DextMP

DextMP predicts moonlighting proteins (MPs) within proteomes by analyzing titles, abstracts, and UniProt functional descriptions to identify multiple independent cellular functions.

Key Features:

• Textual Information Extraction: Systematically extracts titles from scientific literature, abstracts from research papers, and functional descriptions from the UniProt database for proteins.
• Advanced Language Models: Applies a combination of a deep unsupervised learning algorithm, Term Frequency–Inverse Document Frequency (TF-IDF) within a bag-of-words framework, and Latent Dirichlet Allocation (LDA) for topic modeling.
• High Predictive Accuracy: Achieves over 91% accuracy in predicting moonlighting proteins via cross-validation on datasets comprising known MPs and non-MPs.
• Genome-Wide Application: Applied to human, Saccharomyces cerevisiae (yeast), and Xenopus laevis genomes, estimating approximately 2.5–35% of these proteomes contain potential MPs.

Scientific Applications:

• Disease Research: Identifying MPs involved in disease pathways to inform studies of pathogenesis and potential therapeutic targets.
• Functional Annotation: Recognizing proteins with multiple independent roles to improve accuracy of functional annotations in biological databases.
• Computational Biology: Enhancing protein function prediction algorithms by accounting for proteins with multiple distinct functions.

Methodology:

DextMP extracts titles, abstracts, and UniProt functional descriptions, applies language models including a deep unsupervised learning algorithm, TF-IDF (bag-of-words), and LDA, and uses cross-validation to compare model performances for predicting moonlighting proteins.