(Invited) Rapid and Robust Discrimination of Food-Contaminating Microorganisms Guided By Machine Learning

Abstract

Interests in safety and security of foods are more growing than ever with rising global populations. Manufacturing processes for food production need to be strictly managed to exclude the possible contamination by toxic microorganisms including bacteria and fungi. Upon occurrence of microbial contamination, the species of the contaminating microorganisms should be rapidly discriminated to limit further expansion of contamination. Pre-testing of microbial contamination of food ingredients prior to manufacturing processes is also carried out to ensure that specific toxic microorganisms are not detected from the ingredients. Conventionally, sequencing of the barcode regions in the microbial genomes including the genes encoding 16S and 18S rRNA (for prokaryotes and eukaryotes, respectively) has been widely employed for microbial discrimination. More recently, matrix assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF-MS)-based discrimination method emerged. However, these approaches have several drawbacks such as requirements of expensive equipment, highly skilled operators, and relatively long assay time for several days. Therefore, it is not easy to employ these methods in food manufacturers, majorities of which are small-sized companies. To address this issue, we have developed a novel bioimaging-based approach for discrimination of bacteria and fungi, termed “colony fingerprinting”. In colony fingerprinting, bacterial or fungal colonies were formed on transparent agar media. The colonies were irradiated by light emitting diode (LED), and complementary metal-oxide-semiconductor (CMOS) image sensors are used to detect the light penetrating or dispersed by the colonies which generate microbial species-specific optical patterns, termed “colony fingerprints”. We extracted a number of discriminative parameters representing the features of morphology and intensity-distribution of the colony fingerprints with the aid of bioimage informatics tools. Discrimination of microbial species were carried out by machine learning-based analyses of the extracted parameters. As a proof-of-concept study, we developed a colony fingerprinting platform consisting of a CMOS image sensor (pixel size: 3.2 μm, imaging area: 6.55 × 4.92 mm2), pinhole, and blue LED. Colonies of 5 closely related Staphylococcus spp. (S. aureus, S. epidermidis, S. haemolyticus, S. saprophyticus, and S. simulans) were visualized, and 14 types of discriminative parameters such as roundness, solidity, Zernike moment, and image entropy were extracted using MATLAB software. Simple linear discrimination analysis (LDA) using the 14 parameters classified the 25 colony fingerprints of 5 Staphylococcus spp. (a total of 125 colony fingerprints) into correct species with 79.4% accuracy. Further analyses with powerful machine learning approaches including artificial neural network (ANN), support vector machine (SVM), and random forest (RF) resulted in the accuracies of 99.2%, 98.4%, and 100.0%, respectively. It was possible to obtain the colony fingerprints used for discrimination within 11 h on average. This result suggests that colony fingerprinting is promising for accurate and rapid discrimination of closely related bacterial species. We successfully demonstrated the discrimination of 20 bacterial species (5 gram-negative and 15 gram-positive bacteria) based on colony fingerprinting by employing up to 65 discrimination parameters obtained by a size-dependent manner. Therefore, accumulation of discriminative parameters could be useful to improve the discrimination accuracy.Besides bacteria, fungi are another major cause of food contamination. However, colony fingerprinting-based discrimination of fungi is challenging because fungi extend the hyphae (fungal filaments) to random directions, and branch them at arbitrary positions. These morphological features are much different from bacterial colonies. Therefore, the discriminative parameters employed for bacterial discrimination were not applicable, and a novel set of parameters need to be developed. Given the growth features of fungal colonies, we computed the number of hyphae and their branches, and intensity distribution on the images as discrimination parameters. As a result of SVM- and RF-based analyses with 7 types of parameters, 6 closely-related Aspergillus spp. were distinguished with 95% and 100%, respectively. Discrimination of fungi based on colony fingerprinting was completed within 48 h, which was shorter than time required by conventional fugal discrimination methods including MALDI-TOF-MS, Raman spectrometry and FTIR spectrometry.Given the promising features of colony fingerprinting, we newly developed a practicable platform for colony fingerprinting which equips a line imaging sensor with scanning system enabling the visualization of a number of culture plates, leading to high throughput analysis. We examined accuracy and robustness of colony fingerprinting-based bacterial discrimination by changing the colony formation conditions such as temperatures and cell densities on the culture plates.In summary, discrimination of microorganisms was successfully demonstrated with the machine learning-guided colony fingerprinting. In the future study, development of the colony fingerprint library with hundreds of microbial species will make this rapid and robust system a powerful tool to secure food safety.