Reimagining the Synthetic Biology DBTL Cycle with Machine Learning

Abstract

While progress has been made in developing synthetic regulatory circuits for a variety of applications, circuit engineering remains highly unpredictable due to the inherent complexity of the intracellular medium. With recent advances in high-throughput molecular biology, we are witnessing early progress in the use of ML/AI tools for overcoming this complexity and accelerating circuit engineering. The use of ML/AI for circuit design is not currently limited by a lack of computing power or model architectures-it is held back by the lack of high-quality training datasets. One recently reported pipeline, CLASSIC, is a platform for building expansive circuit libraries via hierarchical DNA assembly, and assaying them using a combination of next-generation long- and short-read sequencing. The development of CLASSIC has not only illuminated best practices for high-throughput data collection and model training but also defined principles for incorporating ML/AI into the design/build/test/learn cycle for synthetic biology projects. Here, we detail experimental methods, computational methods, and strategies for using CLASSIC to construct circuit libraries, acquire data sets, and train ML/AI models We highlight generalizable principles that can be applied to data-driven circuit design projects.

Keywords: Artificial intelligence; Bioengineering; Genetic circuits; Genetic engineering; High throughput; Library screening; Machine learning; PyTorch; Synthetic biology.