Context-Aware Poly(A) Signal Prediction Model via Deep Spatial-Temporal Neural Networks.

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• Polyadenylation (Poly(A)) is a critical post-transcriptional modification in eukaryotic mRNA maturation.
• Accurate identification of Poly(A) signals (PASs) is essential for understanding gene expression regulation and mRNA metabolism.

Purpose of the Study:

• To develop a novel deep learning model, MCANet, for accurate genome-wide prediction of Poly(A) signals.
• To adaptively capture spatial-temporal contextual dependencies in Poly(A) signal identification.

Main Methods:

• Proposed a deep dual-dynamic context-aware model (MCANet) utilizing multiscale convolution and self-attention networks.
• Implemented identity connectivity for enhanced feature map contextualization.
• Introduced a fully parametric rectified linear unit (FP-RELU) for improved training and generalization.
• Employed a cross-entropy loss (CL) function to focus on challenging samples.

Main Results:

• MCANet demonstrated superior performance in predicting Poly(A) signals across various datasets.
• Ablation studies confirmed the effectiveness of MCANet's network design for feature learning and prediction.
• The model successfully captured spatial-temporal contextual information for accurate PAS identification.

Conclusions:

• MCANet offers a powerful and effective approach for identifying Poly(A) signals at the genome level.
• The proposed model advances the understanding of mRNA metabolism and translation regulation mechanisms.
• MCANet's architecture provides a robust framework for complex biological sequence analysis.