Over the past few years, companies in food tech product development have begun to utilize machine learning and other AI techniques to accelerate the development of their products. One of those companies is Biocraft, a company focused on developing pet food utilizing cultured meat as its primary protein input.
The company announced in May they would focus exclusively on B2B (they had previously been developing a consumer-facing product under the brand Because Animals), and this week started talking about how they are utilizing AI to assist in product development.
I sat down with Biocraft CEO Shannon Falconer and AI lead Chai Molina to learn more about the company’s AI and the future direction of the company.
Tell us why you decided to investigate how AI could help you develop cultivated meat.
Falconer: My background is my PhD is in biochemistry, and so mostly I was working on drug discovery and antibiotic research. And you know, when AI really hit the pharmaceutical industry in a meaningful way about a decade ago, it dropped the time and the cost of bringing a drug to market, so I’ve been very bullish on integrating this technology into what we’re doing for cultured meat.
I asked Chai if there are any types of tools that are available or that could work for us to actually do what we want to help in dropping our costs, and getting the right ingredient and nutritional profile of our products. Chai looked around and said, “No, there is not.” And so it was then really that we decided if there’s nothing available that we can purchase and use, then we’ve just got to build this ourselves.
Molina: I come to this with a view that this is a mathematical problem, that we just have to find the connections between kind of modeling out how human reasoning sort of works and connecting the dots between pieces of machinery in the cell. To try to understand how we can tweak this Rube Goldberg machine. How we can push it into the direction that we want it to go.
How did you start building the AI model?
Molina: There’s a machine learning component that is along the lines of natural language processing, where we collect our data from lots of publicly available papers and databases. From there, we process the data and basically build out a picture of the machinery inside the cell.
What do you mean by that?
Molina: These databases and papers might show a tiny glimpse of one piece of that machinery inside a cell. In a way, we’re superimposing little pictures and little parts of that machinery to build out the bigger picture. From there, we try to understand if you pull this cable or take this step, what’s it gonna do? There are all these threads of biochemistry in the cell, I like to think of it like dominoes where you push one, and then you see downstream effects. And so that is more of a mathematical modeling approach, involving network theory.
You’re using the analogy of a machine to describe a cell and understand what the domino effects of a certain action or input within a given hypothesis about that cell.
Molina: Yes. Once we have a picture of the machinery in the cell, it’s like, okay, ‘what can we how can we tweak that to make it do what we want?’ Say we want to add a novel medium component for a growth serum for the cells that will hopefully push them in the direction that we want, such as cell proliferation. So, for example, we look at different substances that are safe for consumption and ask how would the addition of these things at least qualitatively impacts the machinery in the cell.
And you’re running these hypotheses in the AI and then testing out promising results in a wet lab?
Falconer: If you’re a wet lab scientist, and you generate a hypothesis, there are so many things to test. Especially when you’re working on something as complicated as media optimization in order to achieve the right cocktail that will elicit proliferation as well as the nutritional profile that you are that you want that you desire. And so the time that it would take to perform all these various experiments empirically, not only of course, is very lengthy and very expensive. And so what this tool does is it allows us to trim down that list of experiments. This tool is able to prioritize for us and give us sort of a ranking order as to which hypotheses are more likely to succeed or fail. And so this shortens the time and the number of experiments.
And then and then the other thing that it does for us is, it allows us to actually get better at identifying sort of the unknowns. What this tool can do is it can identify, say, anywhere between, say, A and Z -anywhere along this line where a human brain cannot read and put into place all of the different connections – what might ultimately elicit the end desired effect. We can then go back and say, Oh, but we now know that five nodes upstream in these completely disconnected papers, we see that this domino will hit this one, and then this one hits this one, etc. And then we can actually achieve this desired effect down the road.
You announced last May you were becoming a B2B company exclusively, and you were sunsetting your CPG products. How has this new focus, combined with the AI development tool, changed your product development speed?
Falconer: Yes, so now we’re exclusively a b2b company, focused on delivering volumes and working with existing pet food manufacturers who already have that massive consumer base and who can disseminate product quickly as soon as we have it available to sell it. And so that’s what we’re focused on right now. I’d say over the past 12 months, with just focusing on this product development. I think we’ve made probably more progress in 12 months than we did in five years. And a big part of that is the development of our AI platform.
If you’d like to learn more about how AI is accelerating next-generation food development, join us October 25th at the Food AI Summit.