Machine learning

Machine learning refers to the use of artificial intelligence models trained on large datasets, and is an important piece of the Nom Nom research toolkit. In addition to the algorithms commonly used by other researchers, we have also developed in-house neural networks for the analysis and classification of microbiome data.

Read classification

Due to the limitations of available databases, it is not always clear whether a particular sequence identified in a microbiome sample belongs to a bacterium, virus, parasite, or the host organism. However, through carefully trained and tested deep learning models,it is possible to predict what domain of organism these sequences originate from with >95% accuracy.


In this project, we differentiated reads from bacteria, dogs, and helminths, parasitic worms that infect the canine gastrointestinal tract. Being able to identify helminths can enable us to better test for parasites that may be leading to health issues in the samples that we receive.

Identification of associations

Risk factors for the development of multifactorial diseases such as obesity may often interact with each other, or otherwise combine in non-linear ways. However, we can use machine learning models to identify relationships that would otherwise be obscured by traditional statistical methods of analysis.


In one analysis currently being prepared for publication, we used two machine learning models to identify ensembles of risk factors for obesity in household dogs. In combination with a traditional regression model, these additional steps confirmed that we were not missing any significant contributing variables in our final analysis. We are also using machine learning to identify lifestyle and diet factors associated with high microbiome diversity. In addition to better understanding sources of diversity, analyses such as these can help generate fresh hypotheses for further investigations.

 

 

Sample classification

A primary research interest at Nom Nom is how microbial composition at baseline can predict responses to interventions such as a diet change or a course of probiotics. Using machine learning classification on baseline data, we have been able to predict with 78% accuracy whether a particular dog will be a high responder to a change from kibble to a fresh diet. A better understanding in this area can lead to more targeted, and ultimately personalized, interventions based on a pet's unique microbial signature.