omicau benchmark protocol: a leakage-safe, compute-efficient evaluation of multi-omics audit and fusion workflows

This working paper and protocol prospectively specifies an independent, compute-efficient benchmark of omicau, a multi-omics audit and fusion workflow. Simulated datasets are independent experimental units evaluated with one frozen five-fold cross-validation partition. Real cohorts use five-fold, three-repeat group-aware outer cross-validation and three-fold inner cross-validation for tuning. External-method comparisons are restricted to a prespecified simulation subset, and explainability analysis is disabled except for one prespecified case study. Null controls are treated as pipeline sanity checks; group leakage is assessed through the grouping warning and the naive-versus-group-aware performance gap; batch and missingness outputs are interpreted as risk flags rather than causal proof. TCGA-BRCA PAM50 is included only as a secondary positive control, while primary real-cohort endpoints are externally defined. All analyses run before this protocol is frozen and archived are pilot-only and cannot support confirmatory claims. Definitive simulation seeds must pass a zero-overlap audit against retained pilot seeds. Primary interval-coverage claims are outside scope.