OmicSelector 2.0

Rigorous biomarker discovery from high-dimensional omics data with zero data leakage.

Overview

OmicSelector is an R package for biomarker discovery that enforces methodologically sound machine learning practices. Built on the mlr3 ecosystem, it guarantees:

Zero Data Leakage: Feature selection occurs strictly inside cross-validation folds
Proper Nested CV: Separation of outer (evaluation) and inner (selection) loops
Feature Stability: Nogueira Stability Index for reproducible biomarker sets
Multi-Objective Selection: Balance performance, stability, and parsimony

Installation

remotes::install_github("kstawiski/OmicSelector")

Quick Start

library(OmicSelector)

# Create pipeline from your data
pipeline <- OmicPipeline$new(
  data = my_data,          # data.frame with features + target
  target = "outcome",      # target column name
  positive = "Case"        # positive class for AUC
)

# Create learner with embedded feature selection
learner <- pipeline$create_graph_learner(
  filter = "anova",        # Feature selection: anova, mrmr, variance, correlation
  model = "ranger",        # Model: ranger, glmnet, svm, log_reg
  n_features = 20          # Number of features to select
)

# Run nested cross-validation
benchmark <- BenchmarkService$new(
  task = pipeline,
  outer_folds = 5,
  inner_folds = 3,
  seed = 42
)
benchmark$add_learner(learner)
result <- benchmark$run()

# Analyze stability and select best signature
stability <- compute_stability_from_resample(result$benchmark_result)
best <- select_best_signature(result, mode = "weighted")

Data Format

Your data should be a data.frame with: - Feature columns: Numeric values (gene expression, miRNA counts, etc.) - Target column: Factor/character (classification) or numeric (regression)

# Example structure:
#   gene_A  gene_B  gene_C  outcome
# 1   2.34    1.56    3.21     Case
# 2   1.12    2.89    0.45  Control

Configuration Options

Feature Selection Methods

Method	Code	Best For
ANOVA F-test	`"anova"`	Default, continuous features
Kruskal-Wallis	`"kruskal"`	Non-normal distributions
Chi-Squared	`"chi_squared"`	Categorical features
Variance	`"variance"`	Pre-filtering
Correlation	`"correlation"`	Quick univariate
Information Gain	`"information_gain"`	Mixed feature types
Gain Ratio	`"gain_ratio"`	Avoiding cardinality bias
mRMR	`"mrmr"`	Reducing redundancy
CMIM/JMIM/JMI	`"cmim"`, `"jmim"`, `"jmi"`	Feature interactions
AUC	`"auc"`	Classification performance
Relief	`"relief"`	Detecting interactions
RF Importance	`"importance"`	Non-linear relationships
Permutation	`"permutation"`	Model-agnostic

Classification Models

Model	Code	Strengths
Random Forest	`"ranger"`	Handles interactions, robust
XGBoost	`"xgboost"`	High performance, handles missing values
LightGBM	`"lightgbm"`	Very fast, memory efficient
Elastic Net	`"glmnet"`	Interpretable coefficients
SVM	`"svm"`	High-dimensional data
Logistic Regression	`"log_reg"`	Baseline, interpretable
k-NN	`"kknn"`	Non-parametric
Naive Bayes	`"naive_bayes"`	Fast, small data
LDA/QDA	`"lda"`, `"qda"`	Dimensionality reduction
Neural Net	`"nnet"`	Non-linear relationships
Decision Tree	`"rpart"`	Interpretable

Key Modules

Module	Description
OmicPipeline	Build mlr3 graphs with preprocessing
BenchmarkService	Nested CV with zero leakage
select_best_signature	Multi-objective signature selection
compute_nogueira_stability	Feature selection stability metrics
FrozenComBat	Batch correction with frozen parameters
fit_platt_scaling	Probability calibration
MultiOmicsStacker	Late integration of multi-omics data

Phase 5: Advanced Features

Module	Description
FilterGOF_KS / FilterHurdle	GOF filters for sparse/zero-inflated data
xai_pipeline	DALEX-based interpretability with correlation warnings
create_stability_ensemble	Bootstrap stability for reproducible biomarkers
make_autotuner_glmnet	Bayesian hyperparameter optimization
smote_augment	SMOTE for class imbalance (inside CV)

Docker

docker build -f Dockerfile.core -t omicselector:2.0 .
docker run -it --rm -v $(pwd):/workspace omicselector:2.0 R

Citation

@article{stawiski2022omicselector,
  title={OmicSelector: automatic feature selection and deep learning
         modeling for omic experiments},
  author={Stawiski, Konrad and Kaszkowiak, Marcin and Mikulski, Damian and others},
  journal={bioRxiv},
  year={2022},
  doi={10.1101/2022.06.01.494299}
}

Authors

Konrad Stawiski, M.D., Ph.D. (konrad.stawiski@umed.lodz.pl)
Marcin Kaszkowiak, M.D.
Damian Mikulski, M.D.

Supervised by: Prof. Wojciech Fendler, M.D., Ph.D.

Department of Biostatistics and Translational Medicine, Medical University of Lodz, Poland