Talk abstracts
Talk on Saturday 09:15-09:30am submitted by Soon Yi
Fine scale structural information substantially improves multivariate regression model for mRNA in-vial degradation prediction
Soon Yi (Department of Biochemistry, Case Western Rserve University School of Medicine), Sara E Ali (Moderna, Inc), Yashrajsinh Jadeja (Moderna, Inc), J. Wade Davis (Moderna, Inc), Mihir Metkar (Moderna, Inc)
Abstract:
Significance
mRNA vaccines require ultra-cold storage, creating significant distribution challenges globally. Current structural metrics for predicting mRNA stability achieve only modest accuracy (correlation coefficients <0.6), limiting rational design of stable therapeutics. Our research aimed to answer the following question: Can novel sequence-derived feature improve mRNA stability prediction to enable better therapeutic design?
Methods
We analyzed curated datasets of mRNA variants with experimentally measured in-solution half-lives. We developed base-pairing probability log-odds (LO), transforming base-pairing probabilities into a bidirectional scale (-∞ to +∞) to capture local structural differences missed by traditional metrics. We created STRAND (Stability Regression Analysis using Nucleotide-Derived features), a four-feature multivariate model combining LO with free energy, average unpaired probability, and GC content. Performance was evaluated using Leave-One-Out Cross-Validation on Nano-Luciferase variant sequences (N=69) and tested on independent eGFP sequences (N=13).
Results
Traditional metrics showed moderate correlations with half-life (Pearson r = -0.52 for free energy, r = -0.47 for average unpaired probability). Sequences with nearly identical traditional metrics exhibited up to 3-fold stability differences. Our log-odds metric achieved superior correlation (Pearson r = -0.79). STRAND demonstrated strong predictive performance on independent test data (Spearman r = 0.82, Pearson r = 0.72, RMSE = 0.19), substantially outperforming published deep learning models including DegScore-XGBoost (Spearman r = 0.37) while using only four interpretable features.
Conclusions
Base-pairing probability log-odds provides orthogonal structural information that significantly improves mRNA stability prediction. Our compact STRAND model offers superior performance compared to complex deep learning approaches while maintaining practical utility. This could facilitate development of more stable mRNA therapeutics with reduced cold-storage requirements, potentially expanding global access.
References:
https://www.biorxiv.org/content/10.1101/2025.08.15.670605v2
Keywords: RNA Structure Prediction, mRNA Vaccine, mRNA Therapeutics