Improving deep models of protein-coding potential with a Fourier-transform architecture and machine translation task

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Distinguishing messenger RNAs (mRNAs) from long noncoding RNAs (lncRNAs) is crucial for genome annotation and understanding gene regulation.
• Current computational methods for protein-coding potential often rely on predefined features, limiting their adaptability.
• Sequence-to-sequence (seq2seq) models have shown success in natural language processing, offering a new paradigm for biological sequence analysis.

Approach:

• Developed a seq2seq deep neural network model incorporating LocalFilterNet (LFNet) to predict protein-coding potential.
• LFNet architecture is designed with an inductive bias for the three-nucleotide periodicity characteristic of coding sequences.
• Trained the model using a dual objective: RNA classification and simultaneous RNA-to-protein translation learning.

Key Points:

• Simultaneously learning translation alongside classification significantly improved the model's performance in distinguishing mRNAs from lncRNAs.
• LFNet effectively captures the inherent three-nucleotide periodicity of coding sequences.
• The model generates nucleotide-resolution importance scores, providing insights into sequence features guiding mRNA-lncRNA discrimination.

Conclusions:

• The developed deep learning approach enhances the accuracy of identifying protein-coding transcripts.
• LFNet offers a powerful tool for genome annotation and regulatory element discovery.
• Novel methods for estimating mutation effects using Integrated Gradients were explored, highlighting challenges in efficient approximation.