Predicting mean ribosome load for 5'UTR of any length using deep learning

Area of Science:

• Computational Biology
• Genetics
• Molecular Biology

Background:

• The 5' untranslated region (UTR) is crucial for regulating mRNA translation and protein levels.
• Accurate modeling of 5'UTR sequences offers insights into translational control and genetic variant interpretation.
• Existing models are limited by the sequence lengths tested in reporter assays.

Purpose of the Study:

• To develop a novel computational model for predicting translation efficiency from 5'UTR sequences of any length.
• To overcome the limitations of previous models regarding sequence length restrictions.
• To improve the interpretation of genetic variants affecting translation.

Main Methods:

• Introduction of a novel neural network operation called "frame pooling".
• Development of a Mean Ribosome Load (MRL) prediction model utilizing frame pooling.
• Evaluation of the model on fixed-length sequences and diverse genome-wide datasets.

Main Results:

• The frame pooling model demonstrates state-of-the-art performance on randomized sequences.
• The model exhibits superior generalization capabilities on longer sequences.
• Accurate prediction of translation-related data across various genome-wide datasets was achieved.

Conclusions:

• Frame pooling enables accurate MRL prediction for 5'UTRs of arbitrary length.
• The developed model offers improved performance and generalization for predicting translational control.
• The open-source model can assist in identifying pathogenic genetic variants, exemplified by a case study on HBB in beta-thalassemia.