CircSI-SSL

CircSI-SSL applies self-supervised learning to identify circular RNA (circRNA) protein-binding sites from sequence-derived representations.

Key Features:

• Self-Supervised Learning: Employs a self-supervised learning paradigm to learn from unlabeled circRNA sequence data.
• Multi-View Feature Coding and RNA_Transformer: Uses multiple feature coding schemes combined with an RNA_Transformer model for cross-view sequence prediction to learn mutual information across multi-view representations.
• Fine-Tuning with Limited Labels: Fine-tunes the pretrained model using a minimal number of labeled samples to improve generalization from scarce annotations.
• Empirical Performance on circRNA Datasets: Validated on six circRNA datasets and reported to outperform existing algorithms, including at a low training:test ratio (1:9).
• Transplantation to linRNA Datasets: Maintains performance on six linear RNA (linRNA) datasets without modifications to network architecture or hyperparameters, indicating scalability.

Scientific Applications:

• Regulatory role analysis: Predicts circRNA protein-binding sites to facilitate study of circRNA regulatory roles in gene expression.
• Disease mechanism and therapeutic target investigation: Supports exploration of circRNA involvement in various diseases and identification of potential therapeutic targets.
• Biomarker discovery: Aids discovery of circRNA-related biomarkers through predicted circRNA–protein interaction sites.

Methodology:

Pretrain an RNA_Transformer via self-supervised cross-view sequence prediction using multiple feature coding schemes to learn mutual information across representations, then fine-tune the model with a minimal number of labeled samples.