Optimization of probiotic therapeutics using machine learning in an artificial human gastrointestinal tract

Susan Westfall,Francesca Carracci,Danyue Zhao,Li Shen,James Simon,Molly Estill,Qing-Li Wu,Giulio Maria Pasinetti

doi:10.1038/s41598-020-79947-y

Abstract

The gut microbiota’s metabolome is composed of bioactive metabolites that confer disease resilience. Probiotics’ therapeutic potential hinges on their metabolome altering ability; however, characterizing probiotics’ metabolic activity remains a formidable task. In order to solve this problem, an artificial model of the human gastrointestinal tract is introduced coined the ABIOME (A Bioreactor Imitation of the Microbiota Environment) and used to predict probiotic formulations’ metabolic activity and hence therapeutic potential with machine learning tools. The ABIOME is a modular yet dynamic system with real-time monitoring of gastrointestinal conditions that support complex cultures representative of the human microbiota and its metabolome. The fecal-inoculated ABIOME was supplemented with a polyphenol-rich prebiotic and combinations of novel probiotics that altered the output of bioactive metabolites previously shown to invoke anti-inflammatory effects. To dissect the synergistic interactions between exogenous probiotics and the autochthonous microbiota a multivariate adaptive regression splines (MARS) model was implemented towards the development of optimized probiotic combinations with therapeutic benefits. Using this algorithm, several probiotic combinations were identified that stimulated synergistic production of bioavailable metabolites, each with a different therapeutic capacity. Based on these results, the ABIOME in combination with the MARS algorithm could be used to create probiotic formulations with specific therapeutic applications based on their signature metabolic activity.

Highlights

The gut microbiota’s metabolome is composed of bioactive metabolites that confer disease resilience
The bacteria was dosed into the bioreactor through the normal feeding cycle every 8 h at a concentration of 1 × 1010 colony forming unit (CFU)/meal resulting in accumulation of bacterial CFUs at − 16, − 8 and 0 h before the ABIOME was allowed to run without inoculation
One major hurdle for the development of therapeutic probiotics is the lack of humanized GI models that can be manipulated in controlled environments to study the influence of probiotics on gut microbiota communities and their metabolome

Summary

Introduction

The gut microbiota’s metabolome is composed of bioactive metabolites that confer disease resilience. Miniaturised in vitro models such as gut-on-a-chip[12] have been designed that can even recreate the anoxic–oxic interface defining the mucosal bacteria’s microenvironment[13] The limitation of these smaller systems is the difficulty in establishing a stable gut microbiota with the same complexity as larger systems and since the volume is so low, studying the gut microbial populations or metabolites using standard microbiological techniques is nearly impossible. To dissect the complexity of the gut microbiota, machine learning and artificial intelligence have become important computational tools to discover trends and synergies in large data sets that are otherwise eclipsed by conventional analytical techniques These algorithms are highly adaptable, trainable and designed to account for limited or missing information in line with the practical confines of preclinical and clinical r esearch[14]. Some groups have begun testing computational tools to assess associations between microbiome metagenomic datasets and disease p henotypes[15,16]; machine learning tools have not been used to predict the gut microbiota’s metabolome towards therapeutic optimization of probiotic formulations

Methods

Results

Conclusion