Translation uniformity and evolutionary conservation are associated with microprotein mass spectrometry detectability

Abstract Background Non-canonical open reading frames (ncORFs) encode a substantial, largely unexplored layer of the human proteome. The short microproteins they produce frequently escape detection by standard mass spectrometry (MS) workflows, creating a “dark proteome” that hampers gene annotation and functional characterisation. To address this, sequence-derived attributes were used to identify factors associated with MS repository detectability and to build an interpretable predictive framework that enables rational prioritisation of ncORFs for proteomic discovery. Results 7,097 ncORF sequences (99% <150 residues) were curated from the TransCODE Consortium, and labelled by PeptideAtlas peptide-level evidence as MS-detected ( n = 1,725) or undetected ( n = 5,372). An ensemble combining ESM-2 protein language model embeddings with logistic regression and nineteen biological features (ESM2-LR) achieved a mean ROC-AUC of 0.721 (SD 0.018, coefficient of variation 2.5% across ten independent runs). Model-agnostic SHAP and permutation importance analyses concordantly ranked cumulative in-frame ribosome footprint uniformity (CIF-HBP) as the strongest associated factor, followed by mammalian ORF-level evolutionary conservation (ORBLv_mammal) and genomic biotype. Predicted HLA class I-binding peptide density, which captures physicochemical properties overlapping with optimal LC-MS/MS analyte characteristics, was also positively associated with detectability. Late-fusion analysis demonstrated that ESM-2 embeddings and biological features contribute complementary predictive signals. Conclusions These findings indicate that sustained translation, purifying selection for ORF structural integrity, and favourable peptide physicochemistry converge to influence ncORF detectability. The ESM2-LR framework provides a reproducible, scalable tool to prioritise candidate microproteins for targeted MS validation, gene annotation, and the interpretation of disease-associated regulatory variants, subject to future external validation on independent datasets.