Latent-Factor Stability Is Not Sufficient: Evaluating SABLE-MO in Simulated and Cancer Multi-Omics Data

Abstract

Reproducibility is often used to assess latent factors learned from multi-omics data. Yet factors that recur across model refits may not reflect the shared biological structure of interest. We investigated this issue using SABLE-MO, a shared/private representation-learning framework with a sep-arate factor-retention procedure. We evaluated the method in simulations with known latent structure. The original retention rule produced no false biological shared-factor acceptance across 800 independent-null datasets, but accepted at least one false factor in 70 of 200 correlated-private datasets (35%). Following an exploratory analysis of this failure, we increased the lower-decile aligned-correlation threshold from 0.40 to 0.60 and tested it on a new, independently seeded simulation bank. False acceptance decreased from 33% to 19%, while recovery of true shared factors changed only slightly (94.375% to 94.0%). Exact modality-support attribution reached 67.5%, below its prespecified criterion. We also analyzed matched RNA expression, DNA methylation, and copy- number data from TARGET osteosarcoma (78 patients), TCGA-SARC (250), and TCGA-KIRC (304). Outcome-blind model selec- tion favored a simpler shared/private representation in TARGET.In both TCGA cohorts, the more complex representation improved cross-view retrieval and missing-view reconstruction relative to the simpler model, but worsened ordinary reconstruction. These external comparisons were performed before factor retention. Our results show that factor reproducibility, representation quality,and biological interpretation require separate assessment. Stricter retention reduced errors in a specific simulated setting but did not eliminate them