SurvMarker: an R package for identifying survival-associated molecular features using PCA-based weighted scores.

Abstract

Background: Identification of prognostic molecular features from high-dimensional molecular data is central to biomarker discovery in cancer and other complex diseases. Principal component analysis (PCA) is widely used for dimensionality reduction in survival studies, yet selecting individual features from principal components (PCs) remains challenging and often relies on arbitrary thresholds. To address this limitation, we developed SurvMarker, an R package that prioritizes survival-associated molecular features using a PCA-based scoring framework.

Results: SurvMarker applies PCA to normalized molecular data, jointly evaluates PCs using multivariable Cox proportional hazards models, and ranks features by aggregating absolute loadings across survival-associated PCs. Feature significance is assessed using an empirical null framework with false discovery rate control. In both synthetic global-null and permutation-based null simulations, SurvMarker showed comparative or better false positive control, particularly in small-n, large-p settings, compared with LASSO Cox, Elastic Net Cox, and Partial Least Squares Cox, while maintaining well-calibrated null p-value distributions. In the TCGA-LAML cohort, SurvMarker achieved the best predictive performance among these methods for gene expression data, with a C-index of 0.78 and an overall time-dependent AUC of 0.882 with similar applicability to miRNA expression data. Compared with sparse PCA-based and fixed per-PC threshold approaches, SurvMarker also achieved better predictive performance and yielded more compact, stable feature sets across different PC settings.

Conclusions: SurvMarker provides a robust, interpretable, and reproducible framework for identifying survival-associated molecular features from high-dimensional data. By combining survival-guided PC selection, weighted feature aggregation across PCs, and empirical null-based inference, it improves false discovery control, stability, and biological relevance, and offers a practical tool for biomarker discovery across multiple omics data types.