PolyA-GLM: A comprehensive framework for De novo polyadenylation site prediction using genome language models

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Polyadenylation (poly(A)) sites are vital for post-transcriptional gene regulation.
• Accurate poly(A) site prediction aids in identifying RNA processing defects in diseases like cancer.
• Traditional methods face limitations in cross-species and cross-cell type generalization.

Purpose of the Study:

• To evaluate genome language models (GLMs) for accurate poly(A) site prediction.
• To leverage GLMs' ability to capture long-range genomic sequence dependencies.
• To develop an end-to-end pipeline for novel poly(A) site discovery.

Main Methods:

• Evaluated DNABERT-2, Nucleotide Transformer, and HyenaDNA using few-shot classification and fine-tuning.
• Assessed model performance in recognizing canonical polyadenylation signals (PASs) and their proximity to cleavage sites.
• Employed systematic signal perturbation for model interpretability validation.

Main Results:

• GLMs effectively recognized canonical PASs and their spatial relationships to cleavage sites.
• HyenaDNA achieved an AUC of 0.751 in few-shot learning, with performance gains after fine-tuning.
• A token-level classification approach enabled precise, position-wise poly(A) site identification.

Conclusions:

• Genome language models show significant promise for advancing poly(A) site prediction.
• The PolyA-GLM pipeline facilitates the discovery of novel regulatory elements.
• GLMs offer a powerful approach to understanding RNA processing and gene regulation.