DECIPHER decodes cell types and predicts age from bulk omics

In a computational study using simulated benchmarks, experimental cell mixtures, and human datasets, researchers developed DECIPHER to infer cell-type composition from bulk omics data across multiple molecular modalities. The framework addresses technical batch effects by learning two distinct components: a domain-constant representation that captures shared biological features and a domain-specific representation that models platform-associated noise. By integrating nonlinear representation learning with differentiable non-negative least-squares optimization, the algorithm estimates cell-type proportions directly from the domain-constant space. Across simulated datasets, experimental cell mixtures, and real-world multi-omics benchmarks, DECIPHER produced robust and competitive deconvolution performance. Beyond identifying cell fractions, the learned domain-constant representations successfully predicted chronological age across independent cohorts and stratified clinical prognosis in lung adenocarcinoma.