Synthetic Biology Platform Development for Manufacture of Next Generation Biologics

After a recombinant protein destined for secretion has been transcribed it must be translocated from the cytosol of the cell to the endoplasmic reticulum (ER) to finish translation. The targeting process any secreted recombinant protein must go through for this to happen is co-translational translocation. Co-translational translocation is dependent on the presence of an N-terminal amino acid sequence known as the signal peptide. The efficiency of this signal peptide can be marred by its interaction with the recombinant protein resulting in poor cleavage specificity. As signal peptide efficiency is often seen as protein specific, this feature, paired with its recombinant counterpart, must be investigated to achieve higher expression levels of complex recombinant proteins in a host cell factory. This specificity poses a problem for vector optimisation in cell line development (CLD) where molecules change regularly. A library of 37 N-terminal signal peptides was curated, composed of native CHO signal peptides, literature-based signal peptides shown to be beneficial in a CHO host and synthetically designed signal peptides. The effect of these signal peptides on protein expression directly upstream of multiple AstraZeneca provided molecules was measured singularly and in combination. General performance, “all-purpose” signal peptides were identified across single chain molecules and monoclonal antibodies (mAbs). Using data from two light chains and one ScFv protein a machine learning model was created to predict the rank of a signal peptide in its molecular context. This model, created using XGBoost regression, was experimentally validated and resulted in three out of six signal peptide predictions falling into the 95% confidence interval range of measured scaled activity when expressing an ScFv. There is great potential to further optimise this model by introducing more molecules and signal peptides, increasing its prediction power, and reducing experimental options for vector optimisation.