Robust ORF Deconvolution from Multiple Ribo-seq Datasets with Group LASSO Penalty

Abstract

Ribo-seq has provided compelling evidence that there are many translated but unannotated ORFs. Despite a growing body of peptidomic support for cryptic ORFs, the computational discovery of weakly translated, short and mutually overlapping ORFs has remained a major bottleneck. Here we introduce PRICE 2, a method designed to resolve these challenging translation events by adapting concepts from isoform deconvolution. PRICE 2 employs a generative model combined with group LASSO regularized regression to jointly infer active ORFs across multiple sequencing runs. In a comprehensive benchmark, PRICE 2 outperformed thirteen state-of-the-art tools in detecting short and overlapping ORFs. Orthogonal validation via immunopeptidomics demonstrated that ORFs identified by PRICE 2 enable the discovery of [~]40% more cryptic MHC-I peptides than the best-performing competing tool at a fixed false discovery rate. By analyzing public ribo-seq data across 45 distinct cell and tissue types, we constructed ORF-stock, a comprehensive database of non-canonical ORFs for immunopeptidomic studies.