Cross-Program mRNA Process Models with Interpretable Machine Learning

The commercialization of mRNA based COVID vaccines has fulfilled the promise of mRNA science as a new generation of transformative medicines. The potential demand for variant-targeted vaccines of infectious diseases and personalized neoantigen expressing cancer vaccines require rapid process development and supply of mRNAs of newly designed sequences. Predictive in silico tools can be used to reduce the time and cost of process development and recommend optimal RNA process for new sequences. In this presentation, we will introduce how historical data, machine learning, and optimization techniques are utilized to deepen process understanding and accelerate process development. A structured relational database has been designed with historical data from process development experiments linked with metadata such as sequence and process attributes. Interpretable machine learning models were used to identify relationship among manufacturability, sequence attributes, and process parameters. Quantitative predictive models were trained and deployed to predict the process outcomes. Digital applications were created as an interface for scientists and engineers to utilize historical data as reference and compare with new experiments, and as an advisory tool to recommend optimal processes with explanations.