Year: 2026
Type: preprint
Abstract: Regression analysis of high-throughput omics data is often made difficult by noise, batch effects, and the loss of information resulting from the discretization of continuous variables. Existing machine learning algorithms often struggle to extract stable biological signals under these heterogeneous conditions. We introduce BI-SCUDO Regression, a rank-based regression framework designed for omics data analysis. By leveraging within-sample rankings and network-based predictions, BI-SCUDO Regression bypasses the need for complex normalization and provides robust, interpretable results. In synthetic benchmarks, BI-SCUDO Regression demonstrated strong resilience to noise, maintaining predictive and feature-selection stability under conditions in which standard models exhibited performance degradation. When applied to cohorts from TCGA datasets, BI-SCUDO Regression consistently outperformed LASSO, Random Forest Regression, and XGBoost in predicting survival, with an average increase in Ratio of Performance to Deviation ranging from 43% to 86%, depending on the comparator. Furthermore, the algorithm consistently identified parsimonious, statistically significant, and biologically relevant biomarkers, with enrichment for cancer-related terms across all stage stratifications. Although BI-SCUDO Regression incurs higher training times due to its genetic optimization strategy, it offers a robust and interpretable approach to outcome prediction and biomarker discovery, with potential applications in precision medicine, where noisy data compromise the performance of conventional regression methods.
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
Awards: G231133531, G972786147
OpenAlex ID: W7236165514
