Published in American Society for Microbiology journal mSystems, the study details how AI and machine learning can be leveraged to improve safety and quality control in the food and beverage space.
The researchers used raw milk to test if AI and shotgun sequencing data could be used to correctly identify abnormal milk samples, such as milk that contains antibiotics, from regular ones. AI would do this by analyzing the type and amount of bacteria present in milk samples – the so-called milk microbiome - similarly to how the risk of illnesses like type 2 diabetes can be predicted with AI by analyzing the human gut microbiome.
Proving that AI can be used for screening in this context would be significant, as traditional methods for anomaly detection in the food industry are limited.
In addition to proving this concept, the cohort – which includes researchers from Pennsylvania State University, Cornell University and IBM Research – also looked if AI could predict the different stages of milk processing, transportation and the season it was milked in by using affordable and publicly available datasets.
To carry out the experiment, the researchers collected 58 bulk tank milk samples to establish baseline samples of the raw milk microbiome; they found that 33 microbes had a stable presence in raw milk, with Pseudomonas, Serratia, Cutibacterium, and Staphylococcus being the most abundant.
Already here the researchers confirmed that traditional methods such as cPCA and MDS were limited in their ability to differentiate between sample classes – yet AI was able to do so while also identifying microbial drivers that separated sample classes.
The study then applied explainable AI to 16S rRNA data - an affordable alternative for whole-genome shotgun sequencing metagenomics - from two publicly available milk microbiome datasets to see if the tool could differentiate between different categories of milk, including transport stage and processing stage.
Why detecting rogue ingredients in food is challenging
Traditional methods for anomaly detection in the food and beverage space include alpha and beta diversity, differential abundance, clustering, contrastive PCA (cPCA) and multidimensional scaling (MDS) – but none of these can fully and accurately separate sample classes, e.g. anomalous sample from a baseline sample.
Machine learning tools, however, have been successfully used to sample the gut microbiome, for example to predict the risk of illnesses such as type 2 diabetes. Conversely, the study detailed here tested if machine learning could detect and identify anomalies in the milk microbiome.
Machine learning techniques managed to predict processing stage, e.g. which milk sample was pasteurized, by correctly detecting the types and abundance of bacteria within. In addition, ML also produced accurate models regarding the stages of milk storage, identifying which samples were of raw milk, tanker milk or silo milk.
Further, the AI-powered screening tools also identified the season in which a milk sample was collected by measuring the abundance of mycoplasma.
“To the best of our knowledge, this study characterized raw milk metagenomes in more sequencing depth than any other published work to date and demonstrates that there is a set of consensus microbes that were found to be stable elements across samples,” the researchers concluded in the paper.
“We demonstrate that our explainable AI approach is able to successfully predict the processing stage and the transport stage a milk sample comes from. This study provides advances in the application of machine learning that can be expanded across the food industry.”
Source:
Development and evaluation of statistical and artificial intelligence approaches with microbial shotgun metagenomics data as an untargeted screening tool for use in food production
Authors: Ganda, E., et al
DOI: https://doi.org/10.1128/msystems.00840-24