From sequence to yield: deep learning for protein production systems
Recent progress in laboratory automation has enabled rapid and large-scale characterization of strains engineered to express heterologous proteins, paving the way for the use of machine learning to optimize production phenotypes. The ability to predict protein expression from DNA sequence promises to deliver large efficiency gains and reduced costs for strain design. Yet it remains unclear which models are best suited for this task or what is the size of training data required for accurate prediction. Here we trained and compared thousands of predictive models of protein expression from sequence, using a large screen of Escherichia coli strains with varying levels of GFP expression. We consider models of increasing complexity, from linear regressors to convolutional neural networks, trained on datasets of variable size and sequence diversity. Our results highlight trade-offs between prediction accuracy, data diversity, and DNA encoding methods. We provide robust evidence that deep neural networks can outperform classic models with the same amount of training data, achieving prediction accuracy over 80% when trained on approximately 2,000 sequences. Using techniques from Explainable AI, we show that deep learning models capture sequence elements that are known to correlate with expression, such as the stability of mRNA secondary structure. Our results lay the groundwork for the more widespread adoption of deep learning for strain engineering across the biotechnology sector.