Determinants of cysteine coverage by chemoproteomics

Abstract

Covalent chemoproteomics has become a powerful route to expanding the druggable proteome, enabling proteome-wide screening of compounds that covalently modify nucleophilic residues, most commonly cysteine. Yet typical experiments detect only 5-18% of the ~250,000 cysteines encoded by the human proteome, and the determinants of this limited coverage remain poorlydefined. Here we combine cysteine featurisation, explainable machine learning (ML) and robust statistical filtering to establish what governs cysteine detectability in iodoacetamide- desthiobiotin (IA-DTB) chemoproteomics. Using a scalable feature-generation pipeline, we computed physicochemical and structural properties for cysteines and the corresponding tryptic peptides and trained models that predict detectability with high accuracy. Shapley analysis identifies protein abundance and intrinsic peptide properties governing LC-MS behaviour as the dominant determinants, whereas cysteine structural descriptors contribute comparatively little, indicating that detection is constrained downstream of probe reactivity. We further introduce a distance-based statistical filter that improves reproducibility across replicates, and show that cell line selection, subcellular fractionation and alternative electrophiles together expand coverage by >30%. More broadly, this work moves optimisation of chemoproteomics workflows from empirical trial-and-error towards rational, data-driven design.