AutoWrinkleID: a machine learning pipeline for biofilm wrinkle identification and quantitative analysis

Biofilms, which are prevalent in both natural and engineered systems, consist of microbial communities encased in a self-produced extracellular matrix. This extracellular matrix can lead to the formation of macroscopic wrinkles on agar substrates due to the mechanical stresses experienced during growth. Biofilm wrinkles are of interest because they may convey information about the growth and structural development of the biofilm, and even have functional roles.

Traditionally, detecting wrinkles in biofilm images has been a labor-intensive process that relies on manual annotation by human experts. This method is not only time-consuming and not scalable, but it is also highly subjective, as different annotators may interpret the wrinkles differently. To address these challenges, researchers have developed a machine learning pipeline called AutoWrinkleID, which automates the detection, characterization, and quantification of biofilm wrinkles across a variety of bacterial species and strains.

The name AutoWrinkleID reflects the automated identification of wrinkles, followed by their detailed analysis. The pipeline starts with bright field images of the biofilms and proceeds to identify the wrinkle structures, characterize them, and quantify their extent. The masks used to delineate the wrinkle regions are generated in two ways: through manual annotations by experts and by using specific fluorescence markers.

These fluorescence markers include constitutive fluorescence, which is inherent to the bacteria, and motility fluorescence, which is induced by a specific stimulus.

The researchers discovered that models trained with data annotated using motility fluorescence markers generally outperformed those trained with manually annotated data or constitutive fluorescence markers. This suggests that motile bacterial phenotypes are closely linked to the formation of wrinkles. To further validate these findings, the team conducted a comparative analysis between masks generated using motility fluorescence and those derived from constitutive fluorescence.

The results consistently showed that the motility-based masks were superior across all bacterial strains examined.

In addition to developing the AutoWrinkleID pipeline, the researchers conducted computational analyses to evaluate its robustness and limitations. They examined how the quality of the training dataset and the choice of fluorescence markers influenced the performance of the wrinkle analysis. Notably, they found that Sholl analysis, a technique originally used to study neuronal dendrites, proved effective for characterizing the unique patterns of biofilm wrinkles.

However, the applicability of Sholl analysis is constrained by the shape of the wrinkles and the quality of the masks used in the analysis.