A rapid, affordable, and reliable method for profiling microbiome biomarkers from fecal images

Abstract

Human and veterinary healthcare professionals are interested in utilizing the gut-microbiome as a target to diagnose, treat, and prevent (gastrointestinal) diseases. However, the current microbiome analysis techniques are expensive and time-consuming, and data interpretation requires the expertise of specialists. Therefore, we explored the development and application of artificial intelligence technology for rapid, affordable, and reliable microbiome profiling in rhesus macaques (Macaca mulatta). Tailor-made learning algorithms were created by integrating digital images of fecal samples with corresponding whole-genome sequenced microbial profiles. These algorithms were trained to identify alpha-diversity (Shannon index), key microbial markers, and fecal consistency from the digital images of fecal smears. A binary classification strategy was applied to distinguish between samples with high and low diversity and presence or absence of selected bacterial genera. Our results revealed a successful proof of concept for "high and low" prediction of diversity, fecal consistency, and "present or absent" for selected bacterial genera.

Keywords: Biological sciences; Microbiology; Microbiome.

Graphical abstract

Figure 1

Fecal smear technique to acquire high quality digital images (A) Paper template with printed square and fiducial marks on a flat surface, with the sample ID written on the top section. (B) Feces applied within the borders of the square. (C) Fecal layer smeared as evenly as possible using a spatula. (D) Photographing the smear including the fiducial marks and sample identification.

Figure 2

General overview of the process of the experimental AI/ML method in comparison to the conventional screening of the microbiome using sequencing techniques For the experimental method, feces is smeared on a paper template, subsequently a photograph is taken and uploaded into the AI algorithm. Within seconds, targeted results such as diversity and genera are obtained. This targeted result is easier to use for clinicians to diagnose, intervene, or treat patients with gastrointestinal problems.

Figure 3

Alpha-diversity prediction model (Shannon-index) with binary classification strategy high versus low diversity showing good predictive accuracy The area under the receiver operating characteristic curve (AUROC), data are presented as mean ± SD, for the diversity prediction model.

Figure 4

Good predictive model performance for bacterial genera: Coprococcus, Phascolarctobacterium, and Ruminococcus The area under the receiver operator characteristic curve (AUROC), data are presented as mean ± SD, for the genus presence predictions. The model shows good predictive performances (AUC >0.7) for the genera Coprococcus, Phascolarctobacterium, and Ruminococcus. Less successful predictive performances were observed for Intestinimonas and Oscillibacter (AUC <0.7).

Figure 5

Fecal consistency prediction model based on Waltham score with binary classification strategy high versus low Waltham score shows good predictive accuracy The area under the receiver operator characteristic curve (AUROC), data are presented as mean ± SD for high versus low Waltham score.