Penguin

Penguin predicts pseudouridine (Ψ) sites from Oxford Nanopore Technologies direct RNA sequencing data to identify pseudouridylated uridines.

Key Features:

• Integration of Machine Learning Models: Implements Support Vector Machine (SVM), Random Forest (RF), and Neural Network (NN) classifiers for pseudouridine site prediction.
• Feature Extraction from Raw Signals: Extracts features from raw Nanopore signal traces and from basecalled k-mers.
• Automated Data Preprocessing Pipeline: Performs read alignment with Minimap2 and signal extraction with Nanopolish as preprocessing steps.
• High Predictive Accuracy: Achieved 93.38% and 92.61% accuracies with SVM in random split and independent validation tests, respectively, exceeding previously reported maximum of 76.0% for pseudouridine predictors.
• Versatility for Other RNA Modifications: Adaptable to the identification of other types of RNA modifications.

Scientific Applications:

• Hek293 dataset analysis: Predicted 6,137,606 pseudouridylated U-mers out of 67,491,289 total U-mers in Hek293 cells.
• HeLa dataset analysis: Predicted 1,193,192 pseudouridylated U-mers out of 229,637,931 total U-mers in HeLa cells.
• Comparative analysis between cell lines: Identified unique genomic pseudouridine locations with an overlap of 0.01% (6,482 shared locations) between Hek293 and HeLa and reported pseudouridylation rates of 9% in Hek293 versus 0.5% in HeLa.

Methodology:

Preprocessing used Minimap2 for alignment and Nanopolish for signal extraction; features were derived from raw Nanopore signals and basecalled k-mers, and SVM, RF, and NN classifiers were trained and evaluated using random split and independent validation testing.