Paul Jensen, Assistant Professor at University of Michigan Biomedical Engineering, and his graduate students have created an artificial intelligence agent that uses game-playing robots to answer scientific questions. BacterAI can assign autonomous scientific experiments for robots that eventually lead to answers that would normally take humans years to answer. Their Deep Phenotyping system has completed 931,038 automated experiments since January 2020. (Image: Marcin Szczepanski, Michigan Engineering)

A team led by University of Michigan assistant professor Paul Jensen, who was at the University of Illinois when the project began, has created an artificial intelligence system to enable robots to conduct autonomous scientific experiments — upwards of 10,000 per day — potentially driving a drastic leap forward in the pace of discovery in applications from medicine to agriculture to environmental science.

That artificial intelligence platform — BacterAI — mapped the metabolism of two microbes associated with oral health — with no baseline information to start with. Bacteria consume some combination of the 20 amino acids needed to support life, but each species requires specific nutrients to grow. The team wanted to know what amino acids are needed by the beneficial microbes in our mouths so they can promote their growth.

Jensen sat down for an exclusive Tech Briefs interview — edited for length and clarity.

Read now to see how BacterAI fared — and what makes it tick.

Tech Briefs: What inspired your research?

Jensen: We’ve been interested in tooth decay for quite a while. Tooth decay is caused by a bunch of bacteria in your mouth that overgrow and start producing lots of acid from the sugar that we eat. We wanted to develop computer models of these bacteria and figure out what makes ‘em tick and what makes ‘em grow, but we needed a lot of experimental data to do that. So, we started building some robotic systems that would do these growth assays to try and figure out what it is that these bacteria need. The bacteria in your mouth need a lot of strange nutrients in order to survive. I like to use the analogy that they’re kind of like children who — when they first leave home and have never cooked for themselves — don’t know how to make anything because they live in our mouth and we feed them three times a day. They end up having all these strange nutrient requirements that other free-living bacteria don’t.

So, we started with all the robots and then we built a really high-throughput robotic system that could do thousands of experiments a day. And we quickly found that the robots worked faster than the humans did. So, we couldn’t come up with new experiments for the robots to do every day, and we couldn’t process all the data that the robots produced every day. The robots were always waiting on the scientists in the lab to figure out what’s the next step? What does the last set of experiments mean? The robot sat, not moving most of the time, and we decided that the only way to use the robots full-time was to get humans out of the loop.

We started to develop an AI system that could take the results from one day and think about it and come up with what would be the next best experiment to do. And as soon as we did that, then the humans were out of the picture, and we found that the robots can just run by themselves. They don't need us for any planning; they’re actually very efficient. They’re better at picking experiments than we are.

Tech Briefs: I’m sure there were too many to count, but what was the biggest technical challenge you faced throughout the work?

Jensen: There are a lot of issues with quality control. When we humans finish doing an experiment, we can take it back at the end of the day and look at it and kind of think, ‘Did this work? Is this what I expected? Did something go wrong when all of this was getting set up?’ There’s a lot of intuition that goes into that, and we had to automate that as well, because otherwise we would spend all of our day checking 10,000 experiments every single day. So that was the bigger challenge — actually doing the learning and figuring out what experiments should come next to figure out if the previous day’s experiments were actually correct. Did something go wrong? Because the robots know very little about all the other things that can happen biologically. They just do exactly as they’re told, and they don’t really think about other things that could have gone wrong before the bacteria reached them or when we’re making the media that they’re using. All of those things needed to be checked up automatically.

Deep Phenotyping system has completed 931,038 automated experiments since January 2020. (Image: Marcin Szczepanski, Michigan Engineering)

Tech Briefs: Can you explain in simple terms how the technology works?

Jensen: The technology is the same type of AI that’s used for gameplay — learning how to play chess, learning how to play video games, everything from Pong to StarCraft uses this same type of artificial intelligence. And the idea is that, if you think about playing chess, the way this AI works is it starts by just moving pieces randomly. It doesn’t have any strategy; it doesn’t really know what it should be doing. It just makes random moves, and then you play an entire game and see if you win or lose. And if it wins, the computer says, ‘Ah, I made some good moves during that game.’ If it loses, it says, ‘Those moves probably weren’t too good, so the next time I play a game, I’m either going to try and repeat the good moves or try and avoid the bad moves.’

It seems like a very simple strategy, but if you do this over and over and over again, eventually the computers can figure out patterns of good moves, right? It just takes a lot more gameplay; humans normally don’t have the patience. When you’re learning to play chess, you don’t play tens of millions of games. We do the same thing with the robots.

Every day the AI system makes a bunch of moves, saying, ‘I want to grow bacteria in these conditions.’ And it also makes predictions, saying, ‘I think it will grow in this set of conditions and it won’t grow in this set.’ And then the robots run the experiment and the result comes back and says, ‘These were the conditions where you were right and these were the conditions where you were wrong.’ And it learns from that in the same way.

Tech Briefs: The paper I read quotes you as saying, ‘Understanding how bacteria grow is the first step toward re-engineering our microbiome.’ So, what’s the next step? Are there plans for more research, further testing, development, etc.?

Jensen: The bacteria we’re studying live in your mouth; there are hundreds of different species of bacteria that live on the surface of our teeth, and some of them are bad, some of them cause cavities, and some of them are good. What we’d like to do is make it so that the good bacteria can grow and the bad bacteria can’t grow. If we do that, we can avoid people getting cavities.

One of the ways to do that is by supplementing what goes into your mouth; you could add things to toothpaste, you could add things to mouthwash, things that favor the growth of the good bacteria and inhibit growth of the bad bacteria. The first step was figuring out what could those things be — what are things that the bad bacteria don’t like and the good bacteria would like?

The second way would be figuring out what’s the delivery strategy. How much should we put in? What and when should we put in? That becomes a really difficult question when you talk about hundreds of different bacteria in this community that you need to balance. We’re going to learn more about the other bacteria besides the two that we did in the study, and then also start to put communities of bacteria together and try and learn their dynamics as a whole.

Tech Briefs: Do you have any advice for engineers aiming to bring their ideas to fruition?

Jensen: I think the biggest thing is that technology is changing so rapidly that it’s OK to start on a project even if you don’t have all of the pieces figured out yet. If you know it’s going to be a multi-year project, a lot of things are going to change by the time the project finishes. In our case, we didn’t have a plan for how we were going to run all of the robots, but we knew we could probably figure that out by the time we got the robotics system built.

So, it’s OK to do that — to sort of dream big while there’s still problems on the horizon for which we don’t have a solution yet but get going now and either we or someone else will come up with a solution that we can repurpose to solve our problem too.

Tech Briefs: Is there anything else that you’d like to mention?

Jensen: This project took a lot of different expertise; we had microbiologists and we had computer scientists and then engineers and roboticists all working together. I think that’s going to become much more common — groups of interdisciplinary people or people who work and have expertise in more than one area of science and engineering are going to be in high demand in the future.