SPROTify: A protein-centric machine learning framework for small protein identification with validation in cancer proteomics.

Abstract

Small proteins, typically under 100 amino acids in length, constitute an emerging class of functional biomolecules involved in diverse biological processes. Despite their importance, they are frequently overlooked in genome annotation, as conventional prediction tools prioritize canonical-length open reading frames (ORFs) or rely on nucleotide sequences that may not capture the protein-level physicochemical properties governing stability and detectability. To address this, we developed SPROTify, a protein-centric machine learning framework designed for high-accuracy small protein prediction using amino acid-derived features. Trained on curated, experimentally validated human data and refined through hyperparameter optimization, the model achieved an accuracy (ACC) and F1-score of 0.92, with an AUC of 0.97 on an independent test set. Its generalizability was further assessed using independent cross-species, alternative-transcript/isoform-associated small-protein (ATISP), and cancer proteomics datasets. SPROTify demonstrated strong and consistent classification performance across these diverse external validation datasets. In six cancer proteomics cohorts, 334 small-protein candidates were identified, including 36 differentially expressed small proteins (DESPs) exhibiting strong cancer-type specificity. Structural modeling of these DESPs revealed unique fold characteristics and tissue-specific patterns, suggesting specialized functional roles in tumor biology. These findings underscore the potential of SPROTify as a robust, protein-centric engine for prioritizing small proteins. All scripts and datasets for SPROTify are freely available at https://github.com/CSBM-Lab/SPROTify.git.