Abstract
Codon optimality couples translation elongation to decay and is among the strongest reported sequence determinants of mRNA half-life. Long non-coding RNAs (lncRNAs) are largely untranslated. Is lncRNA stability sequence-encoded in the way mRNA stability is? We compared the classes on half-lives from the same assays, re-estimating the codon stabilization coefficient (CSC) inside each cross-validation fold to remove label leakage. Performance is the cross-validated coefficient of determination R² - 1 for perfect prediction, 0 for the training mean, negative for worse - and the Spearman rank correlation ρ between predicted and measured half-life. mRNA half-life was sequence-predictable (R² = 0.174, ρ = 0.44), and CSC alone (R² = 0.084) outperformed a six-member translation-independent set (R² = 0.008), an ordering that held in three further datasets, two assays and two species. lncRNA half-life was not: R² = -0.050 in same-assay HeLa lncRNAs (n = 364) and R² = +0.0007 from k-mer composition in the largest published lncRNA dataset (n = 33,285), where that study's own cross-validated ρ is 0.091. That value clears its permutation null: the classes differ in magnitude, not detectability. Pretrained genomic language models, one covering 99.6% of transcripts whole, left mRNAs at R² = 0.042 and lncRNAs at 0.0002-0.0036. Coding-potential strata change nothing; matching on length, GC and n with an identical feature set gives R² = 0.033 versus 0.0002. Signal injection puts the reliable detection limit near 0.5% of variance, above every lncRNA value obtained. mRNA-derived decay models should not be transferred to lncRNAs without re-validation.