How EcoBOT’s automated plant lab knows where it’s guessing 

Establishing causal relationships in plant-microbe studies can be exceptionally difficult, as it requires juggling sterility, standardization and manual labor. A single contaminant or inconsistent measurement can render weeks of work unusable. For example, one study on the effects of plant genotype and soil microbiome on growth in Lotus japonicus found that 15% of variability in shoot length was due to variability in the pots, complicating their conclusions. EcoBOT, an automated plant lab at Lawrence Berkeley National Laboratory, is designed to solve these problems. 

 EcoBOT uses uncertainty quantification, a statistical approach that estimates the confidence level behind each prediction, built on Gaussian process models, which forecast outcomes along with an error margin, and Bayesian optimization, a method that uses those predictions to select the most informative next experiment to run. The team published the EcoBOT methods paper in Frontiers in Plant Science. 

“It’s a powerful way to quantify uncertainty given a particular data set… Based on what you know from these early experiments, [it tells you] what you should do next, and it’s well enough constrained and robust enough, mathematically, that you can pretty much rule out large mistakes like a hallucination,” Marcus Noack, a research scientist in the Applied Mathematics and Computational Research Division at Lawrence Berkeley National Laboratory (LBNL), said in an interview with R&D World. 

 Unlike a large language model, which produces a fluent answer without a reliable measure of its own confidence, the AI in EcoBOT runs as an active-learning loop. The system looks at the results of completed plant growth experiments, quantifies how confident it is about different regions of the experimental space and recommends the next most informative experiment to run.  

EcoBOT uses Bayesian Optimization to select experimental conditions using a variance acquisition function, meaning it picks the next test specifically to reduce uncertainty in areas where the model is least confident.  

“It’s quite interpretable, in that you can follow up on what these decisions meant and why they were made,” Noack said.  

How researchers validated the system 

The researchers used EcoBOT to validate responses in Brachypodium distachyon to nutrient limitation and copper stress, with Bayesian optimization improving model accuracy by more than 30% between experimental rounds.  

Plants dosed with 500 micromolar CuSO4 confirmed significant inhibition of both root and shoot biomass. Results from 40 plants across seven copper concentrations were used to build an initial Gaussian Process model relating copper concentration to plant weight.  

The model’s own uncertainty was used to select five new concentrations for a second round targeting the concentration ranges where the first model was least confident. Those five concentrations were retested, along with the 0 and 500 micromolar conditions.  

The researchers compared model accuracy before and after incorporating the second round of data, using 30 test-train splits. They found that root fresh-weight models improved 34.9% and shoot fresh-weight models improved 31.6%.  

They also checked for batch effects and found no significant difference in control-plant biomass between the two rounds.  

A separate nutrient-deprivation test confirmed that the system reliably detected expected plant stress responses.  

The researchers note that, on sparse datasets with limited experimental cycles, their own prior research shows Bayesian optimization “does not always hold a significant advantage over other methods including random sampling, however it cannot perform worse.” They state that “other methods would likely have worked comparably in this instance.”  

Where EcoBOT still needs human input 

While EcoBOT uses AI to power automated experiments, it still requires human input.  

“Determining and using judgment to assess whether something is real or an artifact is something a person, at least now, and I’d imagine for the foreseeable future, is better at doing,” Peter Andeer, a research scientist at LBNL’s Environmental Genomics and Systems Biology Division (EGSB), said in the interview.  

In fact, EcoBOT was designed to keep experts in the loop, not eliminate human involvement.  

“You don’t want to take people out of the loop, because you want to have a person who’s an expert in the field go in, look at the data and make sure everything checks out,” Andeer said.  

The road to commercialization 

The team built EcoBOT around a liquid-handling robot manufactured by Hamilton, but it is not currently commercially available.  

“We’re continuing to work with the manufacturer of the liquid handling portion of the robot, which is Hamilton… I think if anybody were going to commercialize it, it would probably be them, because they already make most of the components,” Trent Northen, EGSB deputy division director at LBNL, said.  

“We had a vision of these things being in many different locations, and this could be a really nice way to standardize plant microbiome research,” he said.