A poly-omics machine-learning method to predict metabolite production in CHO cells
2 Department of Computer Science and Information Systems, Teesside University, UK
* Author to whom correspondence should be addressed.
The success of therapeutic proteins such as insulin has led to the massive recognition of biological medical products as highly effective clinical drugs. As the use of biologics gains popularity, in industrial biotechnology there is a push to maximise their production. The ovary cells of the Chinese hamster (CHO cells) are the most common production cell line, however - like most mammalian cells - they are very inefficient in producing desired compounds. Culture bioengineering can improve the yield, but identifying the optimal interventions is usually expensive and time-consuming. Machine learning coupled with computational modelling of CHO cells has the potential to effectively elucidate optimal bioengineering steps towards improved production of therapeutic metabolites and proteins.
In this study, we combine machine learning techniques with gene expression profiling and metabolic modelling to estimate lactate production in CHO cell cultures. We train our poly-omic method using gene expression data from varying conditions and associated reaction rates in metabolic pathways, reconstructed in silico. The poly-omic reconstruction is performed by generating a set of condition-specific metabolic models, specifically optimised for lactate export estimation. To validate our approach, we compare predicted lactate production with experimentally measured yields in a cross-validation setting. Importantly, we observe that integration of metabolic information significantly improves the predictive ability of the model when compared to gene expression alone. Our poly-omic method can therefore accurately predict whether CHO cells have optimal conditions for producing target therapeutic compounds and represents a promising tool for the optimisation of the culture engineering process.