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

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Ribosomes synthesize proteins by processing messenger RNA (mRNA) sequences.
• Distinguishing protein-coding mRNAs from long noncoding RNAs (lncRNAs) is crucial for genome annotation.
• Current computational methods often rely on predefined features for protein-coding potential prediction.

Purpose of the Study:

• To develop a novel deep learning approach for predicting protein-coding potential.
• To improve the classification accuracy of mRNAs versus lncRNAs.
• To identify sequence features that guide translation and differentiate RNA types.

Main Methods:

• A sequence-to-sequence (seq2seq) deep neural network model was formulated.
• The model incorporated LocalFilterNet (LFNet) to capture three-nucleotide periodicity in coding sequences.
• Simultaneous training for translation and classification tasks was employed.

Main Results:

• Simultaneously learning RNA-to-protein translation improved classification performance.
• The model generated nucleotide-resolution importance scores, highlighting distinguishing sequence patterns.
• A new method for estimating mutation effects using Integrated Gradients was developed.

Conclusions:

• Deep learning models, particularly seq2seq architectures with inductive biases like LFNet, can effectively predict protein-coding potential.
• Jointly training for translation enhances RNA classification and feature interpretation.
• The developed methods offer insights into RNA sequence regulation and mutation impact.