Case Study: Machine learning opens new doors in bioproduct development

Recently, the synthetic biology world has looked to machine learning to speed up the bioengineering process and come up with inventive solutions. Conventional bioengineering methods are slow, but machine learning can speed up the process dramatically, transforming large amounts of data into predictions that effectively guide the development of bio-based products.

In 2019, biotech company Lygos teamed up with the Agile BioFoundry (ABF) to optimize strain performance through ABF’s multi-omics and machine learning capabilities

Through a series of experiments, the teams generated over 80,000 data points to train machine learning algorithms. Agile BioFoundry’s team then developed artificial neural networks to leverage this data and provide Lygos with actionable recommendations to increase strain performance through Design-Build-Test-Learn cycles. This unprecedented amount of data will be vital to advance the use of machine learning for cell bioengineering.

Lygos and the ABF plan to iterate on this process throughout the next year, generating valuable information that will be the focus of an upcoming publication. Their hope is that by the end of this process, strain performance will be improved significantly, benefitting Lygos’ commercial interests developing bio-based products. 

The project is impacting Lygos’ overall approach to synthetic biology, particularly when it comes to their interest in producing malonic acid, which can be used in plastics, food additives, and more. 

“Our platform for malonic acid at Lygos is quite well established. Because of that, we’ve gone down many scientific avenues to improve performance,” said Mark Held, senior scientist at Lygos. “When you do that for a long time, you start running out of rational places to further engineer. That is where this multi-omics and machine learning technology shows its strength. It has the ability to find things we can’t easily predict.” 

Held said this is evident given what information they’ve already been able to generate, with the system recommending targets that were counterintuitive in nature. 

“In some cases, the algorithms suggested recommendations contrary to operations that we’ve done in the past,” he said. “For a mature program like our malonic acid program, machine learning has a great deal of application and potential. If we hit the numbers we expect to, it will be an absolute game changer for Lygos.”