Gemini Co-Scientist predicted the shape of E. coli colonies from images obtained at neighboring substance concentrations
Gemini Co-Scientist predicted the shape of E. coli colonies from images obtained at neighboring substance concentrations
On August 27, the authors posted an arXiv preprint. They gave the system images of genetically modified E. coli colonies grown at several concentrations of IPTG, a substance that activates the gene in this experimental system, while withholding images from one additional concentration. They then compared the prediction with a colony that the laboratory had grown under the withheld condition. For the bacterial strain whose shape changes in response to IPTG, the predicted changes matched the observed direction for three of the four measured features.
Swarming E. coli colonies spread outward from the inoculation point and form a branching pattern on the nutrient gel. In the pLac-rpoS strain, IPTG activated the rpoS gene, which affects swarming. As the dose increased, the colonies became smaller and their radial branches became denser. The control strain, pLac-gfp, maintained the same shape across doses and was used to test whether the program could recognize stability.
In May, Co-Scientist had already proposed genetic factors for testing in cell experiments based on the literature and helped interpret screening results. In the new experiment, it received data from a completed series and was asked to reconstruct the appearance of a colony using only images from the other conditions.
The laboratory grew and scanned the colonies. The authors withheld images from one concentration at a time. Gemini Co-Scientist received images from the neighboring conditions and wrote a program to reconstruct the withheld point. Gemini 3 Pro Image generated 16 colony variants, and Gemini 2.5 Pro selected one of them. The colony physically grown at the withheld dose served as an independent test of the prediction.
The researchers compared colony shape using mean radius, elongation, edge irregularity, and circularity. For pLac-rpoS, the changes in the first three features followed the laboratory data. For circularity, the generated colonies were more circular than the real colonies. For the pLac-gfp control, the program preserved the stable shape.
The researchers defined the task, grew the cultures, and acquired the images. Co-Scientist developed a method for interpolating between the known doses. The authors propose using this workflow for experimental series in which one biological system is tested under many conditions and the researchers must choose which measurements to perform next.