iDeepC

iDeepC predicts binding sites of RNA-binding proteins (RBPs) on circular RNAs (circRNAs) using deep learning to identify RBP–circRNA interactions when labeled training data are limited.

Key Features:

• Siamese Neural Network Architecture: Employs a Siamese neural network to learn pairwise relationships between circRNA sequences for binding-site prediction.
• Attention Module: Incorporates a lightweight attention module to focus on relevant sequence features in circRNAs and improve interpretability and accuracy.
• Metric Learning: Applies metric learning to learn similarity measures between circRNAs based on RBP-binding characteristics, aiding detection with limited samples.
• Transfer Learning Pipeline: Uses transfer learning by pretraining on labeled data from well-characterized RBPs to improve prediction for poorly characterized RBPs with small sample sizes.

Scientific Applications:

• Gene regulation and disease mechanism studies: Predicts RBP binding sites on circRNAs to support investigation of gene regulatory interactions and RNA-related disease mechanisms.
• Characterization of poorly characterized RBPs: Enables inference of binding sites for RBPs with scarce experimental labels by transferring knowledge from well-characterized RBPs.

Methodology:

Data preprocessing of circRNA sequences; model training of a Siamese network with attention and metric learning using labeled RBP–circRNA datasets with transfer learning from well-characterized RBPs; evaluation of predictive accuracy against benchmark RBP–binding circRNA datasets.