PERSEO: adaptive per-feature distribution modelling for differential analysis of heterogeneous omics data

Abstract Differential analysis is fundamental to biomarker discovery across omics disciplines. However, existing statistical tools enforce a single probability distribution across all features, introducing systematic misspecification errors that compromise statistical power and false discovery control. Here we characterized 131 public datasets from nine omics profiles and found that no profile is governed by a single distribution, while standard normalizations merely displace empirical distributions rather than conforming them to downstream model assumptions. To overcome modality-specific constraints, we present PERSEO, an omics-agnostic R framework for feature-wise differential analysis across any high-throughput layer, irrespective of origin or platform. By operating within generalized additive models for location, scale and shape (GAMLSS) with Jacobian-corrected information criteria, PERSEO adaptively models each feature under its optimal distribution, evaluating counts, continuous intensities, proportions, and zero-adjusted measurements on a unified scale. In benchmarks spanning 43 profile-and-stress scenarios, PERSEO matched reference methods in bulk RNA-sequencing and outperformed specialized tools in metagenomics, proteomics, and metabolomics, where its F1-score advantage reached 0.19 under noise and outlier contamination. On 12 public cohorts, PERSEO resolved cross-method discordances, recovering verified differences while exposing pervasive false discoveries in field-specific tools. PERSEO is freely available at https://github.com/OPSergio/Perseo.