Survival NPDR

Nearest-Neighbor Projected Distance Regression for Biomarker Discovery in Gene-Expression Survival Data

Patrick Kampmeyer, Bryan A. Dawkins, Ryan J. Urbanowicz, Kia Kazemi-Nia, Brett A. McKinney

Survival NPDR extends NPDR to right-censored survival outcomes. It identifies prognostic biomarkers in high-dimensional gene-expression data by regressing pairwise survival-time differences on pairwise attribute differences within a censor-aware nearest-neighbor neighborhood.

---

Repository Structure

survivalNPDR/
├── scripts/
│   ├── simulation/               # Figs 2 & 3, Tables 1 & 2
│   │   ├── main_effect_size_sweep_rsurv.r        # Main-effect benchmark + effect-size sweep
│   │   └── interaction_effect_size_sweep_rsurv.r # Interaction benchmark + effect-size sweep
│   │
│   ├── test_k/                   # Fig 4 — neighborhood size study
│   │   ├── test_k_lognormal.R   # runs both main + interaction k sweeps
│   │   └── plot_k_combined_lognormal.R
│   │
│   ├── real_data/                # Fig 5 and Tables 3 & 4
│   │   ├── preprocess_breast_cancer.R  # Download + preprocess GEO breast datasets
│   │   ├── ovarian_risk_stratification.R                   # Ovarian cancer sNPDR + KM risk stratification
│   │   ├── breast_model_performance.R         # Performance metrics (C-index, logrank, HR) for GSE2034
│   │   └── breast_survival_curve.R         # KM descriptive figure for GSE9893
│   │
│   ├── network/                  # Figs 6 & 7 — interaction network analysis
│   │   └── ovarian_cancer_network.r
│   │
│   ├── figures/                  # Assemble publication figures
│   │   └── plot_ef_vs_effect_size.R
│   │
│   └── analysis_statistical.R   # Friedman + Wilcoxon significance tests
│
├── data/                         # Processed real datasets
│   └── GSE{9893,2034,2990,9891,32062,13876,17260}/
│
├── results/                      # Pre-computed paper result tables (CSV)
└── figures/                      # Pre-computed paper figures (PNG)

---

Installation

1. Install R dependencies

install.packages(c(
  "survival", "dplyr", "rlang", "PRROC", "glmnet",
  "caret", "survcomp", "survivalsvm", "ranger",
  "R.utils", "ggplot2", "tibble", "rsurv", "parallel"
))

# Bioconductor (for real-data preprocessing)
if (!requireNamespace("BiocManager", quietly = TRUE))
  install.packages("BiocManager")
BiocManager::install(c("Biobase", "GEOquery", "curatedOvarianData"))

2. Install the npdr package

# With devtools
devtools::install_github("insilico/npdr")

# Or with pak
pak::pak("insilico/npdr")

---

Quick Start

Your data should be a data.frame with a time column, a status column (1 = event, 0 = censored), and all other columns as features.

library(npdr)

# ── Main-effect detection — signed (asymmetric), adaptive neighborhood ─────────
res <- surv_npdr(
  outcome   = c(time_var = "time", status_var = "status"),
  dataset   = my_data,
  diff.type = "signed"       # signed ΔT ~ signed ΔX
)
head(res[order(res$pval.att), ])

# ── Interaction detection — absolute (symmetric), fixed k = 20 ────────────────
res <- surv_npdr(
  outcome    = c(time_var = "time", status_var = "status"),
  dataset    = my_data,
  diff.type  = "absolute",   # |ΔT| ~ |ΔX|
  nbd.method = "relieff",
  knn        = 20
)

# ── With KM imputation and covariate adjustment ────────────────────────────────
res <- surv_npdr(
  outcome    = c(time_var = "time", status_var = "status"),
  dataset    = my_data,
  diff.type  = "signed",
  km.impute  = TRUE,
  covariates = c("age", "sex", "BMI")
)

# ── Logistic regression with LASSO regularization ─────────────────────────────
res <- surv_npdr(
  outcome    = c(time_var = "time", status_var = "status"),
  dataset    = my_data,
  regression = "binomial",
  regularize = TRUE,
  alpha      = 1             # 1 = LASSO, 0 = ridge
)

Visualize top features

library(ggplot2)

top <- res |> dplyr::arrange(pval.att) |> head(20)

ggplot(top, aes(x = reorder(att, -abs(beta.raw.att)), y = beta.raw.att)) +
  geom_col(fill = "steelblue") +
  coord_flip() +
  labs(x = "Feature", y = "Survival NPDR coefficient",
       title = "Top 20 survival-associated features")

---

Reproducing Paper Results

All scripts use root <- getwd(). Run from the repository root:

cd /path/to/survivalNPDR

Figures 2 & 3 / Tables 1 & 2 — Simulation benchmarks

Each sweep script runs all comparison methods (Cox, LASSO-Cox, survival SVM, RSF, sNPDR-signed, sNPDR-absolute) across a range of effect sizes in parallel.

# Main-effect sweep (default beta_main = 0.00, 0.07, 0.13, 0.20, 0.27, 0.33, 0.40)
Rscript scripts/simulation/main_effect_size_sweep_rsurv.r

# Table 1: main effects at beta_main = 0.13
BETA_MAIN_SWEEP="0.13" Rscript scripts/simulation/main_effect_size_sweep_rsurv.r

# Interaction-effect sweep (default beta_int = 0.00, 0.03, 0.07, 0.10, 0.13, 0.17, 0.20)
Rscript scripts/simulation/interaction_effect_size_sweep_rsurv.r

# Table 2: interaction effects at beta_int = 0.20
BETA_INT_SWEEP="0.20" Rscript scripts/simulation/interaction_effect_size_sweep_rsurv.r

# Assemble Figures 2 and 3
Rscript scripts/figures/plot_ef_vs_effect_size.R

Figure 4 — Choice of neighborhood size k

# Both sweeps run in parallel (uses all available cores − 2)
Rscript scripts/test_k/test_k_lognormal.R   # runs both main + interaction k sweeps
Rscript scripts/test_k/plot_k_combined_lognormal.R 0.5 100
# → figures/auPRC_K_combined_lognormal_no_multisurf_sigma0.50_p100.png

Figure 5 / Tables 3 & 4 — Real cancer dataset benchmarks

# Breast cancer: download and preprocess from GEO (requires GEOquery)
Rscript scripts/real_data/preprocess_breast_cancer.R

# Ovarian cancer sNPDR + KM risk stratification (Table 4)
Rscript scripts/real_data/ovarian_risk_stratification.R

# Breast cancer performance metrics (Table 3)
Rscript scripts/real_data/breast_model_performance.R

Figure 5 — Kaplan-Meier risk stratification

Rscript scripts/real_data/breast_survival_curve.R

Figures 6 & 7 — Interaction networks (ReGAIN)

# Figure 6: breast cancer interaction network (GSE9893)
# Generated from the breast-cancer analysis workflow.

# Figure 7: ovarian ReGAIN interaction network (GSE9891)
Rscript scripts/network/ovarian_cancer_network.r

Statistical significance tests

Rscript scripts/analysis_statistical.R

---

Data

Processed expression matrices and clinical annotations are in data/:

Dataset	Cancer type	Samples	Notes
GSE9893	Breast	155	ERα+
GSE2034	Breast	286	Lymph-node negative
GSE2990	Breast	189	Mixed
GSE9891	Ovarian	260	Stage III/IV
GSE32062	Ovarian	260	High-grade serous
GSE13876	Ovarian	157	High-grade serous
GSE17260	Ovarian	110	High-grade serous

Raw data can be re-downloaded from GEO using scripts/real_data/preprocess_breast_cancer.R (breast) or curatedOvarianData (ovarian).

---

Pre-computed Results

results/ contains pre-computed CSV tables from the paper. figures/ contains the corresponding plots. Re-running any script above will overwrite them.

---

Key Survival NPDR Options

All functionality is accessed through a single function:

surv_npdr(outcome, dataset, diff.type, regression, ...)

Option	Parameter	Values
Difference type	diff.type	"absolute" — symmetric |ΔT| ~ |ΔX| (best for interactions); "signed" — asymmetric ΔT ~ ΔX (best for main effects)
Regression model	regression	"lm" (default) — linear regression; "binomial" — logistic regression
KM imputation	km.impute	TRUE — replace censored times with KM conditional mean before ΔT; FALSE (default)
Covariate adjustment	covariates	e.g. c("age", "sex", "BMI") — pairwise differences included as regressors
Regularization	regularize, alpha	regularize = TRUE enables elastic-net; alpha = 1 → LASSO, alpha = 0 → ridge
Neighborhood	nbd.method, knn	"multisurf" — adaptive k (default); "relieff" + knn = 20 — fixed k

---

Citation

@article{kampmeyer2025survnpdr,
  title   = {Survival {NPDR}: Nearest-Neighbor Projected Distance Regression
             for Biomarker Discovery in Gene-Expression Survival Data},
  author  = {Kampmeyer, Patrick and Dawkins, Bryan A. and Urbanowicz, Ryan J.
             and Kazemi-Nia, Kia and McKinney, Brett A.},
  year    = {2026}
}

---

Related

• NPDR R package — original NPDR for non-survival outcomes
• rsurv — lognormal AFT simulation used in benchmarks