Cesaire J. K. Fouodo
Recent technological advances have enabled the simultaneous collection of multi-omics data, i.e., different types or modalities of molecular data across various organ tissues of patients. For integrative predictive modeling, the analysis of such data is particularly challenging. Ideally, data from the different modalities are measured in the same individuals, allowing for early or intermediate integrative techniques. However, they are often not applicable when patient data only partially overlap, which requires either reducing the datasets or imputing missing values. Additionally, the characteristics of each data modality may necessitate specific statistical methods rather than applying the same method across all modalities. Late integration modeling approaches analyze each data modality separately to obtain modality-specific predictions. These predictions are then integrated to train aggregative models like Lasso, random forests, or compute the weighted mean of modality-specific predictions.
We introduce the R package fuseMLR for late integration prediction modeling. This comprehensive package enables users to define a training process with multiple data modalities and modality-specific machine learning methods. The package is user-friendly, facilitates variable selection and training of different models across modalities, and automatically performs aggregation once modality-specific training is completed. We simulated multi-omics data to illustrate the usage of our new package for conducting late-stage multi-omics integrative modeling.
fuseMLR is an object-oriented package based on R6 version 2.5.1.
Refer to our cheat
sheet
for a quick overview of classes and functionalities.
Install the development version from GitHub with
devtools::install_github("imbs-hl/fuseMLR")The following figure illustrates the general architecture of fuseMLR:
The general architecture of fuseMLR includes the collection classes
Training, TrainLayer, and TrainMetaLayer. TrainLayer and
TrainMetaLayer are stored within a Training instance, while
TrainData, Lrnr, and VarSel (for variable selection) are stored
within a TrainLayer or MetaTrainLayer instance. An Training object
can be used to automatically conduct layer-specific variable selection
and train layer-specfic and meta models.
The following example is based on simulated data available in fuseMLR.
Data have been simulated using the R package InterSIM, version 2.2.0.
library(fuseMLR)
library(UpSetR)
library(ranger)
library(DescTools)Two types of data were simulated: training and testing datasets. Each
consists of four data.frames—gene expression, protein expression,
methylation data, and target variable observations. Individuals are
organized in rows, variables in columns, with an additional column for
individual IDs. In total,
data("entities")
# This is a list containing two lists of data: training and test.
# Each sublist contains three entities.
str(object = entities, max.level = 2L)## List of 2
## $ training:List of 4
## ..$ geneexpr :'data.frame': 50 obs. of 132 variables:
## ..$ proteinexpr:'data.frame': 50 obs. of 161 variables:
## ..$ methylation:'data.frame': 50 obs. of 368 variables:
## ..$ target :'data.frame': 70 obs. of 2 variables:
## $ testing :List of 4
## ..$ geneexpr :'data.frame': 20 obs. of 132 variables:
## ..$ proteinexpr:'data.frame': 20 obs. of 161 variables:
## ..$ methylation:'data.frame': 20 obs. of 368 variables:
## ..$ target :'data.frame': 30 obs. of 2 variables:
Variable selection, training and prediction are the main functionalities
of fuseMLR. We can perform variable selection, train and fuse models
for training studies, and predict new studies.
We need to set up training resources.
training <- Training$new(id = "training",
ind_col = "IDS",
target = "disease",
target_df = entities$training$target)
# See also training$summary()
print(training)## Training : training
## Status : Not trained
## Number of layers: 1
## Layers trained : 0
## n : 70
- Prepare new training layers: Training layers are components of a study
and represent the second component of a
fuseMLRobject.
tl_ge <- TrainLayer$new(id = "geneexpr", training = training)
tl_pr <- TrainLayer$new(id = "proteinexpr", training = training)
tl_me <- TrainLayer$new(id = "methylation", training = training)
# We also prepare the meta layer for the meta analysis.
tl_meta <- TrainMetaLayer$new(id = "meta_layer", training = training)- Add training data (entities) to layers: Exclude the meta layer, as it is modified internally after the training phase.
train_data_ge <- TrainData$new(id = "geneexpr",
train_layer = tl_ge,
data_frame = entities$training$geneexpr)
train_data_pr <- TrainData$new(id = "proteinexpr",
train_layer = tl_pr,
data_frame = entities$training$proteinexpr)
train_data_me <- TrainData$new(id = "methylation",
train_layer = tl_me,
data_frame = entities$training$methylation)
# An overview of the gene expression training data
print(train_data_ge)## TrainData : geneexpr
## Layer : geneexpr
## ind. id. : IDS
## target : disease
## n : 50
## Missing : 0
## p : 133
# An overview of the gene expression training layer
print(tl_ge)## TrainLayer : geneexpr
## Status : Not trained
## Nb. of objects stored : 1
## -----------------------
## key class
## 1 geneexpr TrainData
# An overview of the training resources
print(training)## Training : training
## Status : Not trained
## Number of layers: 5
## Layers trained : 0
## n : 70
- An upset plot of the training data: Visualize patient overlap across layers.
training$upset(order.by = "freq")
We need to set up variable selection methods to our training resources. Note that this can be the same method or different layer-specific methods. For simplicity, we will set up the same method on all layers.
- Preparation parameters of the variable selection method.
same_param_varsel <- ParamVarSel$new(id = "ParamVarSel",
param_list = list(num.trees = 1000L,
mtry = 3L,
probability = TRUE))
print(same_param_varsel)## Class: ParamVarSel
## id : ParamVarSel
## Parameter combination
## $num.trees
## [1] 1000
##
## $mtry
## [1] 3
##
## $probability
## [1] TRUE
- Instantiate the variable selection method and assign training layers.
varsel_ge <- VarSel$new(id = "varsel_geneexpr",
package = "Boruta",
varsel_fct = "Boruta",
param = same_param_varsel,
train_layer = tl_ge)
varsel_pr <- VarSel$new(id = "varsel_proteinexpr",
package = "Boruta",
varsel_fct = "Boruta",
param = same_param_varsel,
train_layer = tl_pr)
varsel_me <- VarSel$new(id = "varsel_methylation",
package = "Boruta",
varsel_fct = "Boruta",
param = same_param_varsel,
train_layer = tl_me)- Perform variable selection on our training resources
set.seed(5467)
var_sel_res <- training$varSelection()
print(var_sel_res)## Layer variable
## 1 geneexpr ACVRL1
## 2 geneexpr AKT1S1
## 3 geneexpr BAX
## 4 geneexpr BCL2L1
## 5 geneexpr CDKN1A
## 6 geneexpr CTNNB1
## 7 geneexpr EEF2
## 8 geneexpr EEF2K
## 9 geneexpr EIF4EBP1
## 10 geneexpr MAPK14
## 11 geneexpr NFKB1
## 12 geneexpr PREX1
## 13 geneexpr PRKCD
## 14 geneexpr PXN
## 15 geneexpr SHC1
## 16 geneexpr SRC
## 17 geneexpr XRCC5
## 18 geneexpr YAP1
## 19 proteinexpr Caveolin.1
## 20 proteinexpr Rab.25
## 21 methylation cg09637363
## 22 methylation cg25060573
## 23 methylation cg23989635
## 24 methylation cg22679003
## 25 methylation cg01663570
## 26 methylation cg09186685
## 27 methylation cg20253551
## 28 methylation cg03386722
## 29 methylation cg00241355
## 30 methylation cg26187237
## 31 methylation cg24991452
## 32 methylation cg20042228
## 33 methylation cg23641145
## 34 methylation cg01228636
## 35 methylation cg17489897
## 36 methylation cg21573601
## 37 methylation cg00059930
## 38 methylation cg19427472
## 39 methylation cg16925486
## 40 methylation cg00239419
## 41 methylation cg23323671
## 42 methylation cg07160163
## 43 methylation cg12507125
For each layer, the variable selection results show the chosen variables. In this example, we perform variable selection. Users can opt to conduct variable selection on individual layers if desired.
We can now train our models using the subset of selected variables. Users can choose to set up layer-specific learners, but for illustration, we will use the same learner for all layers.
- Set up the same leaner parameters.
same_param <- ParamLrner$new(id = "ParamRanger",
param_list = list(probability = TRUE,
mtry = 1L),
hyperparam_list = list(num.trees = 1000L))- Set up learners for each layer. We will use a weighted sum,
implemented internally by
fuseMLR, for the meta-analysis.
lrner_ge <- Lrner$new(id = "ranger",
package = "ranger",
lrn_fct = "ranger",
param = same_param,
train_layer = tl_ge)
lrner_pr <- Lrner$new(id = "ranger",
package = "ranger",
lrn_fct = "ranger",
param = same_param,
train_layer = tl_pr)
lrner_me <- Lrner$new(id = "ranger",
package = "ranger",
lrn_fct = "ranger",
param = same_param,
train_layer = tl_me)
lrner_meta <- Lrner$new(id = "weighted",
lrn_fct = "weightedMeanLearner",
param = ParamLrner$new(id = "ParamWeighted",
param_list = list(),
hyperparam_list = list()),
na_rm = FALSE,
train_layer = tl_meta)- Train the models with the selected variables.
set.seed(5462)
# Retrieve the target variable for resampling reasons. Resampling will be used by
# fuseMLR to generate meta data.
disease <- training$getTargetValues()$disease
trained <- training$train(resampling_method = "caret::createFolds",
resampling_arg = list(y = disease,
k = 10L),
use_var_sel = TRUE)
# Let us now check the status of our training resources.
print(trained)## Training : training
## Status : Trained
## Number of layers: 5
## Layers trained : 4
## n : 70
# Let us check the status of a layer as well.
print(tl_ge)## TrainLayer : geneexpr
## Status : Trained
## Nb. of objects stored : 4
## -----------------------
## key class
## 1 geneexpr TrainData
## 2 varsel_geneexpr VarSel
## 3 ranger Lrner
## 4 rangerMo Model
## On the meta model
tmp_model <- tl_meta$getModel()
print(tmp_model$getBaseModel())## geneexpr proteinexpr methylation
## 0.2336235 0.4249688 0.3414078
## attr(,"class")
## [1] "weightedMeanLearner"
- Retrieve the basic model of a specific layer.
model_ge <- tl_ge$getModel()
print(model_ge)## Class : Model
##
## Learner info.
## -----------------------
## Learner : ranger
## TrainLayer : geneexpr
## Package : ranger
## Learn function : ranger
##
## Train data info.
## -----------------------
## TrainData : geneexpr
## Layer : geneexpr
## ind. id. : IDS
## target : disease
## n : 50
## Missing : 0
## p : 19
Now, we have created training resources, performed variable selection and trained the models with the chosen variables. In this section, we create testing resources and make predictions for new data.
- Create the testing object.
testing <- Testing$new(id = "testing", ind_col = "IDS")- Create new layers.
nl_ge <- TestLayer$new(id = "geneexpr", testing = testing)
nl_pr <- TestLayer$new(id = "proteinexpr", testing = testing)
nl_me <- TestLayer$new(id = "methylation", testing = testing)- Instantiate and add new training data to new layers.
new_data_ge <- TestData$new(id = "geneexpr",
new_layer = nl_ge,
data_frame = entities$testing$geneexpr)
new_data_pr <- TestData$new(id = "proteinexpr",
new_layer = nl_pr,
data_frame = entities$testing$proteinexpr)
new_data_me <- TestData$new(id = "methylation",
new_layer = nl_me,
data_frame = entities$testing$methylation)- An upset plot of the training data: Visualize patient overlap across layers.
testing$upset(order.by = "freq")
- Predict the testing object.
predictions <- training$predict(testing = testing)
print(predictions)## $predicting
## Predicting : testing
## Nb. layers : 4
##
## $predicted_values
## IDS geneexpr proteinexpr methylation meta_layer
## 1 patient23 0.3899925 0.59490714 0.17449127 0.40350102
## 2 patient77 0.4222480 0.49866468 0.14058889 0.35856210
## 3 patient62 0.7736817 0.96993571 NA 0.90031823
## 4 patient43 0.3301254 NA NA 0.33012540
## 5 patient8 0.7346532 0.85212024 0.83084960 0.81741522
## 6 patient74 0.5555329 0.69613294 0.54980476 0.61332790
## 7 patient29 0.3362905 0.46632857 0.27474444 0.37054031
## 8 patient17 0.3997508 0.34566627 NA 0.36485176
## 9 patient25 0.2778349 0.45710040 0.09602103 0.29194448
## 10 patient54 0.7863302 NA 0.84842857 0.82319925
## 11 patient60 0.7415452 0.84184524 0.79166865 0.80128213
## 12 patient44 0.3717952 NA NA 0.37179524
## 13 patient1 0.8211250 0.93648373 NA 0.89556234
## 14 patient76 0.6840357 NA 0.60861508 0.63925695
## 15 patient16 0.6923929 NA 0.69758968 0.69547832
## 16 patient27 0.3699286 NA 0.20396786 0.27139431
## 17 patient58 0.5648032 0.75506071 0.76331349 0.71342965
## 18 patient52 0.4850317 0.13170119 NA 0.25703870
## 19 patient10 0.2517774 NA NA 0.25177738
## 20 patient72 0.7087675 0.94573135 0.62339762 0.78032379
## 21 patient39 NA 0.08271032 NA 0.08271032
## 25 patient46 NA 0.22185635 0.51994405 0.35464937
## 26 patient97 NA 0.70887738 0.86840476 0.77994413
## 27 patient31 NA 0.28228095 NA 0.28228095
## 31 patient87 NA 0.29549603 0.26039048 0.27985710
## 33 patient59 NA 0.14183968 0.38772183 0.25137601
## 34 patient2 NA 0.55434325 0.79045437 0.65952675
## 53 patient85 NA NA 0.15880913 0.15880913
## 60 patient3 NA NA 0.57764881 0.57764881
- Prediction performances for layer-specific available patients, and all patients on the meta layer.
pred_values <- predictions$predicted_values
actual_pred <- merge(x = pred_values,
y = entities$testing$target,
by = "IDS",
all.y = TRUE)
x <- as.integer(actual_pred$disease == 2L)
# On all patients
perf_estimated <- sapply(X = actual_pred[ , 2L:5L], FUN = function (my_pred) {
bs <- BrierScore(x = x[complete.cases(my_pred)],
pred = my_pred[complete.cases(my_pred)], na.rm = TRUE)
return(bs)
})
print(perf_estimated)## geneexpr proteinexpr methylation meta_layer
## 0.1147740 0.1645159 0.0815040 0.1231755
- Prediction performances for overlapping individuals.
# On overlapping patients
perf_overlapping <- sapply(X = actual_pred[complete.cases(actual_pred),
2L:5L],
FUN = function (my_pred) {
bs <- BrierScore(x = x[complete.cases(actual_pred)], pred = my_pred)
return(bs)
})
print(perf_overlapping)## geneexpr proteinexpr methylation meta_layer
## 0.12551583 0.13679344 0.06749516 0.09629256
Note that our example is based on simulated data for usage illustration; only one run is not enough to appreciate the performances of our models.
© 2024 Institute of Medical Biometry and Statistics (IMBS). All rights reserved.