pdac-efs-bench2024

This repository contains code and data for benchmarking multi-omics feature selection methods in pancreatic ductal adenocarcinoma (PDAC), with a focus on survival prediction.

Citation

Zobolas, J., George, A.-M., López, A., Fischer, S., Becker, M., & Aittokallio, T. (2025). Optimizing Prognostic Biomarker Discovery in Pancreatic Cancer Through Hybrid Ensemble Feature Selection and Multi-Omics Data. https://arxiv.org/pdf/2509.02648

Environment Setup

Restore R libraries required to run the benchmark using renv:

renv::restore(exclude = c("BiocGenerics", "BiocManager", "BiocVersion", "ComplexHeatmap", "IRanges", "S4Vectors", "MASS", "ipred", "class"))

Data and Preprocessing

We use 3 multi-omics PDAC datasets. Preprocessing steps include:

• Patient and feature filtering
• Metadata curation
• Resampling definition for benchmark
• Creation of mlr3 survival tasks per omic data type

All related code and metadata are found in the data/ directory.

Multi-omics Benchmark Overview

The benchmarking pipeline consists of three main steps, located in the bench/ directory.

Hybrid Ensemble Feature Selection (hEFS)

Script: bench/run_efs.sh: Runs the Ensemble Feature Selection procedure (bench/efs.R) across:

• All datasets
• All omics
• All resampling iterations

Stores results as EnsembleFSResult objects in bench/efs/. Contact the main author to share with you these intermediate results.

Omic-wise Feature Selection

Script: bench/run_fs.sh: Performs feature selection per omic and per subsampling iteration (100 total).

Two methods available:

• Cox Lasso
• Pre-computed Ensemble Feature Selection (loaded from step 1). We automatically select the number of features via the Pareto front method.

Output: bench/fs.rds - a table with:

• dataset_id, omic_id, rsmp_id
• Selected features for each method

Multi-omics Integration & Benchmarking

Script: bench/run_mm_bench.R: Combines selected features across omics (via late integration/fusion) per subsampling iteration, then trains and evaluates survival models on training/test splits. Available models are Cox Proportional Hazards, Cox Lasso, Random Survival Forests and BlockForest.

Output: bench/result.rds - a table with:

• dataset_id: Identifier of the dataset used
• fs_method_id: Feature selection method applied
• rsmp_id: Subsampling (resampling) iteration identifier
• model_data_config: Configuration used for model training, indicating the model type and which omics and/or clinical data were included (all means clinical + all omics)
• task_nfeats: Number of selected features used in the task
• task_feats: The specific features selected
• Performance scores for the test sets (Harrell's C-index, etc.)