Rank-based relational methods for interpretable biomedical modeling: a taxonomic review

Rank-based relational methods represent biomedical samples through within-sample ordering relations between features rather than through absolute measured values. In biomedical modeling, such representations can provide compact and interpretable decision rules and may reduce sensitivity to some normalization choices and sample-level monotonic shifts. This review examines three historically related but methodologically distinct strands of work: supervised Relative Expression Analysis (RXA) for cohort-level prediction, Relative Expression Ordering (REO) for reference-based single-sample inference, and system-level rank-conservation approaches such as Differential Rank Conservation (DIRAC). Existing surveys have mainly focused on gene-pair classifiers or individual methodological families. Here, we organize these methods within a common taxonomic framework based on three design dimensions: model definition, computational strategy, and implementation-level scalability. This framework is used to compare representative methods from simple pairwise rules to higher-order, hierarchical, evolutionary, ensemble, pathway-level, and emerging multi-omics extensions. We clarify their assumptions, computational trade-offs, software availability, validation requirements, and practical limitations across transcriptomic and other omics settings. Overall, we position rank-based relational methods as complementary tools for interpretable biomedical modeling, particularly when the goal is not only prediction but also transparent characterization of feature relationships and their disruption across phenotypic or disease states.