Optimizing mRNA sequences with deep neural networks

The COVID-19 pandemic has presented new challenges to individuals world-wide. Since the first reports of infections in the US more than 90 million individuals have become infected and more than 1 million have died. SARS-CoV-2 genome has various open reading frames (ORFs) encoding 16 non-structural proteins (NSPs), 4 structural proteins and several accessory proteins. The genome of RNA virus can easily generate mutations as virus spreads. The constant emergence of new mutations in SARS-CoV-2 is the major challenge for the ongoing development of antiviral drug and broad neutralizing antibodies. The two mRNA vaccines from Pfizer/BioNTech and Moderna are moderate effective, 45 to 75 percent at protecting people from in preventing infection from the delta variant, and both of them have received emergency use authorization. More seriously, the omicron variant was first detected in southern Africa and quickly expanded to the whole world. According to a recent study, traditional dosing regimens of COVID-19 vaccines available in the US do not produce antibodies capable of recognizing and neutralizing the Omicron variant. The global data shows the coronavirus pandemic is far from over. Thus, more variants are expectable and some of them may escape the immune response produced after vaccination. How to keep the efficacy of existing mRNA vaccines on variants is challenging us. Aside from SARS- CoV-2, mRNA medicines against cancer and other infectious disease, such as Ebola, Zika virus, and influenza, are advancing through clinical trials. The goal of this project is to develop an integrated deep learning model to optimize 5'UTR, codon usage, and 3'UTR at same time that enables users to design the optimal mRNA sequence to enhance protein expression level, thus to improve the efficacy of mRNA medicines. mRNA medicines hold great promise for the treatment of a wide variety of disease, extending from prophylactics to therapeutics for infectious diseases, cancer, and genetic disease. mRNA medicines have several beneficial features: safety, efficacy, production, and speed. Multiple factors are involved to regulate the stability and efficiency of mRNA, including 5' untranslated region (UTR), 3' UTR, codon et al, and several in silico approaches have being developed to optimize these factors respectively However, as these factors always function together during the translation of mRNA, and individual optimization is insufficient. Thus, a novel integrated deep learning model for these factors is needed to comprehensively enhance the stability and efficiency of mRNA medicine. In silico optimization of mRNA vaccine provides a fast methodology to investigate all possible integration of the ORF, 5' UTR and 3'UTR and identify the optimal mRNA vaccine. The interdisciplinary team proposed to develop the following aims: (1) developing deep learning models for 5' UTR, codon, and 3' UTR respectively, and integrated model for systemic optimization of 5' UTR, codon, and 3' UTR; and (2) experimentally validate the integrated models by designing the 5'UTR, codon, and 3'UTR sequence for representative.