmethod2=NULL,
cutP1=0.1, cutP2=0.1, fdr2=NULL, FScore=param1,
classifier, predMode,
paramlist, innerCore=param2)## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 10.00 10.00 11.00 11.98 13.00 22.00
##
## Levels of predicted Y = 2
##
## AUC ACC R2
## R2 0.9 0.9 0.695print(result)## AUC ACC R2
## R2 0.9 0.9 0.695Other machine learning models can be employed with the following respective
parameter settings. For the classifier "SVM", parameters can be tuned using
an internal cross validation if tuneP=TRUE. For generalized regression model
glmnet, elastic net is specified by the input argument alpha=0.5.
Alternatively, alpha=1 is for the lasso and alpha=0 is the ridge. For the
unsupervised learning supervisedStage1=FALSE, regular PCA
typePCA="regular" is applied and followed with random forest classification
classifier2=TRUE.
## SVM
supervisedStage1=TRUE
classifier <- "SVM"
predMode <- "classification"
paramlist <- list(tuneP=FALSE, kernel="radial",
gamma=10^(-3:-1), cost=10^(-3:1))
## GLM with elastic-net
supervisedStage1=TRUE
classifier <- "glmnet"
predMode <- "classification"
paramlist <- list(family="binomial", alpha=0.5,
ypeMeasure="mse", typePred="class")
## PCA + random forest
supervisedStage1=FALSE
classifier <- "randForest"
predMode <- "classification"
paramlist <- list(ntree=300, nthreads=10) 5.2.2 Interface to biological stratification strategies
For stratification of predictors using biological information, various
strategies can be applied. Currently, BioMM() integrates gene
ontological pathway based stratification, which not only accounts for epistasis
between first level variables within the functional category, but also considers
the interaction between pathways. Therefore, this may provide better information on functional insight.
5.2.2.1 Example
End-to-end prediction based on pathway-wide stratification on genome-wide DNA methylation data is demonstrated below. PCA is used at stage-1 to reconstruct pathway level data, then the random forest model with 10-fold cross validation is applied on stage-2 data to estimate the prediction performance.
## Parameters
supervisedStage1=FALSE
classifier <- "randForest"
predMode <- "classification"
paramlist <- list(ntree=300, nthreads=10)
param1 <- MulticoreParam(workers = 1)
param2 <- MulticoreParam(workers = 10)
## If BioMM is installed from Github, please use the following assignment:
## param1 <- 1
## param2 <- 10
set.seed(123)
result <- BioMM(trainData=studyData, testData=NULL,
pathlistDBsub, featureAnno,
restrictUp=100, restrictDown=10, minPathSize=10,
supervisedStage1, typePCA="regular",
resample1="BS", resample2="CV", dataMode="allTrain",
repeatA1=40, repeatA2=1, repeatB1=40, repeatB2=1,
nfolds=10, FSmethod1=NULL, FSmethod2=NULL,
cutP1=0.1, cutP2=0.1, fdr2=NULL, FScore=param1,
classifier, predMode, paramlist, innerCore=param2)## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 12.00 20.00 37.00 45.57 55.00 164.00
##
## Levels of predicted Y = 2
##
## AUC ACC R2
## R2 0.725 0.725 0.259print(result)## AUC ACC R2
## R2 0.725 0.725 0.2595.3 Stage-2 data exploration
5.3.1 Generation of stage-2 data
Here we demonstrate using supervised random forest method on genome-wide DNA methylation. Gene ontological pathways are used for the generation of stage-2 data.
## Pathway level data or stage-2 data prepared by reconBySupervised()
stage2dataA <- readRDS(system.file("extdata", "/stage2dataA.rds",
package="BioMM"))
head(stage2dataA[,1:5])## label GO:0000027 GO:0000045 GO:0000050 GO:0000060
## GSM951223 1 0.663 0.557 0.707 0.653
## GSM951231 1 0.655 0.710 0.737 0.686
## GSM951249 1 0.776 0.568 0.757 0.533
## GSM951273 1 0.664 0.510 0.741 0.642
## GSM951214 1 0.662 0.632 0.530 0.582
## GSM951270 1 0.419 0.366 0.474 0.415dim(stage2dataA)## [1] 40 51#### Alternatively, 'stage2dataA' can be created by the following code:
## Parameters
classifier <- "randForest"
predMode <- "probability"
paramlist <- list(ntree=300, nthreads=40)
param1 <- MulticoreParam(workers = 1)
param2 <- MulticoreParam(workers = 10)
## If BioMM is installed from Github, please use the following assignment:
## param1 <- 1
## param2 <- 10
set.seed(123)
## This will take a bit longer to run
stage2dataA <- reconBySupervised(trainDataList=pathlist, testDataList=NULL,
resample="BS", dataMode="allTrain",
repeatA=25, repeatB=1, nfolds=10,
FSmethod=NULL, cutP=0.1, fdr=NULL, FScore=param1,
classifier, predMode, paramlist,
innerCore=param2, outFileA=NULL, outFileB=NULL)5.3.2 Visualization
5.3.2.1 Explained variation of stage-2 data
The distribution of the proportion of variance explained for the individual
generated feature of stage-2 data for the classification task is illustrated
plotVarExplained() below. Nagelkerke pseudo R-squared measure is used to
compute the explained variance. The argument posF=TRUE indicates that only
positively outcome-associated features are plotted, since negative
associations likely reflect random effects in the underlying data [6].
param <- MulticoreParam(workers = 1)
## If BioMM is installed from Github, please use the following assignment:
## param <- 1
plotVarExplained(data=stage2dataA, posF=TRUE,
core=param, horizontal=FALSE, fileName=NULL)## png
## 2
5.3.2.2 Prioritization of outcome-associated functional patterns
plotRankedFeature() is employed to rank and visualize the outcome-associated
features from stage-2 data. The argument topF=10 and posF=TRUE are used to
define the top 10 positively outcome-associated features. Nagelkerke pseudo
R-squared measure is utilized to evaluate the importance of the ranked
features as indicated by the argument rankMetric="R2". The size of the
investigated pathway is pictured as the argument colorMetric="size".
param <- MulticoreParam(workers = 10)
## If BioMM is installed from Github, please use the following assignment:
## param <- 1
topPath <- plotRankedFeature(data=stage2dataA,
posF=TRUE, topF=10,
blocklist=pathlist,
rankMetric="R2",
colorMetric="size",
core=param, fileName=NULL)
5.3.2.3 The significance of CpGs in pathways of interest
cirPlot4pathway() illustrates the significance of the individual CpGs falling
into a set of pathways. Here the top 10 outcome-associated pathways are investigated.
Negative log P value is used to define the significance of each CpG within these pathways.
core <- MulticoreParam(workers = 10)
## If BioMM is installed from Github, use the following assignment:
## core <- 10
pathID <- gsub("\\.", ":", rownames(topPath))
pathSet <- pathlist[is.element(names(pathlist), pathID)]
pathMatch <- pathSet[match(pathID, names(pathSet))]
cirPlot4pathway(datalist=pathMatch,
topPathID=names(pathMatch), core, fileName=NULL)## [1] 1
## [1] 2
## [1] 3
## [1] 4
## [1] 5
## [1] 6
## [1] 7
## [1] 8
## [1] 9
## [1] 10## Circular_Manhattan Plotting pval...
## Plots are stored in: /tmp/Rtmp7yftvq/Rbuild766472049e0d/BioMM/vignettes## png
## 3
5.4 Computational consideration
BioMM with supervised models at both stages incorporating pathway based stratification method will take longer to run than unsupervised approaches. But the prediction is more powerful. Therefore, we suggest the former even if the computation is more demanding, as the adoption of 5G is pushing advances in computational storage and speed. Parallel computing is implemented and recommended for such scenario. In this vignette, due to the runtime, we only showcased the smaller examples and models with less computation.
6 Session information
sessionInfo()## R version 3.6.1 (2019-07-05)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 18.04.3 LTS
##
## Matrix products: default
## BLAS: /home/biocbuild/bbs-3.10-bioc/R/lib/libRblas.so
## LAPACK: /home/biocbuild/bbs-3.10-bioc/R/lib/libRlapack.so
##
## locale:
## [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C LC_TIME=en_US.UTF-8
## [4] LC_COLLATE=C LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
## [7] LC_PAPER=en_US.UTF-8 LC_NAME=C LC_ADDRESS=C
## [10] LC_TELEPHONE=C LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
##
## attached base packages:
## [1] parallel stats graphics grDevices utils datasets methods base
##
## other attached packages:
## [1] CMplot_3.4.0 variancePartition_1.16.0 Biobase_2.46.0
## [4] BiocGenerics_0.32.0 scales_1.0.0 limma_3.42.0
## [7] vioplot_0.3.2 zoo_1.8-6 sm_2.2-5.6
## [10] pROC_1.15.3 e1071_1.7-2 glmnet_2.0-18
## [13] foreach_1.4.7 Matrix_1.2-17 rms_5.1-3.1
## [16] SparseM_1.77 Hmisc_4.2-0 ggplot2_3.2.1
## [19] Formula_1.2-3 survival_2.44-1.1 lattice_0.20-38
## [22] ranger_0.11.2 BiocParallel_1.20.0 BioMM_1.2.0
## [25] BiocStyle_2.14.0
##
## loaded via a namespace (and not attached):
## [1] nlme_3.1-141 bitops_1.0-6 pbkrtest_0.4-7 doParallel_1.0.15
## [5] RColorBrewer_1.1-2 progress_1.2.2 tools_3.6.1 backports_1.1.5
## [9] R6_2.4.0 rpart_4.1-15 KernSmooth_2.23-16 lazyeval_0.2.2
## [13] colorspace_1.4-1 nnet_7.3-12 withr_2.1.2 prettyunits_1.0.2
## [17] tidyselect_0.2.5 gridExtra_2.3 compiler_3.6.1 quantreg_5.51
## [21] htmlTable_1.13.2 sandwich_2.5-1 labeling_0.3 bookdown_0.14
## [25] caTools_1.17.1.2 checkmate_1.9.4 polspline_1.1.16 mvtnorm_1.0-11
## [29] stringr_1.4.0 digest_0.6.22 foreign_0.8-72 minqa_1.2.4
## [33] rmarkdown_1.16 colorRamps_2.3 base64enc_0.1-3 pkgconfig_2.0.3
## [37] htmltools_0.4.0 lme4_1.1-21 htmlwidgets_1.5.1 rlang_0.4.1
## [41] rstudioapi_0.10 gtools_3.8.1 acepack_1.4.1 dplyr_0.8.3
## [45] magrittr_1.5 Rcpp_1.0.2 munsell_0.5.0 stringi_1.4.3
## [49] multcomp_1.4-10 yaml_2.2.0 MASS_7.3-51.4 plyr_1.8.4
## [53] gplots_3.0.1.1 grid_3.6.1 gdata_2.18.0 crayon_1.3.4
## [57] splines_3.6.1 hms_0.5.1 zeallot_0.1.0 knitr_1.25
## [61] pillar_1.4.2 tcltk_3.6.1 boot_1.3-23 reshape2_1.4.3
## [65] nsprcomp_0.5.1-2 codetools_0.2-16 glue_1.3.1 evaluate_0.14
## [69] latticeExtra_0.6-28 data.table_1.12.6 BiocManager_1.30.9 vctrs_0.2.0
## [73] nloptr_1.2.1 MatrixModels_0.4-1 gtable_0.3.0 purrr_0.3.3
## [77] assertthat_0.2.1 xfun_0.10 class_7.3-15 tibble_2.1.3
## [81] iterators_1.0.12 cluster_2.1.0 TH.data_1.0-107 References
[1] NIPS ML4H submission: Chen, J. and Schwarz, E., 2017. BioMM: Biologically-informed Multi-stage Machine learning for identification of epigenetic fingerprints. arXiv preprint arXiv:1712.00336.
[2] Breiman, L. (2001). “Random forests.” Machine learning 45(1): 5-32.
[3] Cortes, C., & Vapnik, V. (1995). “Support-vector networks.” Machine learning 20(3): 273-297.
[4] Friedman, J., Hastie, T., & Tibshirani, R. (2010). “Regularization paths for generalized linear models via coordinate descent.” Journal of statistical software 33(1): 1.
[5] Wold, S., Esbensen, K., & Geladi, P. (1987). “Principal component analysis.” Chemometrics and intelligent laboratory systems 2(1-3): 37-52.
[6] Claudia Perlich and Grzegorz Swirszcz. On cross-validation and stacking: Building seemingly predictive models on random data. ACM SIGKDD Explorations Newsletter, 12(2):11-15, 2011.