Reproducibility is a fundamental requirement for generating reliable and impactful scientific findings, particularly in the context of high-dimensional omics data such as transcriptomics and proteomics. A central task in these studies is differential expression analysis, which aims to identify molecular features whose expression levels differ across biological conditions. The datasets are often complex, characterized by a large number of features measured in relatively small sample sizes, technical noise, and sampling variability. Ensuring that detected differentially expressed features are reproducible across independent studies and experimental conditions is therefore critical.

The reproducibility-optimized test statistic (ROTS) framework was developed to address these challenges by prioritizing features that demonstrate high reproducibility. ROTS accommodates linear models, linear mixed-effects models, two-group comparisons, multi-group comparisons, and survival analysis. These modeling approaches enable robust identification of differentially expressed features in longitudinal studies, time-to-event analyses, and other complex experimental designs commonly encountered in clinical and systems biology research. The incorporation of model flexibility allows researchers to assess differential expression while accounting for confounding variables, repeated measures, and time-dependent effects, improving the biological relevance of the findings.

The reproducibility-optimization framework is validated using both simulated datasets and real-world omics studies, demonstrating that it consistently outperforms conventional statistical methods in identifying reproducible and biologically meaningful features, even under challenging conditions with noise and variability. Unlike the conventional approaches that focus on statistical significance, ROTS explicitly optimizes the statistic for reproducibility and generally offers a more robust criterion for feature selection in high-throughput data analysis.

Co-authors: Laura Elo