Microbiome

The DeepMicroCore project: Machine Learning algorithms for the identification of the core microbiome from integrated data on multiple species

---

Abstract

The DeepMicroCore project aims to leverage machine learning (ML), in particular explainable AI (Artificial Intelligence), to identify the core microbiome across substrates and animal species. On a large integrated dataset, ML/AI models are developed to accurately predict substrates and species of origin of any given microbiome: this is a multi-class classification problem where the target is either the tissue of origin (e.g. gut, milk) or the tissue-species combination (e.g. human gut, bovine milk). Variable importance techniques then reveal the relevant microbial species for classification, which are supposed to constitute the core microbiome. These results are benchmarked against current approaches for the identification of core microbiomes, which are based on frequency and abundance.Our integrated dataset currently include ~1000 samples from the milk, rumen and gut microbiomes of cows, pigs, goats. Data were obtained from ten 16S rRNA-gene sequencing research projects and were integrated using a custom bioinformatic pipeline based on a modified Nextflow nf-core/ampliseq analysis workflow. Initial predictive models have been prototyped on a reduced dataset of cow milk, gut and rumen microbiomes, using Lasso-penalised logistic regression (LLR), Random Forest (RF), Extreme Gradient Boosting (XGB) and convolutional neural networks (1d-CNN). Two different data representations were used: i) filtered and normalised count (ASV/OTU) tables, for the LLR, RF and XGB models; ii) direct sequence data, either independent or joined R1/R2 reads, for 1d-CNN. A validation split approach was used to evaluate the performance of predictive models: 80% of the data were used for training, 20% for testing. Within the training set, a 5-fold cross-validation scheme was used to tune the hyperparameters: the degree of penalization in LLR; number and depth of trees, number of resampled variables, and size of terminal nodes for RF and XGB -plus shrinkage parameter for XGB only; network architecture, learning rate, batch size and number of epochs for 1d-CNN. Results showed high accuracy of predictions from all multi-class classification models, with very few misclassifications (average accuracy = 0.984; average Cohen's kappa = 0.97; average Matthew's correlation coefficient = 0.973). Such high accuracy justifies the use of important variables (microbial taxa) to identify the core microbiome of the given substrate and/or animal species. Next steps include the application of predictive models to the larger integrated datasets and the comparison of results with current approaches to the definition of core microbiomes. The results of this work will lead to a better understanding and definition of core microbiomes, and to practical applications like: i) traceability of samples and food products; ii) identification of contamination (e.g. fecal material in milk); iii) detection of dysbiosis; iv) breeding and selection (microbiability); v) investigating the effects of probiotics.

Keywords: 2026

How to Cite:

Biscarini, F., Cozzi, P., Castiglioni, B., Stella, A., Bobbo, T., Bala, K., Škipina, M. & Seničar, M., (2026) “The DeepMicroCore project: Machine Learning algorithms for the identification of the core microbiome from integrated data on multiple species”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2286440. doi: https://doi.org/10.31274/wcgalp.24006

Rights: 1