Human 5' UTR design and variant effect prediction from a massively parallel translation assay

Area of Science:

• Molecular Biology
• Genetics
• Bioinformatics

Background:

• Predicting the impact of cis-regulatory sequences on gene expression is crucial for biological discovery.
• 5' untranslated regions (UTRs) play a significant role in regulating translation and protein expression.

Purpose of the Study:

• To develop a predictive model linking human 5' UTR sequence to translation efficiency.
• To engineer novel 5' UTRs for precise control over protein expression levels.
• To extend the predictive approach to chemically modified RNA for therapeutic applications.

Main Methods:

• Combined polysome profiling of a large library of randomized 5' UTRs with deep learning.
• Utilized a genetic algorithm to engineer 5' UTRs based on the predictive model.
• Validated the model by testing truncated and naturally occurring 5' UTR variants.

Main Results:

• Successfully built a deep learning model accurately predicting ribosome loading from 5' UTR sequence.
• Engineered new 5' UTRs that precisely control protein expression levels.
• Demonstrated the model's predictive power on 35,212 truncated and 3,577 natural 5' UTR variants.
• Identified 45 disease-associated single-nucleotide variants (SNVs) that significantly alter ribosome loading.

Conclusions:

• The developed deep learning model enables accurate prediction and engineering of 5' UTR sequences for translation control.
• This approach is applicable to chemically modified RNA, supporting mRNA therapeutics and synthetic biology.
• Identified disease-SNVs impacting ribosome loading offer insights into molecular mechanisms of human diseases.