DeeReCT-PolyA: a robust and generic deep learning method for PAS identification.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Polyadenylation is crucial for mRNA maturation and gene expression regulation.
• Accurate identification of poly(A) signals (PASs) is essential for transcript annotation and understanding regulatory mechanisms.
• Existing PAS recognition methods are often motif-specific and human-centric, limiting their generalizability and insight into cross-species mechanisms.

Purpose of the Study:

• To develop a robust, motif-agnostic, and interpretable deep learning model for accurate PAS recognition.
• To enhance the generalizability and transferability of PAS identification models across different species.
• To uncover novel poly(A) patterns and compare regulatory mechanisms between species.

Main Methods:

• Developed a deep learning model for PAS recognition requiring no prior knowledge or human-designed features.
• Trained a single model on human PAS motifs and evaluated its performance on human and mouse datasets.
• Employed model interpretability techniques, including visualization of important oligomers/positions and conversion of convolutional filters into sequence logos.

Main Results:

• The proposed deep learning model outperforms state-of-the-art methods trained on specific motifs.
• The model demonstrates strong generalization capabilities to mouse datasets.
• Transfer learning between species further improved prediction accuracy, and novel poly(A) patterns were revealed through model interpretation.

Conclusions:

• The developed deep learning model offers a powerful and versatile tool for accurate PAS identification.
• The motif-agnostic and transferable nature of the model facilitates cross-species comparative genomics.
• Model interpretability provides novel insights into the underlying regulatory mechanisms of polyadenylation.