CircPCBL

CircPCBL identifies plant circular RNAs (circRNAs) from raw RNA sequence data using a deep learning architecture to distinguish circRNAs from long non-coding RNAs (lncRNAs) and accommodate plant-specific splicing characteristics.

Key Features:

• Deep learning architecture: Implements a novel deep learning model that classifies raw RNA sequences for circRNA identification.
• CNN-BiGRU detector: Processes one-hot encoded RNA sequences, capturing local dependencies through convolutional neural networks (CNN) and long short-term memory units (BiGRU).
• GLT detector: Extracts k-mer features with k = 1 to 4 to provide additional sequence information.
• Feature fusion and classification: Concatenates outputs from the two detectors and feeds them into a fully connected layer to produce the final classification.
• Plant-specific sequence handling: Addresses non-GT/AG junction sites and the scarcity of reverse complementary sequences and repetitive elements in flanking introns.
• Detection targets: Distinguishes plant circRNAs from lncRNAs and recovers experimentally reported circRNAs and lncRNAs in tested species.
• Performance metrics: Demonstrated cross-species validation results including an F1 score of 85.40% on a six-species validation set, species-specific scores for Cucumis sativus (85.88%), Populus trichocarpa (75.87%), and Gossypium raimondii (86.83%), and an average accuracy of 94.08% on human datasets.

Scientific Applications:

• Plant circRNA discovery: Identification of circRNAs across diverse plant species despite atypical splice signals.
• CircRNA versus lncRNA classification: Distinguishing circRNAs from long non-coding RNAs in sequencing datasets.
• Cross-species prediction: Applying trained models to independent species with reported cross-species performance metrics.
• Experimental validation support: Recovering experimentally reported circRNAs and lncRNAs in species such as rice and Poncirus trifoliata.
• Potential extension to animal datasets: Indicating applicability to animal/human circRNA identification based on reported accuracy on human datasets.

Methodology:

Input RNA sequences are one-hot encoded and processed by a CNN-BiGRU detector and a GLT detector extracting k-mer (k = 1–4) features; the detector outputs are concatenated and passed through a fully connected layer for final deep-learning-based classification.