ShortStop: a machine learning framework for microprotein discovery.

Area of Science:

• Genomics
• Proteomics
• Bioinformatics

Background:

• The human genome contains millions of small open reading frames (smORFs).
• Thousands of smORFs are actively translated, but distinguishing functional microproteins from non-functional sequences or regulatory elements is challenging.
• Current empirical methods for microprotein identification are time-consuming and costly.

Purpose of the Study:

• To develop a computational framework, ShortStop, for prioritizing functional smORFs.
• To create a machine learning model using distinct reference groups for classification.
• To address the need for scalable tools in microprotein research.

Main Methods:

• Developed ShortStop, a computational framework classifying smORFs into Swiss-Prot Analog Microproteins (SAMs) and Physicochemically Resembling In Silico Microproteins (PRISMs).
• SAMs represent known microproteins, while PRISMs serve as a proxy for non-functional sequences.
• Utilized machine learning to classify smORFs based on biochemical properties.

Main Results:

• ShortStop achieved high precision (90-94%) and recall (87-96%).
• Classified ~8% of translating smORFs as SAMs and 92% as PRISMs.
• Identified novel microprotein candidates, including one in the StAR gene, and differentially expressed microproteins in lung cancer.

Conclusions:

• ShortStop provides a scalable solution for characterizing microproteins and identifying functional smORFs.
• The framework aids in benchmarking discovery tools and advancing microprotein research.
• Offers a practical approach to differentiate functional microproteins from translational noise.