Package 'IntegratedLearner'

Title Meta level Objective function: NNLS for gaussian; Rank loss for binary observations

Description

Title Meta level Objective function: NNLS for gaussian; Rank loss for binary observations

Usage

auc_obj(b, X, Y)

Arguments

b	Weights vector
X	Design matrix (data frame)
Y	Outcome variable

Value

1 - AUC

Examples

X <- data.frame(m1 = c(0.1, 0.4, 0.8, 0.9), m2 = c(0.2, 0.3, 0.7, 0.95))
Y <- c(0, 0, 1, 1)
auc_obj(b = c(0.5, 0.5), X = X, Y = Y)

---

Integrated machine learning for multi-omics prediction (continuous/binary outcomes)

Description

Performs integrated machine learning to predict a binary or continuous outcome based on two or more omics layers (views). This function implements the core IntegratedLearner engine for non-survival outcomes.

Usage

IL_conbin(
  feature_table,
  sample_metadata,
  feature_metadata,
  feature_table_valid = NULL,
  sample_metadata_valid = NULL,
  folds = 5,
  seed = 1234,
  base_learner = "sl_bart",
  base_screener = "All",
  run_screening = FALSE,
  screen_pct = NULL,
  filter_method = NULL,
  filter_pct = NULL,
  prevalence_pct = NULL,
  meta_learner = "sl_nnls_auc",
  run_concat = TRUE,
  run_stacked = TRUE,
  run_intermediate = FALSE,
  cooperative_rho = c(0, 0.1, 0.25, 0.5, 1),
  cooperative_s = "lambda.min",
  cooperative_type_measure = NULL,
  drop_poor_performing_layers = FALSE,
  verbose = FALSE,
  print_learner = TRUE,
  refit.stack = FALSE,
  family = stats::gaussian(),
  ...
)

Arguments

feature_table	An R data frame containing multi-layer features (in rows) and samples (in columns). Column names of feature_metadata must match the row names of sample_metadata.
sample_metadata	An R data frame of metadata variables (in columns). Must have a column named subjectID describing per-subject unique identifiers. For longitudinal designs, this variable may be non-unique. Additionally, a column named Y must be present which is the continuous or binary outcome of interest. Row names of sample_metadata must match the column names of feature_table.
feature_metadata	An R data frame of feature-specific metadata across views (in columns) and features (in rows). Must have a column named featureID as a unique per-feature identifier and a column named featureType describing the source layers. Row names of feature_metadata must match the row names of feature_table.
feature_table_valid	Feature table from validation set for which prediction is desired. Must have the exact same structure as feature_table. If missing, uses feature_table.
sample_metadata_valid	Sample-specific metadata table from independent validation set when available. Must have the exact same structure as sample_metadata.
folds	How many folds in the V-fold nested cross-validation? Default is 5.
seed	Specify the seed for reproducibility. Default is 1234.
base_learner	Base learner for late fusion and early fusion. Default: sl_bart.
base_screener	Deprecated for this backend; currently ignored and kept only for backward compatibility.
run_screening	Logical; if TRUE, run supervised screening within each training fold (and again on full training data) before fitting base models.
screen_pct	Percentage of features to retain during screening ((0,100]). Applied after any optional filtering.
filter_method	Optional feature-filter method before model fitting. Supported values are 'prevalence' (top detected features by prevalence) and 'variance' (top features by caret-based variance ranking via nearZeroVar + empirical variance). If NULL, defaults to 'prevalence' when filtering is requested.
filter_pct	Optional retention percentage (in (0,100]) for the selected filter_method. Keeps the top filter_pct percent features.
prevalence_pct	Optional retention percentage (in (0,100]) for prevalence-based filtering before model fitting. Deprecated alias of filter_pct with filter_method = 'prevalence'.
meta_learner	Meta-learner for late fusion (stacked generalization). Defaults to sl_nnls_auc.
run_concat	Should early fusion be run? Default is TRUE.
run_stacked	Should stacked model (late fusion) be run? Default is TRUE.
run_intermediate	Should standalone cooperative learning via multiview be run? Default is FALSE. When TRUE, only the cooperative multiview model is fit; base learners, early fusion, and late fusion are skipped.
cooperative_rho	Non-negative agreement-penalty values to evaluate for cooperative learning.
cooperative_s	Penalty value used when predicting from the selected cooperative model. Defaults to "lambda.min".
cooperative_type_measure	Optional multiview CV loss/score. If NULL, an outcome-appropriate default is used.
drop_poor_performing_layers	Logical; if TRUE, layers with single-layer performance below the screening threshold are removed from early and late fusion only. The threshold is AUC < 0.5 for binary and R$^2$ < 0.5 for continuous outcomes. Single-layer outputs are still retained and reported.
verbose	logical; TRUE for printing SuperLearner progress. Default FALSE.
print_learner	logical; Should a detailed summary be printed? Default TRUE.
refit.stack	logical; For late fusion, refit predictions on the entire data are returned if specified. Default FALSE.
family	Allows gaussian() for continuous and binomial() for binary outcomes. Survival outcomes must be handled via IL_survival().
...	Additional arguments (currently unused).

Details

IL_conbin takes a training set (feature_table, sample_metadata, feature_metadata) and, optionally, a corresponding validation set, and returns predicted values based on the validation set. It also performs V-fold nested cross-validation to estimate the prediction accuracy of various fusion algorithms.

Two integration paradigms are supported: early and late. The software includes multiple ML models based on the SuperLearner R package as well as several data exploration capabilities and visualization modules in a unified estimation framework.

Although IL_conbin() is typically called internally by IntegratedLearner() after extracting multi-view feature tables from MultiAssayExperiment objects, advanced users may call IL_conbin() directly when they already have multi-layer feature tables and metadata in matrix/data.frame form.

Value

A list-like IntegratedLearner object containing fitted layer-specific, stacked, and concatenated models, cross-validated performance (AUC or R\302\262), and predictions for training and validation sets. When run_intermediate = TRUE, coefficient-based cooperative feature-importance scores from the selected multiview model are returned in cooperative_feature_importance, cooperative_feature_importance_by_layer, cooperative_feature_importance_signed, and cooperative_feature_importance_signed_by_layer.

Author(s)

Himel Mallick, [email protected]

See Also

IntegratedLearner, IL_survival()

Examples

set.seed(1)
n <- 20
feature_table <- as.data.frame(rbind(
  matrix(rnorm(3 * n), nrow = 3, dimnames = list(paste0("L1_F", 1:3), paste0("S", 1:n))),
  matrix(rnorm(2 * n), nrow = 2, dimnames = list(paste0("L2_F", 1:2), paste0("S", 1:n)))
))
sample_metadata <- data.frame(
  subjectID = paste0("ID", 1:n), Y = rnorm(n),
  row.names = colnames(feature_table)
)
feature_metadata <- data.frame(
  featureID = rownames(feature_table),
  featureType = c(rep("Layer1", 3), rep("Layer2", 2)),
  row.names = rownames(feature_table)
)
fit <- IL_conbin(
  feature_table = feature_table,
  sample_metadata = sample_metadata,
  feature_metadata = feature_metadata,
  folds = 3, base_learner = "SL.mean", run_stacked = FALSE,
  run_concat = FALSE, print_learner = FALSE, family = stats::gaussian()
)
names(fit)

---

Integrated machine learning for multi-omics multiclass classification

Description

Native multiclass backend used by IntegratedLearner() when family = binomial() and the outcome has more than two classes. This backend preserves the existing API while using multiclass-native probability modeling and multiclass stacking.

Usage

IL_multiclass(
  feature_table,
  sample_metadata,
  feature_metadata,
  feature_table_valid = NULL,
  sample_metadata_valid = NULL,
  folds = 5,
  seed = 1234,
  base_learner = "glmnet",
  base_screener = "All",
  run_screening = FALSE,
  screen_pct = NULL,
  filter_method = NULL,
  filter_pct = NULL,
  prevalence_pct = NULL,
  meta_learner = "glmnet",
  run_concat = TRUE,
  run_stacked = TRUE,
  run_intermediate = FALSE,
  cooperative_rho = c(0, 0.1, 0.25, 0.5, 1),
  cooperative_s = "lambda.min",
  cooperative_type_measure = NULL,
  verbose = FALSE,
  print_learner = TRUE,
  family = stats::binomial(),
  eps = 1e-15,
  run_majority = TRUE,
  run_weighted = TRUE,
  ...
)

Arguments

feature_table	An R data frame containing multi-layer features (in rows) and samples (in columns). Column names of feature_metadata must match the row names of sample_metadata.
sample_metadata	An R data frame of metadata variables (in columns). Must have a column named subjectID describing per-subject unique identifiers. For longitudinal designs, this variable may be non-unique. Additionally, a column named Y must be present which is the continuous or binary outcome of interest. Row names of sample_metadata must match the column names of feature_table.
feature_metadata	An R data frame of feature-specific metadata across views (in columns) and features (in rows). Must have a column named featureID as a unique per-feature identifier and a column named featureType describing the source layers. Row names of feature_metadata must match the row names of feature_table.
feature_table_valid	Feature table from validation set for which prediction is desired. Must have the exact same structure as feature_table. If missing, uses feature_table.
sample_metadata_valid	Sample-specific metadata table from independent validation set when available. Must have the exact same structure as sample_metadata.
folds	Number of subject-level folds for out-of-fold predictions. Default 5. Fully nested evaluation of the fusion pipeline is separate.
seed	Specify the seed for reproducibility. Default is 1234.
base_learner	Native multiclass base learner. Default "glmnet".
base_screener	Deprecated for this backend; currently ignored and kept only for backward compatibility.
run_screening	Logical; if TRUE, run supervised screening within each training fold (and again on full training data) before fitting base models.
screen_pct	Percentage of features to retain during screening ((0,100]). Applied after any optional filtering.
filter_method	Optional feature-filter method before model fitting. Supported values are 'prevalence' (top detected features by prevalence) and 'variance' (top features by caret-based variance ranking via nearZeroVar + empirical variance). If NULL, defaults to 'prevalence' when filtering is requested.
filter_pct	Optional retention percentage (in (0,100]) for the selected filter_method. Keeps the top filter_pct percent features.
prevalence_pct	Optional retention percentage (in (0,100]) for prevalence-based filtering before model fitting. Deprecated alias of filter_pct with filter_method = 'prevalence'.
meta_learner	Native multiclass stacking learner. Default "glmnet". Legacy sl_nnls_auc and SL.nnls aliases map to multinomial glmnet, not scalar layer-weight optimization.
run_concat	Should early fusion be run? Default is TRUE.
run_stacked	Should stacked model (late fusion) be run? Default is TRUE.
run_intermediate	Ignored for multiclass outcomes with a message; cooperative multiview learning is not currently supported for multiclass fits.
cooperative_rho	Non-negative agreement-penalty values to evaluate for cooperative learning.
cooperative_s	Penalty value used when predicting from the selected cooperative model. Defaults to "lambda.min".
cooperative_type_measure	Optional multiview CV loss/score. If NULL, an outcome-appropriate default is used.
verbose	logical; TRUE for printing SuperLearner progress. Default FALSE.
print_learner	logical; Should a detailed summary be printed? Default TRUE.
family	Allows gaussian() for continuous and binomial() for binary outcomes. Survival outcomes must be handled via IL_survival().
eps	Small positive constant used to stabilize probabilities.
run_majority	Logical; return multiclass majority voting. Default TRUE; ignored for other outcomes. Ties use the first class in training class order.
run_weighted	Logical; return multiclass probability averaging with nonnegative layer weights summing to one, learned by minimizing out-of-fold multiclass log loss. Default TRUE; ignored for other outcomes.
...	Additional arguments passed to native base-model fitting; these are not passed to the stacked meta-model or fusion-weight optimizer.

Details

Majority voting returns vote fractions under majority; these are not calibrated class probabilities. Ties use the first class in training class order. Learned averaging returns probabilities under weighted_average and stores final layer weights under model_fits$model_weighted$weights. Weights minimize multiclass log loss on out-of-fold layer probabilities. Training outputs cross-fit the weights across the existing folds. As with stacking, fully nested evaluation is needed for an unbiased assessment of the entire fitted pipeline. Results are also grouped under train_out and valid_out, each with single, early, and late entries. The late list contains all enabled methods: stacked, majority, and weighted_average. Each method includes split-specific probabilities, classes, and metrics; learned averaging also includes final weights and, for training, the weights fitted outside each fold. Existing flat outputs remain available. late_methods lists the returned late-fusion methods. Fusion entries contain train_prob, train_class, a metrics table, and scalar train_accuracy, train_balanced_accuracy, train_auc, and train_logloss; validation entries use the corresponding valid_ names. AUC is the macro average of available one-vs-rest class AUCs. Single-layer entries contain probability lists and class data frames. Without validation data, valid_out is NULL. Disabled early fusion gives early = NULL, and disabled late methods are omitted from late.

Feature importance retains a univariate class-association calculation: for each feature and class, the mean in the remaining classes is subtracted from the mean in that class and divided by the feature's overall training standard deviation plus 1e-8. The signed scores form a feature-by-class matrix in feature_importance_signed_by_class. feature_importance_signed_by_layer_by_class groups these matrices by omics layer and ranks features by mean absolute score across classes. feature_importance_global is the unsigned mean absolute score across classes, ranked highest first. These associations use the training features after optional filtering, including features excluded by screening; they are shared across fusion methods and are not model-specific attributions.

Value

A fitted multiclass IntegratedLearner object containing class probabilities, predicted classes, fitted single-layer and fusion models, and training/validation multiclass performance summaries when available.

Examples

set.seed(1)
n <- 24
feature_table <- as.data.frame(rbind(
  matrix(rnorm(3 * n), nrow = 3, dimnames = list(paste0("L1_F", 1:3), paste0("S", 1:n))),
  matrix(rnorm(2 * n), nrow = 2, dimnames = list(paste0("L2_F", 1:2), paste0("S", 1:n)))
))
y <- rep(c("A", "B", "C"), length.out = n)
sample_metadata <- data.frame(
  subjectID = paste0("ID", 1:n), Y = y,
  row.names = colnames(feature_table)
)
feature_metadata <- data.frame(
  featureID = rownames(feature_table),
  featureType = c(rep("Layer1", 3), rep("Layer2", 2)),
  row.names = rownames(feature_table)
)
fit <- IL_multiclass(
  feature_table = feature_table,
  sample_metadata = sample_metadata,
  feature_metadata = feature_metadata,
  folds = 3, base_learner = "multinom", meta_learner = "multinom",
  run_stacked = FALSE, run_concat = FALSE,
  print_learner = FALSE
)
fit$family
fit$late_methods
fit$train_out$late$weighted_average$weights
head(fit$feature_importance_signed_by_layer_by_class[[1]])

---

IntegratedLearner Survival Engine

Description

BioC-friendly survival backend used by IntegratedLearner() for time-to-event outcomes. This preserves the historical ILsurv() interface while using model fitting and fusion routines that do not depend on mlr3proba/mlr3extralearners.

Usage

ILsurv(
  feature_table,
  sample_metadata,
  feature_metadata,
  valid_feature_table = NULL,
  valid_sample_metadata = NULL,
  base_learner = "surv.coxph",
  folds = 5,
  seed = 123,
  run_screening = FALSE,
  screen_pct = NULL,
  drop_poor_performing_layers = FALSE,
  verbose = FALSE,
  do_early_fusion = TRUE,
  t_vec = NULL,
  t_vec_probs = c(0.05, 0.25, 0.5, 0.75, 0.95),
  layer_score = c("sum", "mean", "l2"),
  eps = 1e-12,
  weight_lambda = 0.02,
  weight_penalty = c("l2_to_uniform", "entropy"),
  weight_cap = 1,
  optim_maxit_cox = 4000,
  optim_maxit_ibs = 300,
  ibs_shrink_to_uniform = 0,
  run_intermediate = FALSE,
  cooperative_rho = c(0, 0.1, 0.25, 0.5, 1),
  cooperative_s = "lambda.min",
  cooperative_type_measure = NULL,
  ...
)

IL_survival(
  feature_table,
  sample_metadata,
  feature_metadata,
  valid_feature_table = NULL,
  valid_sample_metadata = NULL,
  base_learner = "surv.coxph",
  folds = 5,
  seed = 123,
  run_screening = FALSE,
  screen_pct = NULL,
  drop_poor_performing_layers = FALSE,
  verbose = FALSE,
  do_early_fusion = TRUE,
  t_vec = NULL,
  t_vec_probs = c(0.05, 0.25, 0.5, 0.75, 0.95),
  layer_score = c("sum", "mean", "l2"),
  eps = 1e-12,
  weight_lambda = 0.02,
  weight_penalty = c("l2_to_uniform", "entropy"),
  weight_cap = 1,
  optim_maxit_cox = 4000,
  optim_maxit_ibs = 300,
  ibs_shrink_to_uniform = 0,
  run_intermediate = FALSE,
  cooperative_rho = c(0, 0.1, 0.25, 0.5, 1),
  cooperative_s = "lambda.min",
  cooperative_type_measure = NULL,
  ...
)

Arguments

feature_table	An R data frame containing multi-layer features (in rows) and samples (in columns). Column names of feature_metadata must match the row names of sample_metadata.
sample_metadata	An R data frame of metadata variables (in columns). Must have a column named subjectID describing per-subject unique identifiers. For longitudinal designs, this variable may be non-unique. Additionally, a column named Y must be present which is the continuous or binary outcome of interest. Row names of sample_metadata must match the column names of feature_table.
feature_metadata	An R data frame of feature-specific metadata across views (in columns) and features (in rows). Must have a column named featureID as a unique per-feature identifier and a column named featureType describing the source layers. Row names of feature_metadata must match the row names of feature_table.
valid_feature_table	Validation feature table (features x samples).
valid_sample_metadata	Validation sample metadata containing time and event.
base_learner	Survival base learner.
folds	How many folds in the V-fold nested cross-validation? Default is 5.
seed	Specify the seed for reproducibility. Default is 1234.
run_screening	Logical; if TRUE, run supervised screening within each training fold (and again on full training data) before fitting base models.
screen_pct	Percentage of features to retain during screening ((0,100]). Applied after any optional filtering.
drop_poor_performing_layers	Logical; if TRUE, layers with single-layer performance below the screening threshold are removed from early and late fusion only. The threshold is AUC < 0.5 for binary and R$^2$ < 0.5 for continuous outcomes. Single-layer outputs are still retained and reported.
verbose	logical; TRUE for printing SuperLearner progress. Default FALSE.
do_early_fusion	Logical; run early fusion model.
t_vec	Optional explicit time points for COX-based late-fusion feature construction.
t_vec_probs	Quantiles used to derive t_vec when t_vec is NULL.
layer_score	Layer summary on cumhaz increments: 'sum', 'mean', or 'l2'.
eps	Numerical lower bound for survival probabilities.
weight_lambda	COX-weight optimization regularization strength.
weight_penalty	COX-weight optimization regularizer: 'l2_to_uniform' or 'entropy'.
weight_cap	Optional cap on any one layer weight (COX method).
optim_maxit_cox	Maximum iterations for COX-weight optimization.
optim_maxit_ibs	Maximum iterations for IBS-weight optimization.
ibs_shrink_to_uniform	Optional convex shrinkage of IBS weights toward uniform.
run_intermediate	Should standalone cooperative learning via multiview be run? Default is FALSE. When TRUE, only the cooperative multiview Cox model is fit; survival base learners, early fusion, and late fusion are skipped.
cooperative_rho	Non-negative agreement-penalty values to evaluate for cooperative learning.
cooperative_s	Penalty value used when predicting from the selected cooperative model. Defaults to "lambda.min".
cooperative_type_measure	Optional multiview CV loss/score. If NULL, Harrell's C is used.
...	Additional base-learner hyperparameters passed to base_learner. You may also pass model_args = list(...) as a named list where each entry is keyed by learner ID.

Details

Risk-to-survival Cox calibration is fitted using training outcomes and OOF risk scores, then held fixed for validation, including coefficients, baseline survival, missing-risk replacement, and the Brier time grid. A training Kaplan-Meier curve is used if Cox calibration is not estimable. Validation outcomes are used only for performance evaluation, including estimating censoring weights, and never to fit a prediction calibration model. IBS fusion is scored directly on its weighted survival curves; single-layer, early-fusion, Cox-fusion, and cooperative models use their frozen training calibration. Reported training probability-error metrics remain internal summaries because calibration and fusion fitting use training outcomes; they are not fully nested evaluation estimates.

Value

List with train_out and valid_out. Single-layer survival outputs include train_out$single$metrics, train_out$single$train_risk, and train_out$single$feature_importance_signed_by_layer, a named list of unweighted signed feature-importance scores by layer. Late-fusion results are returned under train_out$late$IBS, train_out$late$COX, and the analogous validation entries. When run_intermediate = TRUE, train_out$cooperative includes coefficient-based cooperative feature-importance scores from the selected multiview Cox model. brier_time_grid gives the shared training-derived grid used for training and validation Brier/IBS evaluation. survival_calibration stores the training-fitted risk-to-survival mappings. Single, early, late, and cooperative entries include train_surv/valid_surv probabilities when the corresponding model is fitted (lists for single layers, matrices for fusion models).

Examples

set.seed(1)
n <- 20
feature_table <- as.data.frame(rbind(
  matrix(rnorm(3 * n), nrow = 3, dimnames = list(paste0("L1_F", 1:3), paste0("S", 1:n))),
  matrix(rnorm(2 * n), nrow = 2, dimnames = list(paste0("L2_F", 1:2), paste0("S", 1:n)))
))
sample_metadata <- data.frame(
  subjectID = paste0("ID", 1:n),
  time = rexp(n, rate = 0.1),
  event = rbinom(n, 1, 0.6),
  row.names = colnames(feature_table)
)
sample_metadata$Y <- survival::Surv(sample_metadata$time, sample_metadata$event)
feature_metadata <- data.frame(
  featureID = rownames(feature_table),
  featureType = c(rep("Layer1", 3), rep("Layer2", 2)),
  row.names = rownames(feature_table)
)
fit <- ILsurv(
  feature_table = feature_table,
  sample_metadata = sample_metadata,
  feature_metadata = feature_metadata,
  folds = 3
)
names(fit)
fit$brier_time_grid

---

Integrated machine learning for multi-omics prediction and classification

Description

Performs integrated machine learning to predict a binary, continuous, or time-to-event outcome based on two or more omics layers (views). The IntegratedLearner function takes a training MultiAssayExperiment and, optionally, a validation MultiAssayExperiment, inspects standard Bioconductor containers and accessors, extracts multi-layer feature tables, and returns predicted values based on the validation set. It also performs V-fold nested cross-validation to estimate the prediction accuracy of various fusion algorithms. Two integration paradigms are supported: early and late. The software includes multiple ML models based on the SuperLearner R package as well as several data exploration capabilities and visualization modules in a unified estimation framework.

Usage

IntegratedLearner(
  MAE_train = NULL,
  MAE_valid = NULL,
  PCL_train = NULL,
  PCL_valid = NULL,
  experiment = NULL,
  assay.type = NULL,
  outcome_col = "Y",
  subject_id_col = "subjectID",
  na.rm = FALSE,
  folds = 5,
  seed = 1234,
  base_learner = "sl_bart",
  base_screener = "All",
  run_screening = FALSE,
  screen_pct = NULL,
  filter_method = NULL,
  filter_pct = NULL,
  prevalence_pct = NULL,
  meta_learner = "sl_nnls_auc",
  run_concat = TRUE,
  run_stacked = TRUE,
  run_intermediate = FALSE,
  cooperative_rho = c(0, 0.1, 0.25, 0.5, 1),
  cooperative_s = "lambda.min",
  cooperative_type_measure = NULL,
  drop_poor_performing_layers = FALSE,
  verbose = FALSE,
  print_learner = TRUE,
  refit.stack = FALSE,
  family = stats::gaussian(),
  run_majority = TRUE,
  run_weighted = TRUE,
  ...
)

Arguments

MAE_train	A MultiAssayExperiment containing the training data. Each experiment corresponds to one view (omics layer), usually stored as a SummarizedExperiment or TreeSummarizedExperiment. The colData of every selected experiment must contain a column named by outcome_col describing the outcome of interest (binary, continuous, or survival-encoded), and a subject identifier column named by subject_id_col.
MAE_valid	Optional MultiAssayExperiment containing an independent validation/test set for which prediction is desired. It must contain the same set of experiments used for training, with identical assay names and identical feature (row) names for each experiment. Additional experiments in mae_test are ignored.
PCL_train	Optional list of per-layer feature matrices (legacy PCL mode for backward compatibility).
PCL_valid	Optional validation list of per-layer feature matrices.
experiment	Optional character or integer vector specifying which experiments (layers) to extract from MAE_train (and MAE_valid, if supplied). If NULL, all experiments in MAE_train are used.
assay.type	Optional character vector of assay names, one per experiment. If NULL, each selected experiment must contain exactly one assay; otherwise, the user must supply assay.type.
outcome_col	Outcome column name in MAE/PCL sample metadata. Defaults to "Y" for backward compatibility.
subject_id_col	Subject identifier column name in MAE/PCL sample metadata. Defaults to "subjectID" for backward compatibility.
na.rm	Logical; if TRUE, features with missing values are dropped after extraction.
folds	How many folds in the V-fold nested cross-validation? Default is 10.
seed	Specify the arbitrary seed value for reproducibility. Default is 1234.
base_learner	Base learner for late fusion and early fusion. Check out the SuperLearner package page for all available options. Default is `sl_bart`.
base_screener	Deprecated for IL_conbin and IL_multiclass; kept for backward compatibility and currently ignored in those backends.
run_screening	Logical; if TRUE, run supervised screening on training data within each CV fold (and again on full training data before final fitting), then apply the same selected features to fold-validation or external validation data.
screen_pct	Percentage of features to retain during screening ((0,100]).
filter_method	Optional feature-filter method for non-survival runs. Supported values: 'prevalence' and 'variance'.
filter_pct	Optional retention percentage (in (0,100]) for the selected filter_method. Keeps the top filter_pct percent features.
prevalence_pct	Optional retention percentage (in (0,100]) for prevalence-based feature filtering before model fitting. Deprecated alias of filter_pct with filter_method = 'prevalence'.
meta_learner	Meta-learner for late fusion (stacked generalization). Defaults to `sl_nnls_auc`. For multiclass outcomes this legacy default maps to multinomial glmnet; use run_weighted for learned scalar layer weights. Check out the SuperLearner package page for all available options.
run_concat	Should early fusion be run? Default is TRUE. Uses the specified base_learner as the learning algorithm.
run_stacked	Should stacked model (late fusion) be run? Default is TRUE.
run_intermediate	Should standalone cooperative learning via multiview be run? Default is FALSE. When TRUE, IntegratedLearner runs only the cooperative multiview model and skips base learners, early fusion, and late fusion. Supported for continuous, binary, and survival outcomes; ignored with a message for multiclass outcomes.
cooperative_rho	Non-negative agreement-penalty values to evaluate for cooperative learning. A separate cv.multiview() fit is trained for each value, and the best value is selected by cross-validation.
cooperative_s	Penalty value used when predicting from the selected cooperative model. Defaults to "lambda.min".
cooperative_type_measure	Optional multiview CV loss/score. If NULL, an outcome-appropriate default is used.
drop_poor_performing_layers	Logical; if TRUE, layers with poor single-layer performance are removed from early and late fusion only. The thresholds are AUC < 0.5 for binary, R$^2$ < 0.5 for continuous, and C-index < 0.5 for survival outcomes. Single-layer outputs are still retained and reported.
verbose	logical; TRUE for SuperLearner printing progress (helpful for debugging). Default is FALSE.
print_learner	logical; Should a detailed summary be printed? Default is TRUE.
refit.stack	logical; For late fusion, post-refit predictions on the entire data are returned if specified. Default is FALSE.
family	Currently allows `gaussian()` for continuous, `binomial()` for binary, or an appropriate Cox/survival family for time-to-event outcomes. Determines whether IL_conbin or IL_survival is used internally.
run_majority	Logical; return multiclass majority voting. Default TRUE; ignored for other outcomes. Ties use the first class in training class order.
run_weighted	Logical; return multiclass probability averaging with nonnegative layer weights summing to one, learned by minimizing out-of-fold multiclass log loss. Default TRUE; ignored for other outcomes.
...	Additional arguments passed to the underlying engines.

Details

In a typical Bioconductor workflow, users construct assay-level data in SummarizedExperiment, combine those assays in MultiAssayExperiment, inspect the object with experiments(), assay(), colData(), and sampleMap(), and then pass the MAE into IntegratedLearner() for supervised multi-omics prediction. Internally, IntegratedLearner() converts the MultiAssayExperiment into tabular multi-view matrices and then calls IL_conbin for continuous/binary outcomes or IL_survival for time-to-event outcomes, depending on the specified family. Legacy PCL-style inputs remain available for backward compatibility, but the primary user-facing workflow is MAE-based.

Value

A list-like IntegratedLearner object containing the trained model fits (layer-specific, stacked, and/or concatenated models), cross-validated performance estimates, and predicted values for training and, if supplied, validation data. When run_intermediate = TRUE, the object also includes coefficient-based cooperative feature-importance scores from the selected multiview model. Multiclass fits also provide train_out and valid_out grouped into single, early, and late results. All enabled late methods (stacked, majority, weighted_average) include probabilities, predicted classes, and metrics; learned averaging also reports layer weights. Existing flat multiclass outputs remain available.

Author(s)

Himel Mallick, [email protected]

Examples

PRISM_MAE <- readRDS(system.file(
  "extdata", "PRISM_MAE.rds", package = "IntegratedLearner", mustWork = TRUE
))
mae <- PRISM_MAE
names(MultiAssayExperiment::experiments(mae))
head(as.data.frame(S4Vectors::DataFrame(MultiAssayExperiment::colData(mae)))[, 1:4])
head(MultiAssayExperiment::sampleMap(mae))
se <- MultiAssayExperiment::experiments(mae)[[1]]
SummarizedExperiment::assayNames(se)
fit <- IntegratedLearner(
  MAE_train = mae,
  experiment = names(MultiAssayExperiment::experiments(mae))[1:2],
  assay.type = rep("abundance", 2),
  folds = 2, base_learner = "SL.mean",
  run_stacked = FALSE, run_concat = FALSE, print_learner = FALSE,
  family = stats::binomial()
)
names(fit)

---

NNLS function to optimize weights of several base learners

Description

NNLS function to optimize weights of several base learners

Usage

NNLS(x, y, wt)

Arguments

x	x
y	y
wt	wt

Value

Solution of the quadratic programming problem

Examples

x <- cbind(c(0.1, 0.4, 0.6, 0.9), c(0.2, 0.5, 0.7, 0.8))
y <- c(0.15, 0.45, 0.65, 0.85)
fit <- NNLS(x = x, y = y, wt = rep(1, nrow(x)))
fit$solution

---

Plot the summary curves produced by an IntegratedLearner object

Description

Plots outcome-appropriate performance summaries for the training set and, if available, the validation set produced by an IntegratedLearner fit. Depending on the outcome family, this may include ROC curves, R-squared bar plots, multiclass one-vs-rest ROC curves, or survival AUC / Kaplan-Meier panels. Multiclass plots include all enabled single-layer and fusion outputs, including stacking, majority voting, and learned probability averaging. Majority ROC curves use vote fractions, which are not calibrated probabilities.

Usage

## S3 method for class 'learner'
plot(
  x,
  y = NULL,
  label_size = 8,
  label_x = 0.3,
  vjust = 0.1,
  rowwise_plot = TRUE,
  ...
)

Arguments

x	Fitted IntegratedLearner object.
y	Unused (required for S3 signature).
label_size	Optional numeric label size for subplot tags. Default is 8.
label_x	Optional single value or vector of x positions for subplot labels, relative to each subplot. Defaults to 0.3 for all labels.
vjust	Adjusts the vertical position of each label. More positive values move the label further down on the plot canvas. Can be a single value (applied to all labels) or a vector of values (one for each label). Default is 0.1.
rowwise_plot	If both train and test data are available, should the train and test plots be arranged row-wise? Default is TRUE. If FALSE, plots are aligned column-wise.
...	Additional arguments (currently unused).

Value

A list whose $plot entry is a ggplot2/cowplot composite object, along with the underlying tabular data used to generate the plot.

Examples

PRISM_MAE <- readRDS(system.file(
  "extdata", "PRISM_MAE.rds", package = "IntegratedLearner", mustWork = TRUE
))
mae <- PRISM_MAE
se <- MultiAssayExperiment::experiments(mae)[[1]]
head(SummarizedExperiment::assay(se, SummarizedExperiment::assayNames(se)[1]))
fit <- IntegratedLearner(
  MAE_train = mae,
  experiment = names(MultiAssayExperiment::experiments(mae))[1:2],
  assay.type = rep("abundance", 2),
  folds = 2, base_learner = "SL.mean",
  run_stacked = FALSE, run_concat = FALSE,
  print_learner = FALSE, family = stats::binomial()
)
plt <- plot(fit)
names(plt)

---

Make predictions using a trained 'IntegratedLearner' model

Description

This function makes predictions using a trained 'IntegratedLearner' model for new samples for which predictions are to be made

Usage

## S3 method for class 'learner'
predict(
  object,
  feature_table_valid = NULL,
  sample_metadata_valid = NULL,
  feature_metadata = NULL,
  outcome_col = NULL,
  subject_id_col = NULL,
  ...
)

Arguments

object	Fitted 'IntegratedLearner' object
feature_table_valid	Feature table from validation set. Must have the exact same structure as feature_table.
sample_metadata_valid	OPTIONAL (can provide feature_table_valid and not this): Sample-specific metadata table from independent validation set. If provided, it must have the exact same structure as sample_metadata.
feature_metadata	Matrix containing feature names and their corresponding layers. Must be same as that provided in IntegratedLearner object.
outcome_col	Optional outcome column name in sample_metadata_valid. If NULL, uses the mapping stored in object$column_map (or falls back to "Y").
subject_id_col	Optional subject identifier column name in sample_metadata_valid. If NULL, uses the mapping stored in object$column_map (or falls back to "subjectID").
...	Additional arguments (currently unused)

Value

For gaussian and binomial fits, a list with at least

yhat.test

A data frame or matrix of predictions for each single layer and, when requested during training, the stacked and concatenated fusion predictions.

Y_test

Observed validation outcomes, returned only when sample_metadata_valid is supplied.

AUC.test, accuracy.test, balanced_accuracy.test, metrics.test

Binary-outcome performance summaries, returned only when validation outcomes are supplied for binomial fits.

R2.test

Named vector of validation-set R$^2$ values, returned only when validation outcomes are supplied for gaussian fits.

For multiclass fits, returns the output of predict_multiclass.learner(), which contains class probabilities, predicted classes, and optional validation metrics. The flat outputs are prob.test, class.test, and yhat.test, with metrics.test when outcomes are available. Grouped results are returned in valid_out$single, valid_out$early, and valid_out$late. The latter includes all methods enabled during fitting: stacked, majority, and weighted_average, as listed in late_methods. Each fusion method includes valid_prob, valid_class, and a metrics table, plus valid_accuracy, valid_balanced_accuracy, valid_auc, and valid_logloss when labels are available. Without labels, metrics is NULL and scalar metrics are absent. Learned averaging also includes the training-derived layer weights; prediction does not refit them. Majority outputs are vote fractions; ties use training class order, and the fractions are not calibrated probabilities. Disabled early fusion returns early = NULL, and disabled late methods are omitted from late.

Examples

set.seed(1)
sample_ids <- paste0("S", seq_len(18))
layer1 <- matrix(rnorm(3 * length(sample_ids)),
  nrow = 3,
  dimnames = list(paste0("L1_F", 1:3), sample_ids)
)
layer2 <- matrix(rnorm(2 * length(sample_ids)),
  nrow = 2,
  dimnames = list(paste0("L2_F", 1:2), sample_ids)
)
signal <- colMeans(layer1)
sample_metadata <- data.frame(
  Y = as.numeric(signal > stats::median(signal)),
  subjectID = sample_ids,
  row.names = sample_ids
)
feature_table <- as.data.frame(rbind(layer1, layer2))
feature_metadata <- data.frame(
  featureID = rownames(feature_table),
  featureType = c(rep("Layer1", 3), rep("Layer2", 2)),
  row.names = rownames(feature_table)
)
fit <- IntegratedLearner(
  PCL_train = list(
    feature_table = feature_table,
    sample_metadata = sample_metadata,
    feature_metadata = feature_metadata
  ),
  folds = 2, base_learner = "SL.mean",
  run_stacked = FALSE, run_concat = FALSE,
  print_learner = FALSE, family = stats::binomial()
)
pred <- predict(
  fit,
  feature_table_valid = feature_table,
  sample_metadata_valid = sample_metadata,
  feature_metadata = feature_metadata
)
names(pred)

---

Predict function for sl_bart

Description

Predict function for sl_bart

Usage

## S3 method for class 'sl_bart'
predict(object, newdata, family, X = NULL, Y = NULL, ...)

Arguments

object	Fitted sl_bart model object
newdata	Data frame for prediction
family	Family object passed through (unused)
X	Training design matrix (unused)
Y	Training outcome (unused)
...	Additional arguments (unused)

Value

Prediction from the sl_bart model

Examples

if (
  requireNamespace("bartMachine", quietly = TRUE) &&
  nzchar(Sys.which("java")) &&
  any(grepl("jdk.incubator.vector", getOption("java.parameters", character()), fixed = TRUE))
) {
  set.seed(1)
  X <- data.frame(x1 = rnorm(20), x2 = rnorm(20))
  Y <- rnorm(20)
  fit <- sl_bart(
    Y = Y, X = X, newX = X[1:3, ],
    family = stats::gaussian(),
    obsWeights = rep(1, nrow(X)), id = seq_len(nrow(X)),
    num_trees = 5, num_burn_in = 5, num_iterations_after_burn_in = 20
  )
  head(predict(fit$fit, newdata = X[1:3, ], family = stats::gaussian()))
}

---

Predict function for sl_nnls_auc

Description

Predict function for sl_nnls_auc

Usage

## S3 method for class 'sl_nnls_auc'
predict(object, newdata, ...)

Arguments

object	Fitted sl_nnls_auc model
newdata	Validation layer-level predictions
...	Additional arguments (unused)

Value

Prediction from the meta-learner

Examples

set.seed(1)
X <- data.frame(m1 = runif(20), m2 = runif(20))
Y <- rnorm(20)
fit <- sl_nnls_auc(
  Y = Y, X = X, newX = X[1:4, ],
  family = stats::gaussian(), obsWeights = rep(1, nrow(X))
)
head(predict(fit$fit, newdata = X[1:4, ]))

---

Wrapper for bartMachine learner

Description

Support bayesian additive regression trees via the bartMachine package.

Usage

sl_bart(
  Y,
  X,
  newX,
  family,
  obsWeights,
  id,
  num_trees = 50,
  num_burn_in = 250,
  verbose = FALSE,
  alpha = 0.95,
  beta = 2,
  k = 2,
  q = 0.9,
  nu = 3,
  num_iterations_after_burn_in = 1000,
  serialize = TRUE,
  seed = 5678,
  ...
)

Arguments

Y	Outcome variable
X	Covariate dataframe
newX	Optional dataframe to predict the outcome
family	'gaussian' for regression, 'binomial' for binary classification
obsWeights	Optional observation-level weights (supported but not tested)
id	Optional id to group observations from the same unit (not used currently).
num_trees	The number of trees to be grown in the sum-of-trees model.
num_burn_in	Number of MCMC samples to be discarded as 'burn-in'.
verbose	Prints information about progress of the algorithm to the screen.
alpha	Base hyperparameter in tree prior for whether a node is nonterminal or not.
beta	Power hyperparameter in tree prior for whether a node is nonterminal or not.
k	For regression, k determines the prior probability that E(Y|X) is contained in the interval $(y_{min}, y_{max})$, based on a normal distribution. For example, when k=2, the prior probability is 95 percent. For classification, k determines the prior probability that E(Y|X) is between (-3,3). Note that a larger value of k results in more shrinkage and a more conservative fit.
q	Quantile of the prior on the error variance at which the data-based estimate is placed. Note that the larger the value of q, the more aggressive the fit as you are placing more prior weight on values lower than the data-based estimate. Not used for classification.
nu	Degrees of freedom for the inverse chi^2 prior. Not used for classification.
num_iterations_after_burn_in	Number of MCMC samples to draw from the posterior distribution of f(x).
serialize	If TRUE, bartMachine results can be saved to a file, but will require additional RAM.
seed	Seed for reproducibility passed to bartMachine.
...	Additional arguments (not used)

Value

A list with elements pred (predictions for newX) and fit (the fitted model object).

Examples

if (
  requireNamespace("bartMachine", quietly = TRUE) &&
  nzchar(Sys.which("java")) &&
  any(grepl("jdk.incubator.vector", getOption("java.parameters", character()), fixed = TRUE))
) {
  set.seed(1)
  X <- data.frame(x1 = rnorm(20), x2 = rnorm(20))
  Y <- rnorm(20)
  fit <- sl_bart(
    Y = Y, X = X, newX = X[1:3, ],
    family = stats::gaussian(),
    obsWeights = rep(1, nrow(X)), id = seq_len(nrow(X)),
    num_trees = 5, num_burn_in = 5, num_iterations_after_burn_in = 20
  )
  fit$pred
}

---

Elastic net regression with optimized alpha and lambda

Description

Selects alpha by cross-validation, then uses the shared glmnet fitting implementation.

Usage

sl_enet(
  Y,
  X,
  newX,
  family,
  obsWeights,
  id,
  alpha = seq(0, 1, 0.1),
  nfolds = 10,
  nlambda = 100,
  useMin = TRUE,
  loss = "deviance",
  ...
)

Arguments

Y	Outcome variable
X	Covariate dataframe
newX	Dataframe to predict the outcome
family	'gaussian' for regression, 'binomial' for binary classification. Untested options: 'multinomial' for multiple classification or 'mgaussian' for multiple response, 'poisson' for non-negative outcome with proportional mean and variance, 'cox'.
obsWeights	Optional observation-level weights
id	Optional id to group observations from the same unit (not used currently).
alpha	Numeric grid of elastic net mixing parameters between 0 and 1.
nfolds	Number of folds for internal cross-validation to optimize lambda.
nlambda	Number of lambda values to check, recommended to be 100 or more.
useMin	If TRUE use lambda that minimizes risk, otherwise use 1 standard-error rule which chooses a higher penalty with performance within one standard error of the minimum (see Breiman et al. 1984 on CART for background).
loss	Loss function, can be 'deviance', 'mse', or 'mae'. If family = binomial can also be 'auc' or 'class' (misclassification error).
...	Any additional arguments are passed through to cv.glmnet.

Value

A list with elements pred (predictions for newX) and fit (cross-validated glmnet fit metadata).

References

Friedman, J., Hastie, T., & Tibshirani, R. (2010). Regularization paths for generalized linear models via coordinate descent. Journal of statistical software, 33(1), 1.

Hoerl, A. E., & Kennard, R. W. (1970). Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1), 55-67.

Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society. Series B (Methodological), 267-288.

Zou, H., & Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 67(2), 301-320.

See Also

predict.SL.glmnet cv.glmnet glmnet

Examples

set.seed(1)
X <- data.frame(x1 = rnorm(20), x2 = rnorm(20))
Y <- rnorm(20)
fit <- sl_enet(
  Y = Y, X = X, newX = X[1:3, ], family = stats::gaussian(),
  obsWeights = rep(1, nrow(X)), id = seq_len(nrow(X)),
  alpha = c(0, 0.5, 1), nfolds = 3, nlambda = 10
)
head(fit$pred)

---

Elastic net regression, including lasso and ridge with a fixed alpha

Description

Penalized regression using elastic net. Alpha = 0 corresponds to ridge regression and alpha = 1 corresponds to Lasso.

See vignette('glmnet_beta', package = 'glmnet') for a nice tutorial on glmnet.

Usage

sl_glmnet2(
  Y,
  X,
  newX,
  family,
  obsWeights,
  id,
  alpha = 0.5,
  nfolds = 10,
  nlambda = 100,
  useMin = TRUE,
  loss = "deviance",
  ...
)

Arguments

Y	Outcome variable
X	Covariate dataframe
newX	Dataframe to predict the outcome
family	'gaussian' for regression, 'binomial' for binary classification. Untested options: 'multinomial' for multiple classification or 'mgaussian' for multiple response, 'poisson' for non-negative outcome with proportional mean and variance, 'cox'.
obsWeights	Optional observation-level weights
id	Optional id to group observations from the same unit (not used currently).
alpha	Elastic net mixing parameter, range 0 to 1. 0 = ridge regression and 1 = lasso.
nfolds	Number of folds for internal cross-validation to optimize lambda.
nlambda	Number of lambda values to check, recommended to be 100 or more.
useMin	If TRUE use lambda that minimizes risk, otherwise use 1 standard-error rule which chooses a higher penalty with performance within one standard error of the minimum (see Breiman et al. 1984 on CART for background).
loss	Loss function, can be 'deviance', 'mse', or 'mae'. If family = binomial can also be 'auc' or 'class' (misclassification error).
...	Any additional arguments are passed through to cv.glmnet.

Value

A list with elements pred (predictions for newX) and fit (cross-validated glmnet fit metadata).

References

Friedman, J., Hastie, T., & Tibshirani, R. (2010). Regularization paths for generalized linear models via coordinate descent. Journal of statistical software, 33(1), 1.

Hoerl, A. E., & Kennard, R. W. (1970). Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1), 55-67.

Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society. Series B (Methodological), 267-288.

Zou, H., & Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 67(2), 301-320.

See Also

predict.SL.glmnet cv.glmnet glmnet

Examples

set.seed(1)
X <- data.frame(x1 = rnorm(20), x2 = rnorm(20))
Y <- rnorm(20)
fit <- sl_glmnet2(
  Y = Y, X = X, newX = X[1:3, ], family = stats::gaussian(),
  obsWeights = rep(1, nrow(X)), id = seq_len(nrow(X)),
  nfolds = 3, nlambda = 10
)
head(fit$pred)

---

Horseshoe regression

Description

Horseshoe regression

Usage

sl_horseshoe(
  Y,
  X,
  newX,
  family,
  prior = "horseshoe",
  N = 20000L,
  burnin = 1000L,
  thinning = 1L,
  ...
)

Arguments

Y	Outcome variable
X	Covariate data frame
newX	Dataframe to predict the outcome
family	'gaussian' for regression, 'binomial' for binary classification. Untested options: 'poisson' for for integer or count data
prior	prior for regression coefficients to use. 'Horseshoe' by default. Untested options: ridge regression (prior='rr' or prior='ridge'), lasso regression (prior='lasso') and horseshoe+ regression (prior='hs+' or prior='horseshoe+')
N	Number of posterior samples to generate.
burnin	Number of burn-in samples.
thinning	Desired level of thinning.
...	other parameters passed to bayesreg function

Value

A list with elements pred (predictions for newX) and fit (the fitted Bayesian regression object).

Examples

if (
  requireNamespace("bayesreg", quietly = TRUE) &&
  identical(Sys.getenv("INTEGRATEDLEARNER_RUN_BAYESREG_EXAMPLES"), "true")
) {
  set.seed(1)
  X <- data.frame(x1 = rnorm(20), x2 = rnorm(20))
  Y <- rnorm(20)
  fit <- sl_horseshoe(
    Y = Y, X = X, newX = X[1:3, ],
    family = stats::gaussian()
  )
  fit$pred
}

---

Lasso regression with a fixed alpha

Description

Uses the shared glmnet fitting implementation with alpha = 1 by default.

Usage

sl_lasso(
  Y,
  X,
  newX,
  family,
  obsWeights,
  id,
  alpha = 1,
  nfolds = 10,
  nlambda = 100,
  useMin = TRUE,
  loss = "deviance",
  ...
)

Arguments

Y	Outcome variable
X	Covariate dataframe
newX	Dataframe to predict the outcome
family	'gaussian' for regression, 'binomial' for binary classification. Untested options: 'multinomial' for multiple classification or 'mgaussian' for multiple response, 'poisson' for non-negative outcome with proportional mean and variance, 'cox'.
obsWeights	Optional observation-level weights
id	Optional id to group observations from the same unit (not used currently).
alpha	Elastic net mixing parameter, range 0 to 1. 0 = ridge regression and 1 = lasso.
nfolds	Number of folds for internal cross-validation to optimize lambda.
nlambda	Number of lambda values to check, recommended to be 100 or more.
useMin	If TRUE use lambda that minimizes risk, otherwise use 1 standard-error rule which chooses a higher penalty with performance within one standard error of the minimum (see Breiman et al. 1984 on CART for background).
loss	Loss function, can be 'deviance', 'mse', or 'mae'. If family = binomial can also be 'auc' or 'class' (misclassification error).
...	Any additional arguments are passed through to cv.glmnet.

Value

A list with elements pred (predictions for newX) and fit (cross-validated glmnet fit metadata).

References

Friedman, J., Hastie, T., & Tibshirani, R. (2010). Regularization paths for generalized linear models via coordinate descent. Journal of statistical software, 33(1), 1.

Hoerl, A. E., & Kennard, R. W. (1970). Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1), 55-67.

Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society. Series B (Methodological), 267-288.

Zou, H., & Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 67(2), 301-320.

See Also

predict.SL.glmnet cv.glmnet glmnet

Examples

set.seed(1)
X <- data.frame(x1 = rnorm(20), x2 = rnorm(20))
linpred <- 0.5 * X$x1 - 0.25 * X$x2
Y <- stats::rbinom(nrow(X), 1, stats::plogis(linpred))
fit <- sl_lasso(
  Y = Y, X = X, newX = X[1:3, ], family = stats::binomial(),
  obsWeights = rep(1, nrow(X)), id = seq_len(nrow(X)),
  nfolds = 3, nlambda = 10
)
head(fit$pred)

---

Combined SuperLearner function for both NNLS/AUC maximization

Description

Combined SuperLearner function for both NNLS/AUC maximization

Usage

sl_nnls_auc(Y, X, newX, family, obsWeights, bounds = c(0, Inf), ...)

Arguments

Y	Outcome matrix from metalearner
X	Layer-level predictions used to train the metalearner
newX	Layer-level predictions for validation data
family	Family object
obsWeights	Observation weights
bounds	Lower/upper bounds for weights (binomial case)
...	Additional arguments passed through

Value

Estimated meta-learner coefficients and predictions

Examples

set.seed(1)
X <- data.frame(m1 = runif(20), m2 = runif(20))
Y <- rnorm(20)
fit <- sl_nnls_auc(
  Y = Y, X = X, newX = X[1:4, ],
  family = stats::gaussian(), obsWeights = rep(1, nrow(X))
)
head(fit$pred)

---

Update IntegratedLearner fit object based on layers available in the test set

Description

Allow update of IntegratedLearner if only a subset of omics layers are available in test set. If all layers and features match, it calls predict.learner()

Usage

## S3 method for class 'learner'
update(
  object,
  feature_table_valid,
  sample_metadata_valid = NULL,
  feature_metadata_valid,
  outcome_col = NULL,
  subject_id_col = NULL,
  seed = 1234,
  verbose = FALSE,
  ...
)

Arguments

object	fitted 'IntegratedLearner' object
feature_table_valid	Feature table from validation set. It should be a data frame with features in rows and samples in columns. Feature names should be a subset of training data feature names.
sample_metadata_valid	OPTIONAL (can provide feature_table_valid and not this): Sample-specific metadata table from independent validation set. If provided, it must have the exact same structure as sample_metadata. Default is NULL.
feature_metadata_valid	Matrix containing feature names and their corresponding layers. Must be subset of feature_metadata provided in IntegratedLearner object.
outcome_col	Optional outcome column name in sample_metadata_valid. If NULL, uses object$column_map$outcome_col (or "Y").
subject_id_col	Optional subject ID column name in sample_metadata_valid. If NULL, uses object$column_map$subject_id_col (or "subjectID").
seed	Seed for reproducibility. Default is 1234.
verbose	Should a summary of fits/ results be printed. Default is FALSE
...	Additional arguments (unused)

Value

An updated IntegratedLearner prediction object with yhat.test and, when validation outcomes are supplied, the observed outcomes and corresponding performance summaries.

Examples

set.seed(1)
sample_ids <- paste0("S", seq_len(18))
layer1 <- matrix(rnorm(3 * length(sample_ids)),
  nrow = 3,
  dimnames = list(paste0("L1_F", 1:3), sample_ids)
)
layer2 <- matrix(rnorm(2 * length(sample_ids)),
  nrow = 2,
  dimnames = list(paste0("L2_F", 1:2), sample_ids)
)
signal <- colMeans(layer1)
sample_metadata_valid <- data.frame(
  Y = as.numeric(signal > stats::median(signal)),
  subjectID = sample_ids,
  row.names = sample_ids
)
feature_table_valid <- as.data.frame(rbind(layer1, layer2))
feature_metadata_valid <- data.frame(
  featureID = rownames(feature_table_valid),
  featureType = c(rep("Layer1", 3), rep("Layer2", 2)),
  row.names = rownames(feature_table_valid)
)
fit <- IntegratedLearner(
  PCL_train = list(
    feature_table = feature_table_valid,
    sample_metadata = sample_metadata_valid,
    feature_metadata = feature_metadata_valid
  ),
  folds = 2, base_learner = "SL.mean",
  run_stacked = FALSE, run_concat = FALSE,
  print_learner = FALSE, family = stats::binomial()
)
upd <- update(
  object = fit,
  feature_table_valid = feature_table_valid,
  sample_metadata_valid = sample_metadata_valid,
  feature_metadata_valid = feature_metadata_valid
)
names(upd)

---