Machine Learning for Designing Next-Generation mRNA Therapeutics.

Area of Science:

• Molecular Biology
• Bioinformatics
• Synthetic Biology

Background:

• Messenger RNA (mRNA) therapeutics and vaccines have rapidly advanced, yet optimization of noncoding sequences like 5' untranslated regions (UTRs) remains underexplored.
• Current mRNA designs often use unmodified UTRs borrowed from highly expressed genes, limiting control over translation efficiency and stability.

Purpose of the Study:

• To demonstrate the potential of model-driven design for unprecedented control over 5' UTR function in mRNA therapeutics.
• To develop and validate quantitative models linking 5' UTR sequences to translation efficiency and ribosome loading.

Main Methods:

• Utilized polysome profiling and high-throughput sequencing to quantify ribosome loading across millions of synthetic 5' UTR sequences.
• Developed Optimus 5-Prime, a convolutional neural network model trained on experimental data to predict 5' UTR function.
• Validated model predictions using held-out data, large libraries of human 5' UTR fragments, and independent translation reporter assays.

Main Results:

• Accurate predictive models of biological regulation were learned from synthetic and diverse 5' UTR datasets, even with chemical modifications like pseudouridine.
• 5' UTRs demonstrated consistent functional impacts across different coding sequences and chemical modification contexts.
• Optimus 5-Prime, combined with advanced design algorithms, enabled the generation of de novo 5' UTR sequences with precisely controlled translation efficiencies.

Conclusions:

• Model-driven design, integrating synthetic biology and machine learning, provides a powerful approach to optimize mRNA 5' UTRs for therapeutic applications.
• Advanced design algorithms offer improved speed and sequence diversity compared to traditional methods like genetic algorithms.
• This methodology can be extended to optimize other gene regions and diverse applications within mRNA technology.