Deep learning models for predicting RNA degradation via dual crowdsourcing

Area of Science:

• Biotechnology
• Computational Biology
• Molecular Biology

Background:

• Messenger RNA (mRNA) therapeutics show great promise, exemplified by COVID-19 vaccines.
• The widespread application of mRNA is hindered by the inherent instability of RNA molecules, primarily due to in-line hydrolysis.
• Accurate prediction of RNA degradation is crucial for developing more stable RNA therapeutics.

Purpose of the Study:

• To develop accurate predictive models for RNA degradation, specifically in-line hydrolysis.
• To assess the efficacy of machine learning approaches in predicting RNA stability.
• To support the design of stabilized messenger RNA molecules for therapeutic applications.

Main Methods:

• A crowdsourced machine learning competition ('Stanford OpenVaccine') was organized on Kaggle.
• Generated a dataset of 6,043 diverse RNA constructs (102-130 nucleotides) using crowdsourcing on the Eterna platform.
• Collected single-nucleotide resolution measurements of RNA degradation for model training and validation.

Main Results:

• The winning machine learning model achieved 41% accuracy at the nucleotide level, with predictions within experimental error.
• Models demonstrated successful generalization to predict degradation of longer mRNA molecules (504-1,588 nucleotides).
• The developed models showed improved accuracy compared to previously published methods for predicting RNA degradation.

Conclusions:

• Machine learning models can accurately represent and predict in-line hydrolysis, a key factor in RNA instability.
• These models are valuable tools for designing stabilized messenger RNAs, advancing RNA-based therapeutics.
• Integrating crowdsourcing platforms for both data generation and machine learning can accelerate scientific discovery for urgent problems.