Deep learning models for predicting RNA degradation via dual crowdsourcing.

Area of Science:

• Biochemistry
• Computational Biology
• Machine Learning

Background:

• Messenger RNA (mRNA) therapeutics show great promise, but their limited stability due to in-line hydrolysis hinders distribution.
• Predicting RNA degradation is crucial for developing more stable and effective mRNA-based medicines.

Approach:

• A crowdsourced machine learning competition on Kaggle, "Stanford OpenVaccine," was used to predict RNA degradation.
• Generated 6043 diverse RNA constructs (102-130 nucleotides) via crowdsourcing on the Eterna platform for single-nucleotide resolution measurements.
• Top models integrated natural language processing and data augmentation with RNA secondary structure predictions.

Key Points:

• The winning machine learning models achieved 41% accuracy at the nucleotide level, within experimental error.
• Models demonstrated generalization by accurately predicting degradation on longer mRNA molecules (504-1588 nucleotides).
• The study highlights the power of integrating crowdsourcing for both data generation and machine learning model development.

Conclusions:

• Machine learning models can accurately predict in-line hydrolysis, enabling the design of stabilized messenger RNAs.
• This approach accelerates the development of thermostable mRNA therapeutics, overcoming a key distribution challenge.
• The combined use of crowdsourcing platforms for data creation and machine learning is a viable strategy for rapid scientific discovery.