September 23, 2025 – Biodiesel, Bioplastics, Penicillin: “When you look at key industrial bioprocesses, most people use organisms that make products naturally,” says Jeff Chaika, bioprocess scientist and LINUS PAULING fellow at the Department of Energy (DOE) Northwest Research Institute (PNNL). “There were not many successful engineering stocks for the bioprocess to create new or better products.
The reason for this is a story that is too familiar to many researchers. They have been carefully engineering over the years as they aim to improve biological organisms. It works perfectly in the lab, but suddenly fails when expanded for industrial applications.
It is extremely difficult to accurately predict how a living organism's properties will respond not only to its own genes, but also to its environment (early fields known as “predictive phenomics”). Even microorganisms have an incredible number of factors that affect their properties. In the case of biological innovation, that is, when tension fails, it is often almost impossible for researchers to isolate the problem and properly isolate the course.
“This is a major challenge facing synthetic biology researchers and businesses,” says Czajka.
At PNNL, researchers use AI to tackle the challenge.
Arts and Sciences of AI-Accelerated Bioprocess
Bioproductive researchers often want to engineer biological processes to increase power output. The higher the production, the more biofuels and more antibiotics, and more antibiotics, at the cost.
In one case, PNNL researchers had already designed the Lipomyces starkeyi. Yeasts commonly used in biofuel production produce Morinic acid, a platform chemical with a wide range of uses. Next task: Increase output.
“We were trying to optimize the media to produce as much molyic acid as possible what the yeast was growing,” Czajka said. “But there are many parameters. These yeasts are sensitive to all kinds of environmental conditions and it would have taken a very long time to explore all the options.”
“Right now, the main approach to optimization is trial and error,” he said. “This leads many people to say that the field of bioprocessing is more than science.”
However, Czajka and his colleagues applied a different kind of art. It is a machine learning model called an automated recommended tool (ART). This model was developed by researchers from the Lawrence Berkeley National Laboratory and Sandia National Laboratory through the Joint Bioenergy Research Institute (DOE Bioenergy Research Center) and DOE's Agile Biofoundry (a consortium of national laboratories dedicated to accelerating biomanufacturing). PNNL is a member of both groups.
“Art basically designed an experiment for us,” Czajka said. “When you repeat the experiment, Art learns from the results and suggests the next condition to test and which conditions are the worst.”
“In the end, we improved our production by about 20% in a short time, which was very exciting.”
The best of both worlds
AI models are excellent at predicting stock performance, but are often limited by the low availability of accurate molecular measurements that inform them. This is a phenomenon known as sparse data. Moreover, even highly predictive AI models often generate “black box” solutions.
In these cases, PNNL researchers apply a combination of AI and more traditional genomic scale models. This incorporates the scientific and mechanical understanding of organism function by researchers.
“Notifying machine learning models with mechanical insights improves performance and improves explainability,” Czajka said.
Machine learning offers advantages over genomic scale models.
“It's difficult to build accurate and predictive genomic scale models due to the complexity of gene-environment interactions, and there are many genes that we don't know what they're doing,” Czajka said. “A machine learning model can implicitly capture all its complexity.”
By combining them, researchers get the best in both worlds: mechanistic accuracy of genome-scale modeling and predictive punches of machine learning.
For example, Czajka and his colleagues applied this combination approach to predict the output from various strains of Yarrowia Lipolytica, which have a wide range of uses in the production of biofuels and other biological organisms. The team first used genomic scale models to fill gaps in sparse data, and then used the resulting dataset to train a machine learning model that accurately predicts output concentration.
A brighter future for biological development
“Now, you can experience targets for years, learn or engineer tensions, see how it works and try to regain those insights to design them,” Czajka said. “Ideally, these AI-enhanced modeling efforts will help generate new and improved biological organisms while shortening these cycles, saving time and reducing costs.”
The AI-accelerated processes and tools developed at PNNL have a wide range of possibilities to enhance the development of bioenergy molecules and countless other biological organisms. A better designed strain can lead to better medicine, better fuel, better materials, and more.
However, for Czajka, it is not about individual biological organisms.
“In my opinion, this is all under one umbrella: a transition of tensions from the lab to the industry,” he said.
Research in PNNL's AI-Accelerated Bioproducts development is supported by the Department of Energy, the Department of Energy Efficiency and Renewable Energy Bioenergy Technology, and the PNNL's internally funded predictive phenomenon initiative.
About pnnl
Pacific Northwest National Laboratory draws on outstanding strengths in chemistry, geoscience, biology and data science to promote scientific knowledge and tackle the challenges of energy resilience and national security. Founded in 1965, PNNL is run by Batelle and supported by the U.S. Office of Energy Science and Science. The Science Bureau is the sole and largest advocate for basic research in the US physical sciences and is working to address some of the most pressing challenges of our time.
Source: Oliver Peckham, PNNL
