Artificial Intelligence in Microbiome Research and Beyond: Connecting Human Health, Animal Husbandry, and Aquaculture

Abstract

Technological advancements in computational power and algorithm design have enabled artificial intelligence to become a transformative force in microbiome research. This paper presents a concise overview of recent applications of this computational paradigm in human and animal health, with a particular emphasis on aquaculture. International projects focused on the intestinal microbiome have allowed human research to consistently dominate in terms of application cases, offering insights into various pathological conditions. In contrast, animal research has leveraged artificial intelligence in microbiome analysis to promote sustainable productivity, addressing environmental and public health concerns linked to livestock husbandry. In aquaculture, on the other hand, artificial intelligence has mainly supported management practices, improving rearing conditions and feeding strategies. When considering microbiome manipulation, however, fish farms have often relied on traditional methods, without harnessing the immense potential of artificial intelligence, whose recent applications include biomonitoring and modeling interactions between microbial communities and environmental factors in farming systems. Given the paradigm shift currently underway in both human health and animal husbandry, we advocate for a transition in the aquaculture industry toward smart farming, whose interconnected infrastructure will allow to fully leverage artificial intelligence to seamlessly integrate both biological measurements and rearing parameters.

1. Introduction

Over the last decades, the increased affordability of high-throughput sequencing and the significant growth in computational power have helped microbiome research evolve from a fledgling field to a thriving discipline of interdisciplinary interest with applications that now span human, animal, and environmental health [1,2,3]. At the same time, however, the increasing scope of this research field has necessitated the redefinition of the notion of microbiome, whose original meaning has been progressively amended to better reflect the multitude of research perspectives and thus comprehensively capture the multifaceted nature of this ecosystem, although a consensus definition remains a matter of intense debate [4]. At the moment, the research community distinguishes between the microbiota, which is referred to as a consortium of microorganisms residing in a particular environment, and the microbiome, which builds on the definition of microbiota to include both the biomolecules produced by the microbial community and the environmental conditions of the host. For a complete understanding of the microbiome, however, other factors must also be considered, such as microbial interactions (i.e., within-species and between-species communication patterns within different symbiotic relationships) and ecosystem dynamics (i.e., temporal and spatial structuring) [4,5]. While the current definition of microbiome will inevitably evolve to make room for future insights, recognizing the difference between these two core concepts has given microbiome researchers a common ground to better delineate the experimental framework of research studies conducted in various application areas.

With the recognition of the universal occurrence of microorganisms, scientific research has recently undergone a paradigm shift in the interpretation of the relationship between host and microbiome, from autonomous entities to interdependent metaorganisms acting as a functional unit of biological organization, with significant implications from a coevolutionary perspective [6,7]. Within the boundaries of this association, the microbiome dynamically oscillates between two states, namely eubiosis, which describes a balanced state of the microbial community, and dysbiosis, which denotes a disturbed state in microbial composition [8]. In the context of dysbiosis, in particular, microbiome researchers have also proposed the introduction of the notion of pathobiome to specifically refer to the integration of pathogens into the host environment [9]. However, the indiscriminate usage of these terms has made the quantitative distinction between eubiosis and dysbiosis much more difficult [10]. In an effort to understand the complex interactions among microbiome components (i.e., microbe-microbe, microbe-host, and microbe-environment), research groups have consequently generated vast amounts of data. However, significant problems remain due to the lack of consistent standardization (e.g., clinical protocols, experimental methods, and computational techniques), which has limited both the integration of information from different studies and the efficient transfer of research results from basic research to application areas [4,11].

Over the last decade, the microbiome information generated by multi-omics technologies at different biological levels has been accumulated in increasingly complex datasets, whose dimensionality and heterogeneity have required the introduction of innovative approaches to extract meaningful insights. In this context, microbiome researchers have increasingly turned to machine learning (ML) as a data mining solution [12]. Commonly known as a computer science discipline in which machines are programmed to recognize data patterns and extract knowledge according to learning rules defined in terms of mathematical formulas and statistical assumptions, machine learning has made it possible to detect previously unaccounted associations among features and thus generate a number of desirable outcomes based on the chosen learning approach (Figure 1). In this regard, the type of data available (i.e., labeled or unlabeled) and the research task to be performed (i.e., classification, regression, or clustering) represent the main decision criteria used to select the appropriate learning framework. On the one hand, supervised learning maps the relationships between input variables and designated output variables from labeled training instances to make accurate predictions on unseen data (i.e., discrete values for classification tasks and continuous values for regression tasks). On the other hand, unsupervised learning extracts patterns from unlabeled data without the need to provide ground truth information in the form of predetermined labels in order to cluster data based on feature similarity or reduce data dimensionality [13]. Although these approaches represent the main categories in ML-based data analysis, specialized experimental scenarios could benefit from more advanced frameworks, such as semi-supervised learning, which combines small amounts of labeled data with large amounts of unlabeled data, and reinforcement learning, which teaches agents to make sequential decisions by maximizing cumulative rewards through trial and error in an environment [14].

While machine learning has remained the preferred solution in terms of model implementation and computational time, microbiome researchers have witnessed the emergence of deep learning (DL) as a promising alternative for knowledge discovery in large-scale, high-dimensional data (Figure 1). These models are based on artificial neural networks in which information is transmitted through multiple layers of weighted interconnected units implementing mathematical functions able to transform data into a representation suitable for pattern learning and thus generate an appropriate output [13]. With the increasing application of machine learning and deep learning, explainable artificial intelligence (XAI) has also been introduced in microbiome research to improve the interpretability of model predictions from a biological perspective [15]. Despite the plethora of available approaches, when selecting the appropriate solution, several factors should be taken into careful consideration (e.g., data provenance assumptions, data collection protocols, data size homogeneity, algorithm assumptions, feature selection, model implementation, and model interpretability) to avoid biased results and thus generate sufficiently complex models across different levels of biological organization [16,17,18].

As both machine learning and deep learning offer promising insights into the intricate mechanisms underlying the functioning of biological systems as metaorganisms, microbiome research is expected to benefit from these approaches in many ways, such as linking specific taxa to host phenotypes and monitoring host responses to changes in microbiota composition [19,20]. In this narrative review, we discuss the current applications of these subsets of artificial intelligence (AI) in microbiome research, with a particular emphasis on the gastrointestinal tract, which has become a topic of great interest in academic and non-academic fields. In particular, we outline the current application of artificial intelligence in human and animal health, focusing our attention on aquaculture in the latter case. In order to provide the most up-to-date information possible, a literature search has been conducted by screening different online repositories (Google Scholar, PubMed, and Scopus) up to August 2025.

2. Harnessing Artificial Intelligence for Human Health Insights

With the increasing use of artificial intelligence as an innovative tool for biological knowledge extraction, microbiome research in human health has progressively embraced this computational paradigm within experimental settings. This integration aims to elucidate the role of microbial communities residing in different anatomical regions, most notably the gastrointestinal tract, and their influence on human physiological functions, especially those connected to pathological conditions. Indeed, as modern living conditions continue to reshape human microbial ecosystems, researchers have employed AI solutions to deepen our understanding of microbe-host-environment interactions. Within this integrated health framework, such growing awareness has determined an increase in efforts to harness the microbiome for different purposes, such as personalized medicine, preventive healthcare, and the development of innovative therapeutic strategies.

2.1. Connection Between Intestinal Microbiome and Diseases

Multi-omics investigations have enriched our understanding of the human intestinal microbiome. According to recent estimates, the microbial community that colonizes the gastrointestinal tract consists of more than 10¹³ cells with a collective genome of more than 10⁶ genes [21,22], although the number of unique genotypes could substantially exceed these estimates [23]. Additionally, microbiota dysbiosis is recognized as an important factor in the pathogenesis of various diseases (e.g., metabolic disorders, neurological disorders, cardiovascular diseases, liver diseases, and autoimmune diseases), with microbial metabolites (e.g., phenols, indoles, secondary bile acids, short-chain fatty acids, and branched-chain fatty acids) playing a prominent mediatory role in the modulation of signaling pathways and the regulation of metabolic processes [24,25,26]. When investigating the associations between intestinal microbiota and health conditions, artificial intelligence has introduced sophisticated computational methodologies that have been applied to clinically significant conditions in the modern healthcare landscape (

Table 1. ML and DL applications in microbiome research on the human gastrointestinal tract. The mentioned papers have been conveniently grouped into two categories to provide a high-level categorization. In the algorithm column, one of the following implementation possibilities has been reported: (i) the most performant model or model combination (denoted with +); (ii) the list of trained models in the case of similar performances. For information about the algorithms applied during feature selection, refer to the original paper. (CNN, convolutional neural network; k-NN, k-nearest neighbor; LSTM, long short-term memory; MDS, multidimensional scaling; NMDS, non-metric multidimensional scaling; NMF, non-negative matrix factorization; PAM, partitioning around medoids; PCA, principal component analysis; PCoA, principal coordinate analysis; RF, random forest; SVM, support vector machine; XGB, extreme gradient boosting).

Category	Health Condition	Algorithm	Data Type	Reference
Patient stratification	Breast cancer	XGB	16S rRNA gene	[29]
	Parkinson’s disease	LSTM + SVM	metagenomics	[30]
	Heart failure and depression	RF	metagenomics, metabolomics	[31]
	Obesity	SVM	metagenomics	[32]
	Liver disease	CNN	16S rRNA gene	[33]
Microbiota composition analysis	Microbial dysbiosis	k-NN	metagenomics	[34]
	Dietary intake and allergic rhinitis	NMDS, NMF, PCA, PCoA	16S rRNA gene	[35]
	Geography- and season-based variation	MDS + PAM	16S rRNA gene	[36]
	Pregnancy	PAM	16S rRNA gene	[37]
	Pregnancy	PAM	16S rRNA gene	[38]

), enabling the development of personalized therapeutic plans [27]. As a matter of fact, microbiome engineering (e.g., fecal microbiota transplant, antimicrobial peptide administration, dietary intervention, prebiotic and/or probiotic supplementation) is currently regarded as a promising intervention strategy for disease treatment [3], based on the dominance of environmental factors over host genetics in shaping microbiota composition [23,28].

When extracting information from labeled data, supervised ML has provided microbiome researchers with various algorithms. For instance, Huang et al. [31] used a random forest (RF) classifier to recognize patients with comorbid heart failure and depression from metagenomic and metabolomic data measured in fecal samples collected from 95 individuals. In this cross-sectional study, the model trained on both data types attained an area under the receiver-operating characteristic curve (AUC) of 0.83, with Cloacibacillus and α-tocopherol as taxonomic and metabolic signatures, respectively. Alternatively, Asher and Bashan [34] applied a multi-dimensional k-nearest neighbor (k-NN) regressor to predict the abundance profiles of microbial communities from different mucosal surfaces (including the gastrointestinal tract) after microbiome perturbation, using species assemblage (i.e., the sample-specific configuration of resident species) as the independent variable. Under the assumption that samples with similar species assemblages have similar abundance profiles, the model performed efficiently on metagenomic data from multiple reference datasets. This approach was compared to traditional population dynamics models and showed potential to design specific perturbations to guide the microbiota toward a desired composition. On the other hand, Liu et al. [32] compared the performance of classification and regression algorithms on body mass index (BMI) prediction, as BMI groups and BMI values, using metagenomic data derived from the fecal samples of 2262 Chinese individuals. In this obesity-focused study, the support vector machine (SVM) algorithm achieved the highest performance in both predictions, with a mean absolute error (MAE) of 1.63 for regression and an accuracy of 0.72 for classification. As a data-driven approach, supervised ML has overcome several limitations of traditional microbiome research, such as data integration, pattern recognition, biomarker prioritization, and predictive modeling [39].

When research interest focuses on uncovering hidden structures and natural groupings within data, unsupervised ML offers significant advantages. For example, Yan et al. [38] used the partitioning around medoids (PAM) algorithm to investigate clustering patterns in 16S rRNA gene sequencing data from the fecal samples of 45 Chinese women at different pregnancy stages (no pregnancy, late-term pregnancy, and full-term pregnancy). This approach identified three microbial clusters: cluster 1 (Firmicutes- and Bacteroidota-rich) found in each group, cluster 2 (Firmicutes- and Actinobacteriota-rich) found in late-term and full-term pregnancy, and cluster 3 (Proteobacteria- and Firmicutes-rich) found in late-term pregnancy. Komaki et al. [35], instead, compared different dimensionality reduction techniques, namely principal component analysis (PCA), principal coordinate analysis (PCoA), non-metric multidimensional scaling (NMDS), and non-negative matrix factorization (NMF), to visualize the association of microbiota composition with dietary intake and allergic rhinitis symptoms, using 16S rRNA gene sequencing data from 292 patients. The results showed significant association variability based on the dimensionality reduction solution and the distance matrix selected for community comparison. Given the peculiar properties of microbiome data (e.g., dimensionality, redundancy, sparsity, and compositionality), clustering and dimensionality reduction algorithms represent valuable analytical tools, thanks to the robustness of composite variables to technical and biological noise [40,41].

While ML algorithms have provided valuable insights in microbiome research, DL models have allowed to efficiently address the increasing complexity of biological data, although both approaches are not mutually exclusive. As an example, Yu et al. [30] combined machine learning and deep learning to determine the influence of microbiota alterations on the pathogenesis of Parkinson’s disease, using the abundance profiles measured in the fecal samples of 78 Chinese individuals, based on the bidirectional interactions between the intestinal microbiota and the nervous system [42,43]. In this cross-sectional study, the authors implemented a three-layer architecture: the input layer integrates RF and PCA for feature selection; the middle layer combines long short-term memory (LSTM) and SVM for network training; and the output layer applies soft voting for class prediction. The proposed system performed competitively when compared to other available approaches, with an accuracy of 0.85 and an AUC of 0.92. However, numerous neural networks (NNs) have been developed to leverage multiple data types [27,44], such as feedforward neural networks (FNNs), recurrent neural networks (RNNs), convolutional neural networks (CNNs), generative adversarial networks (GANs), graph neural networks (GNNs), attention networks (ANs), and autoencoders (AEs). Thanks to increasingly approachable interfaces, DL frameworks have equipped researchers with sophisticated tools to mine multi-omics data and thus capture patterns that might not be recognized by traditional ML [20,27].

In an effort to address the black-box functioning of machine learning and deep learning, interpretable frameworks have helped researchers visualize the contribution of individual features to model predictions. For instance, Kibria et al. [29] compared the performance of ML and DL algorithms on 16S rRNA gene sequencing profiles from the fecal samples of 962 Ghanaian individuals to elucidate the connection between microbiota composition and breast cancer risk. Using extreme gradient boosting (XGB) as a target model, with an accuracy of 0.76 and 0.72 in balanced and imbalanced dataset scenarios, respectively, the authors computed the Shapley additive explanations (SHAP) values of the bacterial taxa chosen during feature selection. In this study, SHAP analysis identified Clostridium saccharogumia (recently renamed Thomasclavelia saccharogumia [45]), Eubacterium dolichum, and Megamonas as the features with the highest impact on breast cancer occurrence. Although SHAP has been the predominant framework in microbiome research [46,47], other techniques have been proposed as viable alternatives for feature importance estimation, such as permutation feature importance, local interpretable model-agnostic explanations (LIME), and individual conditional expectation (ICE) [48]. As the need for personalized therapeutic plans continues to increase, XAI frameworks are paving the way for determining the role of intestinal microbiota in disease pathogenesis at single-taxa resolution.

2.2. Interrelationship Between Humans and Fish in Microbiome Research

While recent breakthroughs in microbiome research have contributed to a more comprehensive interpretation of the pathogenesis of human diseases, these advancements have also expanded our understanding of the intestinal microbiome of evolutionarily distant species, including fish [49]. For instance, Shima et al. [50] reflected on the bacterial composition of the intestinal microbiota in humans and fish. They noted that the human microbiota mainly consists of Firmicutes and Bacteroidetes, with an average relative abundance of 22.2% and 73.1%, respectively [51]. In contrast, the fish microbiota primarily consists of Firmicutes and Proteobacteria, with an average relative abundance of 13.5% and 51.7%, respectively [52]. These differences highlight the importance of host habitat [53] and, to a lesser extent, host diet [54,55,56] in shaping fish microbiota composition. From a comparative standpoint, however, our understanding of microbiome composition across both clades has been significantly constrained by an imbalance in research efforts, which have disproportionately focused more on mammals while largely overlooking other organisms such as fish, which represent the most phylogenetically diverse vertebrate group. This research bias has limited deeper insights into host-microbiome relationships across evolutionary lineages. However, given the demonstrated microbial convergence between human and piscine intestinal microbiomes, expanding microbiome studies to also include aquatic species holds considerable promise for understanding the evolution of intestinal microbiomes throughout the vertebrate tree of life [57].

As comparative microbiome studies progressively fill the knowledge gaps between these vertebrate clades, fish microbiome research has found innovative applications and introduced transformative perspectives within the field of human health. For instance, the intestinal microbiota of marine fish is being examined for the sourcing of bioactive compounds, especially bacteriocins (i.e., low-weight ribosome-synthetized antimicrobial peptides), for future applications as therapeutic tools in antibiotic-resistant disease treatment in human health as well as aquaculture [58]. Moreover, owing to its high degree of genetic homology with the human genome, zebrafish (Danio rerio) has become a valuable model organism for multiple human diseases [59,60,61], including neurological disorders [62], immune disorders [63], cardiovascular disorders [64], and cancer [65]. On a different level, microbiome research in fish species is also reshaping our current understanding of the human body. For example, Mani et al. [66] identified Plesiomonas and Agrobacterium communities in the brain of healthy salmonids, with a bacterial load 1000-fold lower than in the gastrointestinal tract. While the connection between the nervous system and the intestinal microbiota is widely recognized, these recent discoveries have motivated researchers to determine the existence of resident microbes in the human brain [67]. Such increasing convergence of human and fish microbiome research is bound to have far-reaching implications, ultimately enhancing both human health and aquaculture sustainability.

3. Enhancing Animal Health Through Artificial Intelligence

As outlined in the previous section, microbiome research has revealed a bidirectional connection between human and animal studies. Indeed, the knowledge derived from human research has been instrumental in advancing our understanding of animal microbiomes, while, in turn, discoveries in animal models have reshaped modern perspectives on human physiology. Given the considerable funding for human-focused research in recent decades, studies on the human microbiome have expectedly deepened our understanding of host-microbe interactions, thus laying the foundations for more precise microbiome exploration across species with varying evolutionary distances. As a result, analytical techniques developed for human microbiome research have been adjusted to investigate microbial ecosystems in animals. Despite the different objectives of these research domains, this cross-species perspective has fostered a more holistic understanding of microbiome-driven health, with human research focusing on elucidating the role of microbial dysbiosis in disease pathogenesis and animal research emphasizing environmentally sustainable production, particularly in response to the challenges posed by climate change.

The necessity for sustainable animal production has contributed to the investigation of the intestinal microbiome of animal species with different domestication status. In the case of livestock animals, whose evolution has been profoundly shaped by human intervention, microbiome manipulation has allowed to increase animal productivity, thus supplementing breeding and selection practices with advanced tools for more precise intervention [49]. In these microbiome studies, research areas can be distilled into the following categories: qualitative and quantitative production of food commodities, disease mitigation in livestock species, effects on human health from the consumption of animal products, and effects on animal microbiomes from dietary changes in feeding regimes [68]. Nevertheless, microbiome research currently encompasses numerous animal species, including companion animals (e.g., cats and dogs) and wildlife animals (e.g., amphibians, marine mammals, and non-human primates), and their influence on human health [68]. Given the role of intestinal microbiomes in host adaptation to environmental disturbances, future research will have a significant impact on both conservation efforts [7,69] and human health management [70]. In the latter case, in particular, the potential emergence of infectious diseases from animal vectors necessitates the development of holistic frameworks incorporating the ecological relationships between human, animal, and environmental components [2]. Unlike human research, however, animal microbiome studies have made limited use of AI models (

Table 2. ML and DL applications in microbiome research on the gastrointestinal tract of livestock groups. In the algorithm column, one of the following implementation possibilities has been reported: (i) the most performant model or model combination (denoted with +); (ii) the specific algorithm component within a pipeline (denoted with ⁕); (iii) the list of trained models in the case of similar performances. For information about the algorithms applied during feature selection, refer to the original paper. (AB, adaptive boosting; BERT, bidirectional encoder representation from transformers; CB, categorical boosting; CNN, convolutional neural network; DT, decision tree; ET, extra tree; FNN, feedforward neural network; GB, gradient boosting; GNB, Gaussian Naïve Bayes; LR, logistic regression; PAM, partitioning around medoids; PLS-DA, partial least squares discriminant analysis; RF, random forest; SVM, support vector machine; XGB, extreme gradient boosting).

Livestock Group	Algorithm	Data Type	Purpose	Reference
Ruminants	FNN + CNN	metagenomics	identification of fungal and protozoan sequences	[71]
	RF	16S rRNA gene	stratification based on MUN values	[72]
	PAM	16S rRNA gene	investigation of rumen bacterial clusters	[73]
Swine	DT, RF, XGB, AB, CB, SVM, GNB + PLS-DA, LR + PLS-DA	16S rRNA gene	stratification based on breeding status	[74]
	BERT ⁕	metagenomics	identification of viral sequences	[75]
	PAM	16S rRNA gene	association between microbial clusters and NUE values	[76]
	GB	ITS and 18S rRNA gene	association between eukaryotic communities and body weight	[77]
Poultry	transformer	16S rRNA gene	pathogen prediction in microbial clusters	[78]
	ET	metagenomics	prediction of antibiotic resistance/susceptibility	[79]

), except for unsupervised learning algorithms, such as principal component analysis and principal coordinate analysis, which are routinely used to visualize composition similarities among microbial communities.

Intestinal Microorganisms for Improved Animal Husbandry

In light of the significant contribution of ruminants to human development throughout history, the rumen microbiome has been extensively investigated to elucidate its connections to food production, human nutrition, and environmental footprint [80,81,82,83,84]. For instance, Yu et al. [72] applied supervised classifiers on 16S rRNA gene sequencing data from the fecal samples of 161 Holstein dairy cows to categorize specimens based on milk urea nitrogen (MUN) concentrations. Utilizing RF-based feature selection together with SHAP analysis, nine bacterial genera were selected as core features, which allowed the system to attain an accuracy of 72.7% and an AUC of 0.76, with RF as the best model. Alternatively, Wang et al. [73] investigated rumen bacterial clusters in 99 Guanzhong dairy goats, using abundance profiles measured in rumen fluids with 16S rRNA gene sequencing, based on the notion of human enterotype, as originally proposed by Arumugam et al. [85], and similar findings from the intestinal microbiome of dairy cows [86]. In this study, PAM-based enterotype analysis allowed us to identify two clusters, the P-cluster (Prevotella-rich) and the R-cluster (Ruminococcus-rich), which showed varying degrees of association with host metabolism. On the other hand, Yan et al. [71] implemented a two-layer architecture for the identification of metagenomic eukaryotic sequences, specifically fungi and protozoans, which account for 20% and 50% of the rumen microbiome, respectively [84], while accounting for less than 1% in the human intestine [87]. Each layer relies on an ensemble model combining FNNs and CNNs to discriminate between prokaryotic and eukaryotic sequences in the first stage, and between fungal and protozoan ones in the second stage. The proposed system outperformed existing tools when tested on metagenomic libraries representing intestinal microbiomes from different ruminant species with varying feeding regimes. Considering the strong connection between animal and human health, understanding the rumen microbiome will be crucial to improve current farm management practices.

Similarly to other livestock species, the porcine intestinal microbiome has been studied in relation to health, growth, and yield [88,89,90,91,92,93]. As an example, Azouggagh et al. [74] tested several supervised classifiers on 16S rRNA gene sequencing data from the fecal samples of 237 Iberian pigs to determine the microbial taxa necessary to categorize specimens based on breeding status. While model performances were sufficiently similar, the results showed better discrimination between purebred individuals, in contrast with previous findings indicating tighter grouping among crossbred samples [94]. Conversely, Ramayo-Caldas et al. [77] investigated the association between eukaryotic communities and body weight after weaning on ITS and 18S rRNA gene sequencing data from the fecal samples of 514 Duroc pigs. Using gradient boosting (GB) coupled with SHAP analysis, the authors identified: as regards fungi [95], a positive contribution from Kazachstania; as regards protists [96], both positive (Blastocystis and Entamoeba) and negative (Trichomitus) associations. Alternatively, Sarpong et al. [76] researched the relationship between microbial clusters and nitrogen utilization efficiency (NUE) on 16S rRNA gene sequencing data from the fecal samples of 892 crossbred pigs. In this work, PAM-based enterotype analysis detected Lactobacillus-rich and Clostridium sensu stricto-rich clusters. However, the authors also reported enterotype change occurrences between sampling periods, as confirmed by previous studies on growth stages, including post-weaning [97], between weaning and finishing [98], and post-finishing [99]. When also considering the virome, instead, Mi et al. [75] developed a pipeline for virus sequence identification in metagenomic data. In particular, the authors integrated PhaTYP, a tool developed by Shang et al. [100] based on bidirectional encoder representations from transformers (BERT), to determine the virulence of viral operational taxonomic units (vOTUs). Using 4650 samples of different geographical origins, the pipeline found Caudoviricetes as the most abundant taxa, with Firmicutes and Bacteroidetes as the most common bacterial hosts. Understanding the porcine intestinal microbiome will therefore benefit both husbandry and animal model research. In the latter case, in particular, thanks to physiological similarities, porcine systems have received much attention as potential models for dietary modulation in the human intestinal microbiota [101].

Unlike other livestock species, the intestinal microbiome has been far less studied in poultry due to high variability between flocks [102,103,104,105,106,107,108]. As a notable study, Baker et al. [79] introduced a three-phase surveillance method, using metagenomic data from the fecal samples of 170 broiler chickens, with Escherichia coli as indicator species for antimicrobial resistance (AMR). In the first stage, an antibiotic-specific feature set consisting of antibiotic-resistant gene (ARG) counts and microbial abundance profiles is derived based on antimicrobial susceptibility testing on 26 antibiotics. In the second stage, resistance/susceptibility prediction is performed with ML classifiers for each antibiotic, together with SHAP analysis to visualize feature influence on prediction results. In the third stage, feature dependency on humidity and temperature is assessed with ML regressors to identify correlations between metagenomic features and temperature/humidity conditions. In addition to extra tree (ET) being the best classifier, with an AUC above 0.90 in ten antibiotics, the system found Alcaligenes faecalis, an emerging pathogen in human health, to be correlated with both temperature and humidity. Alternatively, Ram Das et al. [78] proposed a three-stage transformer-based framework for pathogen prediction in microbial clusters. During data pre-processing, categorical and continuous data are encoded to fixed-dimension vectors. Then, a transformer architecture trains on the Farm Management Practices Dataset [109] and the Microbiome Dataset [110] to predict pathogen prevalence (Salmonella, Listeria, and Campylobacter) under different experimental scenarios. Feature explainability was finally assessed through attention matrices by using PageRank to evaluate feature importance, whose estimates were also compared with those of Deep Learning Important Features (DeepLIFT), and hierarchical agglomerative clustering (HAC) to identify microbial clusters with similar ecological properties. When combining management and microbiome data, the model outperformed traditional methods. Considering the importance of sustainable poultry production, elucidating the intricacies of the avian intestinal microbiome is hence posed to make significant contributions to both predictive microbiology and food safety [111].

4. Trends in Artificial Intelligence Application to Aquaculture

Within the broader domain of animal health, microbiome research in aquaculture has emerged as a transformative frontier, offering valuable insights into the microbe-host-environment interactions that influence fish health, growth performance, and environmental sustainability. In this context, aquatic systems, which differ from their terrestrial counterparts in terms of ecological dynamics and microbial composition, present unique challenges and opportunities for scientific investigation. Indeed, with aquaculture operations becoming more intensive to meet current market demands, a comprehensive understanding of the microbial communities inhabiting the aquatic environment and the gastrointestinal tracts of farmed species has become essential for optimizing seafood production and mitigating disease outbreak. To this end, artificial intelligence, which has been traditionally applied for non-microbiome-related purposes, has gained prominence as a powerful tool for elucidating the largely understudied dynamics of fish microbiota with unprecedented precision. In an effort to develop more targeted microbiome-informed intervention strategies, the integration of AI models into aquaculture studies not only enhances animal health management but also contributes to the growing necessity for more resilient and sustainable food systems.

4.1. Microbial Diversity in Fish Microbiota

The increasing reliance on fish consumption in human nutrition has led to the economic growth of the aquaculture industry, which has given rise to a renewed interest in fish intestinal microbiome [112,113,114,115,116]. Due to the high variability in fish biology, however, the gastrointestinal tract is extremely diversified, although it can be usually divided into different topographical regions (i.e., headgut, foregut, midgut, and hindgut) [117], which are populated by different microbial communities [116]. The microbial distribution along the alimentary canal is primarily determined by the microorganisms found in the immediate surroundings of the fish larva (i.e., egg surface, initial feed, and water) [116,117], reaching its final composition based on species-specific pressures [54,118], with a core microbiota established within the first two months, as observed in both zebrafish [119] and European sea bass (Dicentrarchus labrax) [120]. Indeed, current estimates indicate a microbial presence of 10⁹ colony forming units (CFUs) per gram [121], contrarily to earlier studies suggesting much lower levels [122]. Moreover, intestinal microbial diversity has been found to be influenced by both biological factors (e.g., life stage, trophic level, breeding status, diet, sex, and phylogeny) and environmental factors (e.g., temperature, season, salinity, and habitat) [116,117,123,124,125], including, in the case of aquaculture, antibiotics [126]. While the exact mechanism behind fish-microbiota-environment relationships is still debated, studies have proposed trophic level, habitat, and phylogeny as the main contributors [52,53,127]. As the reciprocal regulation between the intestinal microbiota and the physiological functions of the fish host (e.g., energy balance, feeding behavior, nutrient metabolism, and immune response) continues to be studied [55,128], advances in molecular methodologies will allow researchers to determine the existence of a causal relationship [112].

Despite its extreme diversification, the piscine intestinal microbiota comprises both prokaryotic microorganisms (bacteria and archaea) and eukaryotic microorganisms (protists and fungi) [113]. Within this landscape, Proteobacteria and Firmicutes represent the most abundant phyla, which, together with Bacteroidetes, account for 90% of the total microbial biomass [129], whereas the mammalian intestinal microbiota is characterized by a prevalence of Bacteroidetes and Firmicutes [130]. While these estimates provide useful approximations, the intestinal microbiota composition changes when taking additional parameters into account. For instance, salinity has been suggested to have a significant influence on microbial prevalence, as shown by Aeromonadales in freshwater species and Vibrionales in saltwater species [53]. When considering the trophic level, instead, the microbial balance shifts toward Firmicutes in herbivorous fish and Proteobacteria in non-herbivorous fish [113]. Moreover, microbial diversity is unevenly distributed across the gastrointestinal tract in several respects (e.g., density, composition, and function) [131], which are further complicated when differentiating between allochthonous communities (transient microbiota) and autochthonous communities (resident microbiota) [132,133]. As accommodating such variability in fish microbiome analysis becomes an increasing necessity, artificial intelligence has introduced computationally efficient tools to understand the mechanisms governing the intestinal microbial ecosystem, especially with critical issues emerging in the aquaculture industry (e.g., high intensity, metabolic diseases, water quality, and antibiotic abuse) and microbiome manipulation becoming common practice to improve both fish health and farming productivity [55,112,117,124].

4.2. Benefits of Artificial Intelligence in Fish Welfare and Farming

While exact cross-sector statistics are not yet available, unlike human health and animal husbandry, artificial intelligence has found limited use in the study of the intestinal microbiome of piscine species. Indeed, when conducting literature research on the selected online repositories using field-specific keywords (e.g., aquaculture, microbiome, artificial intelligence, machine learning, deep learning), the number of relevant results was notably low, with currently available applications limited to natural environments and, to an even lesser extent, aquaculture facilities (

Table 3. ML and DL applications in microbiome research on the fish gastrointestinal tract. In the algorithm column, the most performant model has been reported. For information about the algorithms applied during feature selection, refer to the original paper. (BN, Bayesian network; PAM, partitioning around medoids; RF, random forest).

Algorithm	Data Type	Purpose	Reference
PAM	16S rRNA gene	microbial clustering according to trophic levels	[135]
BN	16S rRNA gene	network-based modeling of microbiome relationships with rearing conditions	[136]
RF	16S rRNA gene	contamination monitoring in aquatic environments	[134]

). For instance, Turner Jr et al. [134] leveraged ML and DL algorithms on 16S rRNA gene sequencing data from the fecal samples of 79 teleost fish, namely 21 stoplight parrotfish (Sparisoma viride) as saltwater species and 58 walleye (Sander vitreus) as freshwater species, to determine the suitability of intestinal microbiota for the biomonitoring of compromised aquatic environments (CAEs). The results showed remarkable accuracy regardless of species ecology and anthropogenic stressor, with 96% for NNs on saltwater samples, 92% for GB on freshwater samples, and 92% for RF on the combined dataset, due to similar bacterial shifts observed in both groups, with Firmicutes being the dominant phylum in CAEs and Proteobacteria being the dominant phylum in non-CAEs. In an alternative work, Zhang et al. [135] studied the influence of different factors (habitat preference, host taxonomy, and trophic level) on the intestinal microbiome in lacustrine aquaculture, using 16S rRNA gene sequencing data from the fecal samples of 28 teleost fish and 3 water-sediment samples. In this experimental setting, PAM clustering indicated that cluster variation was more consistent with habitat preferences (benthic vs. pelagic) than host taxonomy (Cyprinidae vs. Engraulidae) or trophic level, as confirmed by a RF-based supervised model, while taking into account the coalescence extent of environmental microbiota. On the other hand, Soriano et al. [136] implemented a computational tool based on Bayesian networks (BNs), which have been previously used for the sustainable development of aquaculture systems [137], and structured learning to model the network structure of piscine intestinal microbiota and the consequent interactions with biotic and abiotic variables associated with aquaculture systems, using 16S rRNA gene sequencing and farming conditions as input data. When tested on microbial datasets under different feeding trials, the system was able to determine the influence of farming conditions on taxa abundance and functional profiles. Given the fluctuations of the intestinal microbiota in response to host and environment factors, artificial intelligence will contribute to the improvement of natural environments as well as aquaculture sustainability.

Despite a limited use in microbiome research, artificial intelligence has been applied in aquaculture for different purposes, including fish recognition, species classification, size measurement, biomass estimation, and behavior analysis [138,139] (

Table 4. ML and DL applications in the aquaculture industry outside of microbiome research. In the algorithm column, the most performant model or model combination (denoted with +) has been reported. For information about the algorithms applied during feature selection, refer to the original paper. (ANFIS, adaptive neuro-fuzzy inference system; CNN, convolutional neural network; GB, gradient boosting; NIR-CV, near-infrared computer vision; RB, random block; RF, random forest; YOLO, You Only Look Once).

Task Type	Algorithm	Purpose	Reference
Image-based	YOLO	automatic identification of parasite infection	[140]
	CNN	automatic identification of feeding behavior	[141]
	NIR-CV + ANFIS	automatic feeding decision making	[142]
Non-image-based	GB	quantification of feeding frequency impact on ADG and FCR	[134]
	RF	automatic quantification of weight dispersion	[137]
	RB	prediction of growth performance metrics	[135]

). Interestingly, Li et al. [140] used the You Only Look Once (YOLO) model to implement a video detection system for parasite infection identification, particularly Ichthyophthirius multifiliis, Gyrodactylus kobayashii, and Argulus japonicus, using 27,930 labeled color images for model training, with goldfish (Carassius auratus) as fish model. When combining transfer learning and network freezing, the model reached an overall mean average precision (mAP) of 95.4% with a speed of 0.13 s per image in GPU time. Supported by such performances, the detection system was connected with a smart drug delivery apparatus to perform automated parasiticide dosing upon parasite number exceeding a predefined threshold. Iqbal et al. [141], instead, leveraged CNNs to develop a support decision system for automatic behavior categorization between normal and starvation states. Using 100 black scrapers (Thamnaconus modestus), fish behavior was recorded, and 2000 color images were extracted to train the neural model, which attained satisfactory results in terms of behavior recognition, with an accuracy of 98% and an AUC of 0.98. In an alternative study, however, Zhou et al. [142] proposed a feeding control architecture based on the combination of near-infrared (NIR) computer vision (CV) and an adaptive neuro-fuzzy inference system (ANFIS), using Tilapia as reference species. In this system, the NIR component quantifies feeding behavior in terms of flocking level and snatching intensity, which, together with user-determined feeding decisions, are passed to the ANFIS component to train and thus improve feeding decision prediction. Following this approach, the ANFIS module achieved 98% of feeding accuracy. As suggested by these studies, artificial intelligence, especially computer vision, has the potential to improve both fish welfare and feeding management. Indeed, with feeding expenses ranging from 30% to 70% of the total production costs [143], the latter case is of particular concern to the fish farming industry, in which underfeeding and overfeeding are known to have immediate repercussions on production efficiency (e.g., growth rate, disease resistance, and water quality) and, ultimately, economic growth [144].

In addition to image-based applications, which have substantially increased in the past few years, aquaculture has predominantly approached artificial intelligence in terms of machine learning (Table 4). For example, Huang et al. [145] combined meta-analysis and machine learning to quantify the impact of feeding frequency on average daily gain (ADG) and feed conversion ratio (FCR), using data from 145 case studies. In particular, the authors applied GB regression, with both biological parameters (habitat type, feeding habits, stomach type, fish size, cultivation period) and environmental parameters (temperature, salinity, feeding method, feeding frequency) as model predictors, reaching an R² of 0.81 for ADG and an R² of 0.90 for FCR. While meta-analysis was useful to define general feeding practices, ML models allowed the development of more tailored strategies incorporating fish characteristics and farming conditions. In another study, Young et al. [146] combined multi-omics data (proteome, metabolome, microbiome, lipid composition, and general composition) and clinical covariates (health indices, hematological parameters, and blood biomarkers) to accurately predict growth performance metrics, including feeding efficiency (FE), in 28 Chinook salmons (Oncorhynchus tshawytscha). With the use of Random Block (RB), an omics-oriented variant of the RF algorithm [147], as regressor model, they showed that the integration of multiple data allowed for more accurate predictions, with major contributions coming from proteomic and metabolomic data, while also highlighting modeling difficulties in FE prediction. On the other hand, Navarro et al. [148] proposed an automated tool to address the issue of fish weight dispersion in marine aquaculture, using 486 European sea bass from in vivo trials for method validation. Supported by previous studies on sea-based fish farming [149], the authors applied discrete event system (DES) to mathematically simulate feed-fish-water dynamics. Then, DES simulation data were used to train an RF regressor for accurate growth distribution prediction, which, when validated on experimental data including initial and final weight, yielded promising results for future applications as an alternative to traditional grading procedures. Although currently outnumbered by DL-oriented research, machine learning continues to represent a foundational approach in aquaculture studies.

5. Overview of Current Limitations and Future Directions

Despite remarkable achievements in recent years, the application of artificial intelligence in microbiome studies continues to be faced with methodological constraints [20,27,40,48]. On the one hand, the coverage depth of microbiome analysis is determined by data collection, which generally relies on metabarcoding and shotgun metagenomics as sequencing technologies, despite having different properties in terms of taxonomic resolution and functional annotation availability. Furthermore, data specificity depends on both anatomical region and sampling procedure. On the other hand, machine learning introduces additional requirements, namely data preprocessing and feature selection, due to the peculiarities of microbiome data (e.g., dimensionality, redundancy, noise, sparsity, and size). During the modeling phase, then, researchers need to determine the appropriate combination among predictive algorithms, estimation protocols, and performance metrics, while, at the same time, taking other concerns (e.g., data leakage, class imbalance, and hyperparameter tuning) into consideration. While deep learning innately combines feature extraction and modeling to improve model efficiency on high-dimensional datasets, automatic ML (AutoML) has also emerged as a promising solution to streamline the ML pipeline. Moreover, researchers have to assess the biological relevance of model predictions, whose interpretation can be assessed through inherently interpretable algorithms or explainable frameworks. While valuable in deriving information about the most influential features in predictive outcomes for the purpose of model generalization and transferability, XAI outcomes must nonetheless be rigorously validated to ensure their reliability [150]. In the context of microbiome research, both in vitro and in vivo validation approaches are especially recommended [151], given the growing promise of ML models in elucidating host-microbiome interactions and the increasing use of ML approaches as non-invasive tools for personalized interventions. Nevertheless, a standardized framework to assess the trustworthiness of XAI explanations has yet to be defined. In this regard, ongoing mathematical research efforts have been laying the groundwork to advance the field from observational associations toward experimentally grounded causal inference [152]. As standardization efforts continue to be made to ease the application of artificial intelligence in microbiome studies, future research is expected to deepen our understanding of the underlying mechanisms of this ecosystem, including understudied communities, such as eukaryotes and viruses [48], with even greater expectations coming from holo-omics [153,154].

While artificial intelligence integration in microbiome research has introduced noticeable challenges that compound those already associated with microbiome analysis, the availability of both data and code represents a foundational component of robust data analysis, serving as an essential mechanism for ensuring transparency and reproducibility through external validation [155]. On the one hand, source code sharing (particularly preprocessing protocols, feature extraction pipelines, and model implementations) facilitates both outcome replication and methodology adaptation. On the other hand, data access enables algorithm benchmarking, bias identification, and procedure refinement across different populations and environmental contexts. In this review, the most representative studies from each research domain were screened to assess the availability of data (

Table 5. Summary of data availability in the scientific studies using artificial intelligence presented in this review. (See Table 1, Table 2, Table 3 and Table 4).

Research Domain	Data Available	Data Unavailable
Human health	[29,30,31,34,36,37,38]	[32,33,35]
Animal husbandry	[71,72,73,74,75,76,77,79]	[78]
Aquaculture	[134,136]	[135,140,141,142,145,146,148]

) and code (

Table 6. Summary of code availability in the scientific studies using artificial intelligence presented in this review. (See Table 1, Table 2, Table 3 and Table 4).

Research Domain	Code Available	Code Unavailable
Human health	[34,37]	[29,30,31,32,33,35,36,38]
Animal husbandry	[71,72,75,79]	[73,74,76,77,78]
Aquaculture	[136]	[134,135,140,141,142,145,146,148]

), revealing interesting trends. For instance, data are more frequently shared, generally due to journal guidelines requiring deposition in publicly accessible repositories. Although exceptions to this trend can be observed in human research, where privacy concerns may limit data dissemination, the aquaculture domain continues to exhibit restricted data sharing practices. In contrast, code availability remains comparatively limited, with shared source code primarily found in studies published by journals that explicitly require code submission or in projects involving open-source applications for public use. In the current research landscape, the absence of open access to data and code constitutes a substantial barrier to transparency and reproducibility in AI-driven microbiome studies, with the risk of undermining scientific trust and ultimately hindering the translation of computational insights into meaningful outcomes [156].

When focusing on the gastrointestinal microbiome, research efforts have primarily concentrated on human health and, in more recent years, animal health, especially livestock welfare. As concerns human health, artificial intelligence has witnessed widespread application in many fields, particularly healthcare, where it continues to elucidate both the interactions among intestinal microbes and the interrelationships between microbiomes colonizing different organs [27,157]. With microbiota dyshomeostasis as an emerging contributor to human diseases, advanced computational approaches have the potential to assist healthcare professionals in many respects, such as microbial signature detection, patient stratification, susceptibility prediction, and treatment response [158]. Indeed, the individual composition of the intestinal microbiome has made a strong case for personalized medical approaches, despite application in clinical practice still being limited by model generalizability [159,160]. In the case of animal health, on the other hand, artificial intelligence has been leveraged to increase livestock productivity, with a significant emphasis on ruminants, swine, and poultry [68]. However, microbiome research in livestock species has been consistently biased toward a small number of globally distributed varieties, at the expense of local breeds with high resistance to harsh environments. Nevertheless, livestock welfare remains an essential issue due to environmental and public health concerns related to greenhouse gas emission, food-borne pathogen spread, and antibiotic resistance rise [161]. When considering the ecological complexity of the microbial communities colonizing the gastrointestinal tract, artificial intelligence has afforded us the opportunity to investigate poorly characterized mechanisms in both humans and animals [162]. Moreover, in an increasingly interconnected world, these studies must be reframed within the ongoing technological revolution brought about by the Internet of Things (IoT) paradigm, which has been driving commendable engineering efforts for the development of innovative data collection technologies, ranging from miniaturized networks in human healthcare [163] to remote systems in animal farming. In the latter case, impressive technological alternatives are currently under evaluation, such as robotic monitoring tools for animal husbandry [164] and cutting-edge prototypes for unmanned microbiome sample collection from aquatic species [165] that could then be repurposed for future application on terrestrial species.

While livestock species have gathered noticeable attention, the fish microbiome still remains vastly underexplored, presenting an untapped potential that could benefit human health. On the one hand, fish intestinal microbiomes are under examination as a possible source of novel biomolecules to address multidrug resistance in human diseases. With the marine ecosystem as the most promising sourcing ecosystem, bacterial biomolecules have attracted particular interest due to the specific adaptations caused by the environmental stresses encountered by marine bacterial communities, with antimicrobial peptides (AMPs), a subclass of ribosomally synthetized and post-translationally modified peptides (RiPPs), as prospective alternatives to conventional antibiotics [58,166]. With general-purpose pipelines for RiPP identification currently available [167], research studies have started to focus on the fish intestinal microbiota for AMP discovery, as confirmed by preliminary results from ML-guided metagenomic screening in zebrafish [168]. On the other hand, fish intestinal microbiomes have been proposed as alternative models to study host-microbe and host–pathogen interactions in their human counterparts, in an effort to better understand the underlying mechanisms of different pathological conditions and identify more effective therapeutic targets [60]. Despite being evolutionarily distant species, such knowledge translatability is favored by the significant overlap between the intestinal microbiota of zebrafish and mammals, with additional support from comparative genomics, which estimates 70% of human genes having at least one zebrafish homolog and 80% of human disease-related genes having at least one zebrafish ortholog [169,170]. Moreover, understanding the distribution of microbial communities in fish organs is gradually reshaping our current knowledge of the human body, as recently demonstrated by the discovery of residential microbial communities in healthy salmonid brains [66], thereby raising reasonable questions of whether such a condition could also be found in humans [67]. However, the use of artificial intelligence in this direction has not yet been explored. Although the composition of the fish microbiome is influenced by environmental and biological factors, this ecosystem offers a valuable window into host-microbiome-environment interactions, which will have significant implications in both fish health and conservation efforts [70,171].

While prioritizing productivity, the aquaculture industry has experienced an inconsistent use of artificial intelligence when compared to other animal farming fields. Indeed, fish production is often influenced by management practices, which are currently based on the manipulation of different parameters, such as diet, environmental conditions, and host genetics. Within this context, the manipulation of the intestinal microbiome through feed additives and novel therapeutics represents a traditional solution to optimize both health and yield in aquaculture systems [55,115]. Nevertheless, at the same time, several research studies have begun to focus on the development of feed formulations with alternative ingredients (e.g., worms, insect larvae, and animal by-products) to reduce the current dependence on fish meal and fish oil as nutritional sources [124,125,132,172,173,174]. However, artificial intelligence has found limited application in this direction, with noticeable examples currently confined to contamination monitoring in aquatic environments, microbial clustering according to trophic levels, and network-based modeling of microbiome relationships with rearing conditions. Such limited application scenarios can be sufficiently explained by the recent interest in fish microbiota in aquaculture, which has traditionally applied artificial intelligence as machine learning for the evaluation of animal growth and health status, and computer vision for the investigation of feeding behavior and disease detection [175].

Similar to the ongoing technological innovations in other farming sectors [176], the adoption of cutting-edge technologies is expected to play a crucial role in the transition toward smart fish farming, particularly precision fish farming, thereby allowing for both sustainable aquaculture and reduced human intervention [177,178]. However, this transition is currently hampered by several industry shortcomings, such as the lack of standardization in data formatting and data-sharing protocols, in addition to the developing costs associated with the necessary infrastructures, whose implementation will require to possess transdisciplinary knowledge (e.g., data analysis, software development, and aquaculture expertise) and address security concerns (e.g., data storage, data transmission, and data privacy) [179,180]. In this landscape of interconnected devices, artificial intelligence holds immense potential for both the improvement of farming practices and the automation of system management, which will ultimately allow to increase profitability and sustainability in the aquaculture production cycle [181].

While the research domains investigated in this review have been progressively incorporating IoT as architectural paradigm, albeit at varying rates, the ongoing technological advancements are already steering these fields toward more integrated frameworks, including the Internet of Everything (IoE). This emerging framework builds upon IoT by embracing a holistic perspective that integrates different domain-specific specializations through the unification of devices, data, processes, and organisms into a cross-domain intelligent network, thereby maximizing connectivity on a global scale. Nevertheless, the convergence of heterogenous technologies across different scales and environments unavoidably introduces challenges (namely, energy efficiency, ubiquitous connectivity, interoperability, miniaturization) requiring the establishment of new core attributes (multi-modality, modularity, tunability, scalability) [182]. Within this architectural framework, together with artificial intelligence, several innovative technologies are likely to drive this transition, such as cloud computing, edge computing, blockchain, advanced connectivity, and digital twins. While some of these technologies have already seen preliminary applications in smart agriculture and livestock monitoring [183], their integration into microbiome research could revolutionize the field by enabling real-time data acquisition, environmental monitoring, and intelligent analysis of microbial ecosystems. In human health, for instance, digital twins could serve as simulated models of individual microbiomes to test personalized interventions, including probiotic treatments, dietary modifications, or pharmaceutical responses. In animal husbandry, real-time data from wearable collars and automated feeding systems could be correlated with microbiome fluctuations, facilitating the early detection of infections or stress before clinical symptoms manifest. In aquaculture, IoE systems could detect early signs of microbial imbalance in tanks, enabling timely interventions. Despite these applications being purely speculative, IoE marks a substantial advancement in computational infrastructure, and future technological progress is likely to yield innovations that presently lie beyond our conceptual horizon.

The extensive technological integration characterizing this interconnected landscape has significantly facilitated cross-domain synergies, including knowledge transferability across research fields. Within animal research, especially aquaculture, which has garnered renewed attention in recent years, comparative studies between human and fish microbiota have offered valuable insights. Indeed, these studies have enhanced our understanding of fish microbiota composition in relation to habitat and diet [50], while, at the same time, contributing to the reassessment of human neural physiology following the discovery of living microbial communities in healthy salmonid brains [66]. Alongside these comparative approaches, which have also helped establish zebrafish as a viable model organism for investigating the human microbiome, the increasing convergence between human health and aquaculture has enabled methodologies to be successfully transferred across domains. For instance, metagenomic techniques, which are routinely used in human microbiome research for different purposes, particularly antibiotic resistance surveillance, have been adapted to investigate resistome transfer in non-intensive aquaculture environments, given current public health concerns regarding the potential transmission of resistomes from aquaculture environments to the human intestinal microbiota [184]. Alternatively, microbiome manipulation, which has emerged as a more promising holistic approach to enhance intestinal health for disease treatment in humans [3], has been progressively introduced in aquaculture by considering the farmed animal and its surrounding environment as an interconnected system. This approach is most evident in efforts to influence fish microbiota through dietary supplementation, although ongoing research is evaluating whether similar outcomes could be obtained via water microbiome manipulation, thus indirectly shaping the microbial communities of farmed species [185]. Additional techniques that have also been successfully transferred across domains include metatranscriptomics and environmental DNA analyses [186,187,188]. While the transferability of these methodologies inevitably necessitates certain adjustments to account for physiological and evolutionary differences between humans and animals, particularly fish, domain datasets can take advantage of common suites of AI algorithms, which are now frequently employed across different microbiome research domains (Figure 2) thanks to the model-agnostic nature of the underlying mathematical frameworks. Nevertheless, careful result interpretation remains essential, especially when applying advanced strategies such as transfer learning with pretrained models across species, given the inherent differences in microbial composition.

While knowledge transferability is a well-established practice in scientific research, framework transferability, particularly those based on artificial intelligence, represents a more advanced and less frequent form of cross-domain integration. For instance, the modeling framework implemented by Soriano et al. [136] to examine the interactions between fish microbiome and biotic/abiotic variables within aquaculture rearing systems can be extended to model microbiome-host network relationships in other vertebrate organisms, including humans, thus offering valuable information on taxa interactions that characterize complex microbial ecosystems, including the gastrointestinal microbiome. Conversely, DeepMicro, a DL model introduced by Oh et al. [189], enables the accurate detection of bacterial species from metagenomic sequencing data, allowing the identification of multiple pathogens, including those characterized by difficult culturability or low abundance. Although DeepMicro has principally been used in human health, it is plausible that the same model could be adapted for similar uses in aquaculture microbiome datasets to predict associations between microbial compositions and disease conditions [19]. Therefore, methodological cross-contamination between research domains can occur in a bidirectional manner. On the one hand, the extensive research in human health provides a rich source of transferable tools for aquaculture studies. On the other hand, insights derived from aquatic ecosystems can refine our understanding of microbiome-host relationships in humans. This reciprocal exchange underscores the potential of framework transferability to bridge disciplinary boundaries and foster integrative scientific advancement.

6. Conclusions

Over the last several decades, microbiome research has undergone a profound transformation. From a notional perspective, the field has progressively refined definitions to accommodate increasingly nuanced viewpoints, as attested by the differentiation between the microbiota, the microbiome, and the pathobiome. From a methodological perspective, the integration of artificial intelligence has significantly advanced analytical capabilities, complementing traditional bioinformatics pipelines routinely used in microbiome analysis. In particular, machine learning has emerged as the preferred approach due to implementation simplicity and computational efficiency, although the generation of large-scale, high-dimensional datasets has also accelerated the integration of deep learning, which has enhanced the scalability and automation of data processing workflows. At the same time, the development of XAI frameworks has attempted to mitigate the black-box nature of AI models, even though the biological validity of XAI outcomes remains contingent upon empirical validation. While artificial intelligence has undeniably been a transformative force in microbiome research, it has also introduced algorithm-based challenges that need to be addressed to ensure accurate, unbiased, and reproducible prediction.

Human-centered microbiome research has traditionally dominated the field due to substantial funding allocated through numerous international initiatives. Within this context, artificial intelligence has played an essential role in elucidating the contribution of microbial communities to the pathogenesis of conditions currently burdening healthcare systems. This domain has hence produced numerous studies that have yielded high-accuracy predictive models with potential future applications in clinical practice. Nevertheless, the successful translation of these models into clinical settings still requires further validation across diverse populations and environmental contexts to enhance model generalizability and ensure unbiased performance across heterogeneous patient cohorts.

Building on the insights gained from human microbiome research, similar methodologies have been increasingly transferred to animal husbandry. Within this context, artificial intelligence has primarily been directed toward enhancing animal productivity and optimizing farming practices. Nonetheless, the relevance of microbiome research in animal models is also supported by the widespread use of animal species as model organisms for human investigation. Furthermore, within the one-health approach, the growing risk of disease transmission from animals to humans has further emphasized the interconnectedness of human, animal, and environmental health. Such interrelationships highlight the necessity of considering ecological connections across species and environments when designing microbiome studies that implement artificial intelligence.

Within the broader scope of animal research, aquaculture has emerged as an expanding sector, whose economic growth has brought renewed interest in microbiome research, particularly concerning fish species. The analysis of fish microbiota imposes unique challenges because of biological and environmental factors; however, artificial intelligence has begun to shed light on the fish microbiota, which remains significantly understudied. This stands in contrast to the longstanding use of artificial intelligence, particularly machine learning, in aquaculture, where its applications have focused on production management tasks, such as biomass estimation, behavior monitoring, species classification, and feeding frequency optimization. Nevertheless, aquaculture represents a promising frontier for microbiome research. On the one hand, the intestinal microbiome of marine fish is being sourced for novel bioactive compounds, including those with the potential to combat antibiotic resistance. On the other hand, comparative genomic studies have revealed a bidirectional relationship between fish and humans: human research informs the understanding of fish microbiota composition, while fish models offer valuable insights into microbiome manipulation strategies and the pathogenesis of various human diseases.

As microbiome research progressively integrates sophisticated computational architectures, including IoT and IoE, it is essential to recognize that, while artificial intelligence does not yet offer definitive answers to complex biological questions, it has nonetheless provided researchers with powerful and efficient tools to uncover patterns that may elude traditional bioinformatics approaches. In light of the promise of this interconnected landscape, however, several challenges remain, such as the need for robust data standardization, enhanced interoperability, and stringent cybersecurity. At the same time, these transformative technologies are fostering cross-domain synergies, enabling both knowledge transferability and framework transferability. This convergence is thus poised to redefine the boundaries of scientific inquiry, ushering in a paradigm shift toward more integrative, data-driven research methodologies.