Current limitations in predicting mRNA translation with deep learning models

Area of Science:

• Bioengineering
• Molecular Biology
• Computational Biology

Background:

• Designing nucleotide sequences for specific properties is crucial in bioengineering, particularly for protein expression in research and mRNA vaccines.
• The 5' untranslated region (5'UTR) of messenger RNA (mRNA) significantly influences protein synthesis rates.
• Recent advancements include deep learning models to predict translation output from 5'UTR sequences, fueled by available large datasets.

Purpose of the Study:

• To assess the accuracy and generalizability of current deep learning models for predicting mRNA translational output.
• To evaluate model performance across different cell types and mRNA types (endogenous vs. reporter).

Main Methods:

• Utilized complementary datasets from two distinct cell types for model evaluation.
• Assessed deep learning models trained on specific datasets against new, diverse datasets.

Main Results:

• Deep learning models demonstrated high accuracy on their training datasets.
• Models showed poor generalization to other datasets, particularly those involving endogenous mRNAs.
• Significant differences between endogenous and reporter mRNA properties limit model applicability.

Conclusions:

• Current deep learning models face limitations in predicting translation control mechanisms and genetic variation impacts.
• Future directions involve integrating high-throughput measurements with machine learning for improved translation control understanding and construct design.