Microbiome

Deep Learning and feature selection in the bioinformatic modelling of functionally annotated microbial communities in aquaculture

---

Abstract

Historically, Deep Learning was proposed for very large datasets that contain sufficient amounts of information for training without the danger of overfitting, but model training and even evaluation are often very demanding of computational resources. At first glance, processing smaller datasets is technically easier. However, to achieve satisfactory performance of prediction or classification, such less informative data requires extensive tuning of model's architecture, including its hyperparameters, that is also a computational hurdle. Since in our study, a small number of observations (125 fish, 75 water, and 50 sediment microbiome samples) impeded accurate statistical inferences with conventional approaches, we explored the behaviour of a Deep Learning based classifier. This study aimed to investigate the effects of supplementation of common carp feed and water with effective microorganisms during the growth period. 25 ponds were divided into experimental groups. In each pond, the intestinal microbiome of five individuals, the sediment microbiome, and the water microbiome were determined using 16S rRNA gene sequencing. The abundance of Amplicon Sequence Variants was processed to predict functional abundances expressed by KEGG metabolic pathways. The final modelling step aimed to assess whether the abundances of KEGG pathways were sufficiently altered by supplementation to allow for re-creating the original experimental setup of ponds. The classification was performed by an architecture composed of four dense layers with 128, 64, 64 and 32 nodes respectively with a dropout layer (25 %) after the second dense layer. This architecture was concluded by three parallel classification layers for grouping based on: feed supplementation (2 classes), pond water supplementation (3 classes), and supplementation time points (2 classes for sediment, 3 classes for water). The importance of features for each classification was estimated by calculating SHapley Additive exPlanations. Since no formal hypothesis testing is involved in Shapley value estimation, 2D-K-means clustering was applied to distinguish between important and non-important features. All computations were performed in Python using the Keras library with the TensorFlow backend. Feature importance was calculated using the SHAP package. Classification based on intestinal microbiome-derived KEGG pathways reached test-accuracy of 0.84 for feed supplementation classes and a low test-accuracy (ACC) of 0.44 for water supplementation classes - indicating that intestinal microbiome metabolism was altered by probiotic feed supplementation, but not by water supplementation. KEGG diversity of water depended on the time point of supplementation (ACC=0.87), but did not allow for the identification of feed supplementation classes (ACC=0.47) nor water supplementation classes (ACC=0.33). Sediment microbiome functional diversity was not influenced by supplementation, as the classification resulted in ACC of twice 0.50 and 0.30 for feed, time point, and water supplementation classes, respectively. In conclusion, fish intestinal microbiome is mainly influenced by feed supplementation, whereas water supplementation does not influence pond microbial communities.

Keywords: 2026

How to Cite:

Szyda, J., Sztuka, M., Jakimowicz, M., Sidorczuk, K., Słomian, D. & Napora-Rutkowski, Ł., (2026) “Deep Learning and feature selection in the bioinformatic modelling of functionally annotated microbial communities in aquaculture”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2285729. doi: https://doi.org/10.31274/wcgalp.23784

Rights: 1