This package contains code to reproduce findings in the DeSurv paper: A survival-driven deconvolution tool to prognostic subtype discovery

To run this code you will need to install the R targets package. Targets is a pipeline management tool that will run DeSurv from original data to final figures.

install.packages("targets")

The pipeline now depends on the development version of the DeSurv package located one directory up in ../DeSurv. Install or update it before running any targets to pick up the latest C++ and R functionality:

# install.packages("devtools")
devtools::install_local("../DeSurv", upgrade = "never", force = TRUE)

Each _targets*.R script will automatically pkgload::load_all("../DeSurv") when that directory is present, so the local checkout is always used.

Hyperparameter tuning is handled inside the DeSurv package via desurv_cv_bayesopt_refine(), which launches an initial Bayesian optimisation run, shrinks the promising region, and iterates until improvements plateau instead of relying on custom grid helpers in this repository. All preprocessing, validation filtering, and model-selection logic now flow through the package interface.

The pipeline configuration now lives in targets_configs.R as three named lists: targets_bo_configs() for Bayesian optimisation settings, targets_run_configs() for full-model runs, and targets_val_configs() for validation settings. Each config gets a hash-based config_id and a readable path_tag (label + short hash) that is included in output paths.

Config validation runs during pipeline setup (see validate_desurv_configs() in R/targets_config.R) so mismatched bo_key/run_key values and missing required fields fail fast. You can also compare two configs with desurv_config_diff(old, new, type = c("bo", "run", "val")) to see which hash-driving fields changed.

Common entries you will likely edit are:

• train_datasets controls the training dataset set in BO configs. Validation uses val_datasets when mode = "external" and can also use mode = "train_split" to validate on a train/test split from the training dataset (see bladder configs for an example).
• ngene_config, which controls the gene-filter size passed into preprocessing. Provide a single value (e.g., c(2000)) to fix the number of genes or a range (e.g., c(2000, 5000)) to let Bayesian optimisation tune ngene alongside the other hyperparameters.
• lambdaw_config/lambdah_config, which control whether the factor penalty terms are tuned. Keep them as singletons (e.g., c(0)) to fix the penalties at that value or provide positive ranges (for example c(1e-5, 1e-2)) to let BO search on a log-scale.

Optional tuning parameters

The configuration constants above feed directly into the Bayesian optimiser. If you set a vector with more than one unique value, the pipeline will add the appropriate bound to desurv_cv_bayesopt_refine(); otherwise the single value is passed as a fixed _grid argument. The helper maybe_add_numeric_bound() (see R/bo_helpers.R) handles this logic and is covered by the unit tests under tests/testthat.

The shared desurv_bo_bounds list defines the non-dataset-specific search space for the optimiser (k, alpha, lambda, nu). Adjust these entries if the default range does not match your experiment; the pipeline will merge them with any ngene_config/lambdaw_config/lambdah_config ranges automatically before calling desurv_cv_bayesopt_refine(). Additional refinement controls live next to the other config entries in targets_configs.R: bo_max_refinements determines how many shrinks are attempted, while bo_coarse_control and bo_refine_control describe the per-stage BO settings (e.g., n_init, n_iter, candidate_pool, seed, cv_verbose). Update them if you need a longer coarse search, different seeds, or a quieter optimiser.

Validation output folders are now nested under results/.../validation/<val_config_id>/ so you can run multiple validation configurations against the same training run without overwriting earlier results.

Full-model runs in targets_run_configs() reference a BO configuration via bo_key. This lets you reuse the same Bayesian optimisation results while changing full-run settings like ninit_full, ntop_config, or convergence controls without retriggering BO. The ntop_config setting now lives with the run configs so you can iterate on downstream thresholds independently of BO.

Optional Enrichment Dependencies

The ORA/KEGG helpers in R/ora_analysis.R depend on both clusterProfiler and org.Hs.eg.db. Make sure those packages are installed and available when running any targets that trigger enrichment; otherwise the pipeline will stop early with a clear error message.

Running tests

Basic automated checks live under tests/testthat. To execute them:

Rscript -e 'testthat::test_dir("tests/testthat")'