In drorberel/Bioc2mlr:

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/"
)

Bioc2mlr

R package to bridge between Bioconductor’s S4 complex genomic data container, to mlr, a meta machine learning aggregator package.

Bioc2mlr is designed to convert Bioconductor S4 assay data containers summarizedExperiment, MultiAssayExperiment into generalized machnine learning environment.

Bioconductor's S4 data containers for genomic assays are popular, well established data structures. Their data architecture facilitates the application of common analytical procedures and well established statistical methodologies to large assay data. They are extensible to encompass new emerging technologies and analytical methods. However, the S4 system enforces strict constraints on the data and these constraints raise barriers for interoperability and integration with software and packages outside of Bioconductor's repository.

mlr is a comprehensive package for machine learning. It aggregates hundreds of supervised and unsupervised models and facilitates analytics such as resampling, benchmarking, tuning, and ensemble. The mlrCPO package extends mlr's pre-processing and feature engineering functionality via composable Preprocessing Operators (CPO) 'pipelines'.

Bioc2mlr is a compact utility package designed to bridge between these approaches. It deploys transformations of SummarizedExperiment and MultiAssayExperiment S4 data structures into mlr's expected format. It also implements Bioconductor's popular feature selection (filtering) methods used by limma package and others, as a CPO. The vignettes present comparisons to the MLInterfaces package, which aims to achieve similar goals, and presents workflows for popular publicly available genomic datasets such as curatedTCGAData.

Vision:

knitr::include_graphics("man/figures/vision.jpg")

Why mlr?

Because it support functional / clustered data.

What about caret/SuperLearner/tidymodels ?

Sure. Work in progress. keep tuned.
caret: Bioc2caret
SuperLearner: Bioc2SuperLearner
tidymodels: Bioc2tidymodels

Installation

# Install development version from GitHub
devtools::install_github("drorberel/Bioc2mlr")

# TBA: Install release version from CRAN
# install.packages("Bioc2mlr")

Current implementations

Two Bioconductor assay container are currently implemented: SummarizedExperiment for a single assay (though may have multiple sub-assays slots), and MultiAssayExperiment for multiple assays. Within the machine-learning framework, the two main steps that are adapted are the pre-processing step, followed by the (multivariate) model fitting.

Tools will be demonstrated for each of these 4 combinations.

| S4 assay data container | Pre-processing (TBA)| Model (multivariate) | |-----------------------------|:-------------:|-----------------------:| | SummarizedExperiment (SE) | limmaCPO | Fun_SE_to_taskFunc | | MultiAssayExperiment (MAE) | UnivCPO | Fun_MAE_to_taskFunc |

Usage

Vignettes

Proof of concept demonstration

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>", message=FALSE)

library(tidyverse)  
library(magrittr)

library(Bioc2mlr)
library(mlr)
select<-dplyr::select

Model-evaluation (ML)

A. SummarizedExperiment (SE)

Convert raw data from SE S4 class, to mlr's "task"

data(Golub_Merge, package = 'golubEsets') # ExpressionSet 
smallG<-Golub_Merge[200:259,]
smallG

library(SummarizedExperiment)
smallG_SE<-makeSummarizedExperimentFromExpressionSet(smallG)

# functional:
task_SE_Functional<-Fun_SE_to_taskFunc(smallG_SE, param.Y.name = 'ALL.AML', param.covariates = NULL, param_positive_y_level = 'ALL', task_return_format = 'functional', task_type = 'classif') ## will work with either 1 or multiple assayS
task_SE_Functional

task_train<-task_SE_Functional %>% subsetTask(subset = 1:40)
task_test <-task_SE_Functional %>% subsetTask(subset = 41:72)
classif.lrn = makeLearner("classif.knn")
model<-train(classif.lrn, task_train)
Predict<-model %>% predict(task_test)
Predict %>% calculateConfusionMatrix()

B. MultiAssayExperiment (MAE)

Convert raw data from MAE S4 class, to mlr's "task"

library(MultiAssayExperiment)
miniACC
# miniACC %>% sampleMap %>% data.frame %>% dplyr::select(primary, assay) %>% table # no replicates within same assay

task_Functional_MAE<-Fun_MAE_to_taskFunc(miniACC, param.Y.name = 'vital_status', param.covariates = c('gender','days_to_death'), param_positive_y_level = '1', task_type = 'classif')
task_Functional_MAE

library(bartMachine)
classif_lrn_bartMachine<-makeLearner("classif.bartMachine")
model_bartMachine<-train(classif_lrn_bartMachine, task_Functional_MAE)
Predict_bartMachine<-model_bartMachine %>% predict(task_Functional_MAE)
Predict_bartMachine %>% calculateConfusionMatrix()

Case studies (TBA):

1. CAVDmetaMAE: proof-of-concept example

CAVD dataspace is an online resouce to access and analyze HIV vaccine experimental assay data. It is annotated, and accessible via either online tool, and R API DataSpaceR.

The CAVDmetaMAE package implement a hypothesis-free approach, to find best candidates of immune biomarkers, that are associated with experimental groups, at each study (separately), and across all studies together (meta-analysis).

Within each study, immune biomarkers will be analyzed by both single assays, and combinations across multiple assays.
https://github.com/drorberel/CAVDmetaMAE
Private repo. Access permission by request.

knitr::include_graphics("man/figures/CAVDmetaMAE_ridge.jpg")

knitr::include_graphics("man/figures/pheatmaps_DT.png")

2. Multi-assay customized feature selection for JDRF data (under review)

Data curation

A. NCBI/GEO -> SEs -> MAE -> task (DataPackageR)

Paper’s reproducible results

B. Biomarker discovery:
B.1 Feature selection:
Fun_lrn_univ_only_makePrep_MaG
Fun_lrn_univ_Clusters_All_makePrep_MaG
B.2 Sensitivity analysis

Customized Multi-assay feature selection

C. (TBA) UnivCPO, UnivClustCPO (refactoring the above makePreprocWrapper()

3. Annotated public datasets (TBA)

TCGA curatedTCGAData
Microbiome curatedMetagenomicData

4. Customized multi-assay CPOs / composable pipelines (TBA)

Omicade4CPO Omicade4
mixomicsCPO mixomics

5. Analysis Workflows: (TBA)

Utilize MAE to collapse genes to sets/modules/pathways
Fortified: pheatmaps, ggfortify

drorberel/Bioc2mlr documentation built on May 16, 2019, 5 a.m.