Trillions of experiments per year…
Welcome to the Robotics Roundtable
We take an article that piqued our interest and discuss it from our unique perspectives. Sevy - the mechanical engineer, Connie - the software engineer, and Sean - the social entrepreneur and finance dork.
If you enjoy, consider subscribing to our newsletter where you can read more about our thinking, the latest robotics news coming out of Los Angeles, and follow our journey as we work to unite robotics in Los Angeles
Best, - Connie, Sevy, Sean
Article Summary
An artificial intelligence system, known as BacterAI, has been developed by a team led by a professor from the University of Michigan. This system enables robots to conduct up to 10,000 autonomous scientific experiments per day, potentially accelerating the pace of discovery in fields such as medicine, agriculture, and environmental science. In a recent study published in Nature Microbiology, BacterAI successfully mapped the metabolism of two oral health-associated microbes, providing valuable insights into their growth requirements. By testing hundreds of amino acid combinations daily and learning from the results, the AI system achieved a 90% accuracy rate in predicting the amino acid needs within a short period. This approach of automated experimentation has the potential to significantly speed up scientific research in various domains. The research was funded by the National Institutes of Health with support from NVIDIA.
Sean’s Corner
Trillions of experiments per year…
Few things are dearer to my heart than accelerating science. But scientific progress is hampered by limitations (you can get an overview here). BacterAI may be a part of the fix.
How?
In combination, AI & robotics can be recursive. Like us, they learn from their learning, compounding their intelligence. Unlike us, they can work continuously, consider massive possibilities sets, at digital speeds, at tiny time & space scales, and work in massive parallel. These abilities allow them to move faster with less resources.
As a result, I predict tectonic shifts based on tech like BacterAI. For example, I believe we will see new business models.
Imagine massive clusters of BacterAI-like systems living in buildings away from expensive University campuses running ongoing experiments in parallel. It’s easy to envision these systems democratizing access to research for countries with cheap energy and tons of land - e.g., the Horn of Africa & KSA (see this New Yorker article for the dangers of such expansion).
Companies like Science Exchange, and science automation companies https://opentrons.com/ and https://strateos.com/ are already harbingers.
Here’s a back of the napkin calculation, there are currently ~18,000 biolabs in the US, assume 10 researchers per lab, 5 research projects each quarter means per year, the US runs roughly 3,600,000 bio-experiments.
The average 40 ft tall warehouse in the US has a total usable volume of 7.2M cu ft. If BacterAI is roughly 9 cu ft, ~2.4M BacterAI systems fit in that space.
The UofM (#GoBlue) system claims to be capable of 10,000 per day, but also suggests a million a year. Let’s use the more conservative number: 1,000,000 experiments per year.
With one system creating 1,000,000 experiments per year, that’s 2.4 TRILLION experiments per year per warehouse.
That’s ~670,000x the output of the entire biology infrastructure in the US - from one average sized warehouse.
I might be drunk on science potential, but that might be more than all biology experiments in human history.
And it gets better…
The above BON calculation assumes no optimization of the BacterAI system (of course, I’m using the BacterAI system as a proxy for any AI-Automation science system). Imagine the potential if systems are tailored to work in parallel.
That setup would facilitate “an AWS of science” where undergrads, grad, researchers, and postdocs spool up research on demand, allowing their AI-Scientist to determine the number of BacterAI clusters needed. This technology can significantly reduce research cost and risk - forcing a funding model change.
And it gets even better…
Success in biology will spur systems for chemistry, material science, physics, complexity, etc. A robotic automation system like BacterAI portends a future with more research conducted per day than in all previous human history, across subjects.
Connie’s Corner
The goal of BacterAI is admirable. Is it great to see machine learning methods breaking out of pure computer science and being applied to biological sciences, and a great precedent for similar applications. In particular, it introduced a method that looks to have wide applicability, as “BacterAI could be replaced with Gaussian processes, logical induction systems, or other machine learning techniques.” The potential for other sciences to pick up a computational learning approach could really speed up scientific progress. However, I do feel that the article and paper were presented too optimistically.
However, I do feel that the article and paper were presented too optimistically .I will say I am not an expert in machine learning nor biology, so take my opinion with a grain of salt. I don’t necessarily agree that there was “no baseline information to start with”. The Markov Model used has all 20 amino acids. While that creates a lot of permutations, it is taken for granted that there are only 2^20 permutations, and that all of the inputs are known, with a fairly simple discrete output. In addition, while the first S. gordonii, a well-known bacterium, had a neural net that started untrained, its model was transferred to a second bacterium, giving the second bacterium baseline information. The jump from two bacteria, to that millions of experiments on that other unstudied 90%, is quite a leap. I don’t think that this is going to revolutionize biology.
Sevy’s Corner
First to give you some context on my background understanding of microbiology and AI, I know the broad ideas but not the intricacies.. That being said, the frontier of chemistry and biology are similar in that they involve a lot of experimentation. Experimentation combining several different chemicals or in the case of BacterAI different amino acids. Microbiologists have guesses in what will work but there is always the option to throw quite a bit onto the wall (a.k.a. mix many different combinations) to see what sticks. The exciting part of this article is the speed of the automation that is happening with laboratory robots . More robots means more throwing against the wall, more time humans can think through the data, more breakthroughs, and ultimately more understanding and useful solutions in medicine and beyond. We still have so much to learn. The AI component adds a way to speed up testing another way. I think there is no replacement for a good test but that we can have AI give us better predictions of what to test, speeds up breakthrough combinations sooner just like the physical robots.
What I would find even more interesting if the AI results could feed back into the physical robots to do tests automatically. That would create this cycle where the work could be checked and the model further refined.
Synthesis
As a team, we agree that a system like BacterAI has the potential to significantly enhance the pace of scientific discovery. However, we are excited about the progress already made by the team. We expect even more output with optimized mechanics and software.
However, we disagree on the potential for impact given the current size limitations and machine learning model. Sean is enthusiastic about the potential. Sevy is neutral, with a slice of optimism. Connie is negative and believes BacterAI might not change anything at all. Regardless of our individual opinions, we are all excited to see work towards accelerating science discovery.
We hope you enjoyed this, if so or if not, you can let us know at welcome@lalovesrobotics.com