Creating a yeast classification model by combining AI and cytometry

The identification of yeast strains is crucial for various applications in the fields of microbiology and biotechnology. Its use can be decisive in fermentation processes or pathogen detection, for example. In this context, Amarok Biotechnologies has succeeded in developing a highly accurate preliminary model for classifying mixed yeast populations. The work was based on advanced technologies, combining imaging flow cytometry with artificial intelligence (AI) to create a statistically significant yeast classification model in mixed populations.

Background and challenges in yeast identification

Traditional identification methods, based on biochemical tests and morphological observations, can be time-consuming and require specialized knowledge. The combination of artificial intelligence and flow cytometry with imaging offers a fast, automated and accurate alternative, enabling detailed morphometric and autofluorescent profiling of yeast cells.

Purpose and experimental approach for combining flow cytometry and AI

The aim was to generate a statistically significant yeast identification model, determined by artificial intelligence using convolutional neural networks or linear discriminant analysis (LDA). This would enable strains to be distinguished using morphometric data and autofluorescence characteristics obtained from high-resolution imaging.

To achieve this, the researchers used an Amnis® ImageStream®X MKII flow cytometer, coupled with Amnis® AI and IDEAS® ML artificial intelligence software. They selected the following five yeast strains: German Ale, B. Brut, Pichia, LG Monaco and 1H.

Two models were combined:

1. The first uses brightfield imaging features, validated by statistical analysis of strain identity and actual class.
2. The second is based on the Linear Discriminant Analysis ( LDA) algorithm, while combining autofluorescence measurements.

The analysis model was created by collecting samples of morphometric features obtained by individual analysis of each strain. The addition of individual autofluorescence and size characteristics (surface area in μm²) removed ambiguity as to which yeast strains had similar morphometric characteristics.

Key results and impact of the yeast classification model

The “super-parameters” calculated enable maximum separation between different strains when analyzing mixed populations. The model’s accuracy rate reaches 97% (weighted average). This indicates good prediction of strain class, since statistical performance was evaluated at 95% and no errors were detected by the AI.

The combination of these advanced techniques has enabled us to assign each strain a unique “multi-parametric digital signature”, based on its most significant characteristics. The result is a precise classification even in cases where morphometric similarities could lead to confusion. This model therefore enables strains of different sizes and aggregation states to be identified without significant error.

Future developments and potential applications in microbiology and biotechnology

This AI model can be re-trained by adding additional strains and parameters such as specific fluorescent antibody markers (Bretta test), viability staining, etc. It has good adaptability, enabling it to be used in a wide range of applications. It is highly adaptable, enabling its use to expand. This technology could therefore revolutionize yeast analysis in research and industry, enabling the identification and monitoring of yeast populations with minimal human intervention.