MSCAN

MSCAN predicts RNA methylation sites using a multi-scale self- and cross-attention deep learning network to detect post-transcriptional RNA modifications across diverse RNA types and support study of their functions and regulatory mechanisms.

Key Features:

• Multi-Scale Attention Mechanism: Leverages a multi-scale self- and cross-attention network to capture dependencies at various sequence scales, addressing limitations of BiLSTM and CNN models that struggle with long-distance sequence dependencies and lack parallel computation.
• Integration of NLP Techniques: Incorporates natural language processing methodologies to process RNA sequences and enhance modification-site prediction accuracy.
• High Predictive Accuracy: Demonstrates prediction of twelve RNA modification types (m6A, m1A, m5C, m5U, m6Am, m7G, Ψ, I, Am, Cm, Gm, Um) with area under the ROC curve ranging from 65.69% to 99.04%, outperforming state-of-the-art models.
• Generalization Capabilities: Exhibits strong generalization across various RNA modification types, indicating robustness across diverse modification contexts.

Scientific Applications:

• Epigenomics: Enables mapping of RNA methylation landscapes to investigate epitranscriptomic regulation of gene expression.
• Disease association studies: Facilitates analysis of altered RNA modification patterns implicated in diseases to explore potential regulatory and pathological roles.

Methodology:

Implements a binary classification framework within a deep learning architecture using self- and cross-attention mechanisms to enhance sequence information interaction at multiple scales.