Th195 - Stabl Pairwise Statistic – a Flexible Approach to Feature Selection in Omics with Relaxed Exchangeability Assumption

Abstract Text:

Background: Omics technologies enable the measurement of thousands of features per sample, offering unprecedented resolution. However, identifying predictive features for clinical outcomes remains challenging. The Stabl framework [1] addresses this by injecting artificial noise to estimate the False Discovery Rate (FDR) but assumes global feature exchangeability, which is problematic given the variability in omics data distributions. We propose the Pairwise Statistic (PS), which relaxes this assumption by requiring only pairwise exchangeability.

Methods: The PS quantifies the difference in selection frequency between a feature and its artificial counterpart. Assuming pairwise exchangeability, the distribution of non-predictive features aligns closely with their artificial counterparts relative to the clinical outcome. The optimal feature selection threshold is determined by minimizing a revised FDP+, calculated as the ratio of positive PS above the threshold to negative PS below it. Performance was evaluated using Monte Carlo cross-validation (5 folds, 5 repeats) on COVID-19 and Onset of Labor (OOL) datasets and compared to the original Stabl.

Results: The PS achieved predictive accuracy comparable to Stabl (COVID-19: 0.85 [0.75–0.94] vs. 0.79 [0.68–0.90]; OOL: 17.8 [14.3–21.6] vs. 17.8 [14.6–21.3]). Feature overlap between methods was high, with 3 of 4 features overlapping for COVID-19 and 18 of 19 for OOL.

Conclusion: The PS provides a robust alternative to Stabl by relaxing assumptions of global exchangeability, achieving comparable predictive performance and identifying overlapping feature sets. It offers a flexible and effective approach for omics-based feature selection and biomarker discovery.

[1] Hédou, J., Marić, I., Bellan, G. et al. Nat Biotechnol (2024). https://doi.org/10.1038/s41587-023-02033-x