Researchers at the Agile BioFoundry (ABF) and Washington University in St. Louis are using machine learning to improve biomanufacturing in non-model yeasts.
Oleaginous yeast species that naturally produce microbial oils are useful as biomanufacturing hosts. However, efforts to engineer these non-model yeasts can be time-consuming and expensive due to limited knowledge of their metabolic networks.
The ABF’s collaboration with Washington University in St. Louis aims to combine knowledge mining, feature extraction, genome-scale modeling, and machine learning to predict the outcomes of non-model yeast biomanufacturing under complex genetic and fermentation conditions. The project, which also includes researchers from Rensselaer Polytechnic Institute and Lincoln University, aims to create a machine learning-computational strain design platform that can predict the fermentation performance of three non-model oleaginous yeast species.
“Machine learning can be used to simulate complex systems that are difficult to explicitly and analytically model,” said Yinjie Tang, professor at Washington University in St. Louis. “This project will use ABF capabilities to advance machine learning applications for predicting non-model yeast factory performance and guiding metabolic engineering.”
Machine learning approaches in biomanufacturing require significant amounts of quality training data. The team will first gather data on three non-model oleaginous yeast species — Yarrowia lipolytica, Lipomyces starkeyi, and Rhodosporidium toruloides — sourcing from published papers and the ABF’s extensive research in yeast engineering. To reduce the burden of data collection, the team is using GPT-4, the multimodal large language model created by OpenAI, to streamline the feature extractions. This effort will also utilize omics data from ABF’s Experiment Data Depot and high quality flux features generated by ABF mechanistic models.
As a part of this effort, ABF researchers are leveraging a newly constructed genome-scale metabolic model of Lipomyces starkeyi that provides useful information on the species’ metabolism. These insights will be incorporated into the project’s machine learning approaches.
Working together with ABF project co-leads Kyle Pomraning and Di Liu, the team will then identify features in the collected data that are suitable for machine learning, leading to the development of a structured database.
This data will also become part of the Impact Database, a publicly accessible platform to store biomanufacturing-relevant data from fermentations. The platform has web interfaces for open-sourced machine learning models for titer prediction and strain design.
The team will use the collected data to train a machine learning model to provide enhanced fermentation predictions. They will also use computational strain design tools that can suggest combinations of genetic modifications. This integrated approach aims to reduce the number of experimental tests.
In collaboration with Rensselaer Polytechnic Institute professor Mattheos Koffas, the researchers will demonstrate the effectiveness of this platform by using it to engineer the three non-model yeasts to produce flavonoids and butanol. Strain engineering for these experimental validations will leverage ABF’s genetic tool capabilities. The team anticipates that the machine learning-computational strain design platform will significantly improve strain design and reduce costs.
“This project aims to provide the biomanufacturing community with an integrated knowledge base and AI platform that predict both genotypes and bioprocess conditions for accelerating microbial biomanufacturing.”
Yinjie Tang, professor at Washington University in St. Louis
This collaboration resulted from a 2022 funding opportunity between the U.S. National Science Foundation and the U.S. Department of Energy’s Bioenergy Technologies Office.
ABF capabilities utilized in this project include: