Leakage-Aware Discovery of Small Biomarker Panels

Discovers small binary-classification biomarker panels from count or expression matrices while prioritizing detectability, expression stability, and univariate discrimination. Candidate filtering and panel selection can be repeated inside nested cross-validation to reduce information leakage. The package provides shared resampling splits, exhaustive small-panel search, logistic model fitting with an automatic ridge fallback for unstable separation-prone fits, out-of-fold evaluation, selection-frequency summaries, and optional 'DESeq2' differential-expression support. The nested model-selection workflow follows Varma and Simon (2006) , and the optional differential-expression analysis uses Love, Huber, and Anders (2014) .

detectPanel

detectPanel is an R package for discovering, validating, and applying small biomarker panels from feature-by-sample count or expression matrices.

It provides an end-to-end workflow for feature filtering, small-panel search, nested cross-validation, model fitting, prediction on new samples, stability assessment, visualization, and export.

Overview

High-dimensional molecular datasets often contain many candidate biomarkers but relatively few samples. detectPanel is designed for binary-outcome problems where the goal is to identify a compact panel while reducing information leakage during model selection and validation.

A typical workflow is:

1. prepare a feature-by-sample assay matrix and sample metadata;
2. filter or exclude unsuitable features;
3. search for compact candidate panels;
4. estimate internal performance with nested cross-validation;
5. refit a final model on all available training samples;
6. apply the fitted model to new samples.

Key features

• Detectability-aware feature filtering
• Feature exclusion by exact name or regular expression
• Small biomarker panel search
• Repeated stratified resampling
• Leakage-aware nested cross-validation
• Logistic panel fitting with stability safeguards
• Automatic L2-penalized fallback for unstable logistic regression fits
• Prediction from fitted detectPanel_result objects
• Panel and feature selection-frequency summaries
• Marker, ROC, heatmap, volcano, and PCA-QC plotting utilities
• Export of analysis results

Installation

From GitHub

install.packages("remotes")
remotes::install_github("Emr-27/detectPanel")

From CRAN

Once the package is available on CRAN:

install.packages("detectPanel")

Quick start

The example below generates a small synthetic count matrix and can be run directly after installing the package.

library(detectPanel)

set.seed(42)

n_samples <- 48
n_features <- 18

group <- rep(c("Control", "Case"), each = n_samples / 2)

counts <- matrix(
  rpois(n_features * n_samples, lambda = 80),
  nrow = n_features,
  dimnames = list(
    paste0("marker", seq_len(n_features)),
    paste0("S", seq_len(n_samples))
  )
)

# Add signal to a few features in the positive class
counts[1:3, group == "Case"] <-
  counts[1:3, group == "Case"] + 80

metadata <- data.frame(
  group = group,
  row.names = colnames(counts)
)

result <- discover_panel(
  counts = counts,
  metadata = metadata,
  outcome = "group",
  positive = "Case",
  panel_size = 3,
  candidate_n = 8,
  detection_threshold = 5,
  min_mean = 10,
  min_median = 5,
  min_detection = 0.50,
  min_group_detection = 0.30,
  min_auc = 0.55,
  outer_v = 3,
  outer_repeats = 1,
  inner_v = 3,
  inner_repeats = 1,
  seed = 42
)

result

The selected panel is available from:

result$final_panel

Prediction

A fitted detectPanel_result object can be applied directly to a new feature-by-sample matrix.

new_counts <- counts[, 1:4, drop = FALSE]

probability <- predict(
  result,
  new_counts,
  type = "response"
)

classification <- predict(
  result,
  new_counts,
  type = "class"
)

probability
classification

New data must contain the features required by the fitted model. The saved model retains the preprocessing and decision information needed for prediction.

Validation and interpretation

detectPanel uses nested cross-validation to estimate internal validation performance while keeping candidate selection and panel search within the training data of each outer split.

Nested validation summaries are available from:

result$nested$outer_summary

The final model is refitted on all available training samples for downstream prediction. Its training performance is apparent performance and should not be used as a substitute for nested cross-validation or independent external validation.

A selected panel should therefore be interpreted as a candidate biomarker model rather than evidence of clinical validation. Independent cohorts and a locked preprocessing and prediction procedure are recommended before prospective or clinical use.

Visualization

Depending on the fitted result and analysis stage, detectPanel provides plotting functions for marker expression, marker ROC curves, panel heatmaps, volcano plots, and PCA-based QC.

Examples include:

plot(result, type = "roc")
plot(result, type = "feature_frequency")

See the function documentation for additional plotting options.

Documentation

Detailed workflows are included as package vignettes:

browseVignettes("detectPanel")

The package includes a vignette covering nested validation and the recommended interpretation of internal validation results.

Function-level documentation is available through R help, for example:

?discover_panel
?fit_panel
?nested_validate_panels
?export_results

Citation

If you use detectPanel in research, please cite the package:

citation("detectPanel")

License

detectPanel is distributed under the MIT License.