<b>Deep conservation does not amplify proteomic detection of reverse-strand shadow ORFs in vertebrates</b>

<jats:p>Reverse-strand overlapping open reading frames ("shadow ORFs") are pervasive on the antisense strand of protein-coding genes, but whether they are translated in vertebrates has not been tested at the class level. We catalogued shadow ORFs — defined here as the longest stop-free stretch in a frame, without an initiation-codon requirement — in 318 deeply conserved human genes (orthologues required in at least four of five non-human vertebrates; 248 reach zebrafish, ~430 Myr), quantified frame openness against two null models, and ranked 1,908 six-frame candidates. To test translation without overclaiming, we ran a pre-specified mass-spectrometry enrichment test on three peptide sets of 958 each, matched on peptide length and set size: T, a high-conservation test set (27 genes, composite score 73.3–94.9); C, a low-conservation reverse control (190 genes, composite &lt; 50); and S, a shuffled decoy set derived from T. Each set was searched under identical criteria across six public proteomes, and results were combined as the union of peptides confident in at least one dataset. T exceeded S (union OR = 2.63, P = 1.1 × 10⁻¹¹), showing that the search separates real reverse-frame sequence from shuffled sequence; this bounds the assay's statistical resolution but does not by itself establish sensitivity to genuine translation. Against C, however, T showed no class-level enrichment (union OR = 0.81, P = 0.076); with 958 peptides per set the design had 80% power to detect an odds ratio of 1.35 or greater, so smaller enrichments remain untested. The result was robust to a ≥ 9-amino-acid length restriction (OR = 0.81) and to gene-level clustering (permutation P = 0.26). Cross-dataset recurrence did not separate conserved from low-conservation candidates (10 T vs 12 C peptides confident in ≥ 3 datasets; P = 0.83), and one shuffled peptide was confident in four of six datasets, quantifying a non-trivial false-positive floor. Strand-specific total RNA-seq detected antisense transcription at 26 of the 27 T loci but at only 0.9–3.9% of sense abundance, and antisense level did not track the reproducible peptide loci. We report the recurring loci, led by EGLN1 and NFIB, as calibrated, uncertainty-annotated leads for targeted validation — leads, not proof of function.</jats:p>