AI-empowered human microbiome research Authors: Tian Zhou, Fangqing Zhao Published: 2025 ABSTRACT AbstractRecent advances in high-throughput microbiome profiling have generated expansive data sets that offer unprecedented opportunities to investigate the role of microbes in human health. However, the complexity and high dimensionality of these data present significant analytical challenges that often exceed the capabilities of traditional computational methods. Artificial intelligence (AI), encompassing both classical machine learning and modern deep learning approaches, has emerged as a powerful solution to these challenges. In this review, we systematically explore AI-driven methodologies in microbiome research, including clustering algorithms, dimensionality reduction techniques, convolutional and recurrent neural networks, and emerging large language models. We assess how these approaches enable the extraction of meaningful biological patterns from complex microbial data from a multiscale perspective, facilitating insights into community dynamics, host–microbe interactions and functional genomics. Additionally, we explore the transformative impact of AI on translational applications across both academic research and real-world clinical settings, including disease diagnostics, therapeutic development and precision microbiome engineering. By critically evaluating the current capabilities and limitations of AI in this context, this review aims to chart a path forward for the integration of AI into microbiome research, ultimately accelerating innovations in personalised medicine and deepening our understanding of host–microbiome relationships.Keywords: Microbiome, AI (Artificial Intelligence) KEY MESSAGES Artificial intelligence (AI) offers a paradigm shift from traditional microbiome analysis, enabling higher-resolution, scalable and more predictive insights into complex host–microbiome interactions. Artificial intelligence (AI) offers a paradigm shift from traditional microbiome analysis, enabling higher-resolution, scalable and more predictive insights into complex host–microbiome interactions. AI methods are integral throughout the microbiome analysis pipeline, performing diverse tasks from multi-omic data integration and preprocessing to feature extraction and classification, effectively addressing data heterogeneity and complexity. AI methods are integral throughout the microbiome analysis pipeline, performing diverse tasks from multi-omic data integration and preprocessing to feature extraction and classification, effectively addressing data heterogeneity and complexity. By modelling microbiome data across biological scales, AI enables multiscale understanding of microbial function and its regulatory dynamics. By modelling microbiome data across biological scales, AI enables multiscale understanding of microbial function and its regulatory dynamics. AI-driven models are revolutionising microbiome-based diagnostics and therapeutics, enabling biomarker discovery, disease prediction, patient stratification and the design of personalised interventions. AI-driven models are revolutionising microbiome-based diagnostics and therapeutics, enabling biomarker discovery, disease prediction, patient stratification and the design of personalised interventions. Despite its transformative potential, AI in microbiome research faces challenges in interpretability, generalisability and data governance, demanding robust frameworks and cross-disciplinary collaboration for responsible advancement. Despite its transformative potential, AI in microbiome research faces challenges in interpretability, generalisability and data governance, demanding robust frameworks and cross-disciplinary collaboration for responsible advancement. Introduction Micro-organisms, including bacteria, archaea, fungi and viruses, inhabit every niche of the human body, forming diverse, site-specific and dynamic ecosystems that are crucial for human health and disease throughout the human lifespan.1 2 Disruptions to these microbial communities, a state known as dysbiosis,3 have been linked to various conditions, including IBD,4 allergies,5 asthma6 and several types of cancer.7 Over the past few decades, extensive efforts have been devoted to uncovering the mechanisms by which microbial communities influence host physiology and pathology.8 The advent of high-throughput omics technologies has greatly facilitated our ability to characterise microbiota across genomic, transcriptomic, proteomic and metabolomic layers9 (figure 1a). Amplicon sequencing technologies, including 16S rRNA gene sequencing and internal transcribed spacer sequencing, remain widely used due to their cost-effectiveness and scalability, while metagenome sequencing offers a more comprehensive view,10 enabling high-resolution taxonomic and functional characterisation.11 Long-read sequencing improves assembly and detection of structural variants; however, its adoption is currently limited by higher costs and elevated error rates.12 As metagenomic data sets continue to expand rapidly, effectively managing, integrating and interpreting these complex data sets has become a key bottleneck in microbiome research. Frameworks for human microbiome investigation. (A) Hypothesis-driven and statistics-based microbiome investigation. This approach begins with sample collection from diverse human habitats, followed by multi-omics sequencing (metagenomics, metatranscriptomics, metaproteomics, metabolomics). Raw data are processed via assembly tools to generate contigs, annotated taxonomically and functionally using reference databases, and analysed statistically. Machine learning (ML) models (eg, classifiers, regressors) are applied to high-dimensional data for stratification, prediction and latent feature extraction. (B) Data-driven and AI-based microbiome investigation. Multicohort, multi-omics data sets are harmonised in centralised repositories. AI models (deep learning, generative AI) extract microbial features, decode complex interactions, predict dynamics and link microbiome signatures to clinical outcomes. AI, artificial intelligence; LDA, linear discriminant analysis; PCoA, principal coordinate analysis. Traditionally, microbiome research has been predominantly driven by statistical and alignment-based computational methods, typically structured into upstream and downstream analytical stages. Upstream processing typically involves a combination of alignment-based tools, k-mer-based classification methods and assembly pipelines. These workflows encompass quality control, host-derived sequence removal, taxonomic classification and functional annotation.13,16 For instance, k-mer-based approaches such as Kraken 2 enable fast taxonomic profiling by leveraging short sequence patterns rather than full alignments.14 Meanwhile, assembly based workflows reconstruct metagenome-assembled genomes (MAGs) from short reads to enhance taxonomic and functional resolution.17 However, alignment-based strategies face limitations in scalability and are highly dependent on the breadth and completeness of reference databases.18 Downstream analyses often employ statistical techniques on abundance tables to identify microbial features linked to host phenotypes.19 These approaches have advanced microbiome research by uncovering disease associations and identifying potential therapeutic targets (figure 1a). To better model the microbiome, early studies aiming to understand microbial ecology often employed mathematical modelling to explore community dynamics and interspecies interactions.20 Building on this foundation, the application of deep learning (DL), a major branch of artificial intelligence (AI), has transformed microbiome research by enabling more sophisticated modelling of complex biological sequences. Techniques such as convolutional neural networks (CNNs), recurrent neural networks (RNNs) and long short-term memory networks (LSTMs, a sophisticated RNN variant that employs gating mechanisms to mitigate gradient vanishing or exploding issues, thereby enhancing long-term dependency modelling) have been increasingly applied to microbiome data sets. Unlike traditional alignment-based methods, DL approaches can learn high-dimensional representations directly from raw sequence data, capturing intricate evolutionary signals and uncovering associations between microbial communities and host phenotypes. In this context, the emergence of AI offers powerful new tools for unravelling such complex microbiome data, presenting a timely and promising direction for both basic research and translational applications in human microbiome science. In this review, we follow a hierarchical terminology where AI encompasses a broad class of computational techniques that mimic human intelligence. Machine learning (ML) is a subfield of AI focused on learning from data, while DL is a specialised subset of ML that uses multilayered neural networks. The terms are distinguished and applied consistently as shown in table 1. We outline a paradigm shift in microbiome research from conventional statistical and alignment-based approaches to AI-driven methodologies and summarise strategies across the analytical pipeline, from data preprocessing and representation to downstream interpretation. We further evaluate AI applications at multiple biological scales, spanning microbial communities to molecular functions, and highlight recent advances in generative AI, particularly large language models (LLMs), in the rational design of microbial systems. Finally, we discuss the translational potential and application limitations of AI in microbiome research, offering critical insights into its future development and responsible integration. Table 1Summary of key AI-related terms and methods used in this reviewTermFull name/categoryDefinitionFeaturesCommon use cases in microbiome researchAIArtificial intelligenceA broad field of computer science focused on building systems capable of intelligent behaviourIncludes ML, DL and expert systems, etc.General prediction, decision supportMLMachine learningA subfield of AI that allows algorithms to learn patterns from data without being explicitly programmedSupervised/unsupervised learning; includes support vector machine, random forest, etc.Classification of microbial taxa, disease prediction, clustering samplesClusteringUnsupervised machine learning methodGroups data points based on similarity without predefined labelsK-means, hierarchical clustering, etc.Microbial community type (enterotype) detectionDimensionality reductionUnsupervised machine learning methodProjects high-dimensional data into a lower-dimensional spacePCA, t-SNE, UMAP; aids visualisation and noise reductionVisualising microbiome samples or gene feature spacesDLDeep learningA subfield of ML based on artificial neural networks with multiple layersRequires large data; strong in feature learning and pattern recognitionFunctional prediction, sequence modelling, metagenomic annotation(D)NN(Deep) neural networkA network of interconnected computational units (‘neurons’) used for learning complex patternsForms the basis of DL; includes CNN, RNN, etc.Microbial phenotype prediction, host–microbe interaction modellingCNNConvolutional neural networkA DL architecture specialised for spatial data, such as images or sequence patternsExtracts local patterns via convolutional filtersImage-like microbial analysis, gene sequence motif detectionRNNRecurrent neural networkA DL model that processes sequential data by maintaining a memory of previous stepsSuitable for time-series or language-like dataLongitudinal microbiome dynamics, sequence predictionAEAutoencoderA type of unsupervised neural network designed to learn efficient, compressed representations (encodings) of input dataLearns low-dimensional latent representations; useful for denoising, feature reductionDimensionality reduction of microbiome profiles, latent space clustering of microbial communitiesTransformerSelf-attention-based DL architectureA model that processes all sequence elements simultaneously via attention mechanismReplaces RNNs in many NLP tasks; highly scalableMetagenomic sequence embedding, protein sequence modelling (eg, ESM, ProtBERT)LLMLarge language modelTransformer-based models trained on massive corpora to model languageStrong zero-shot and transfer capabilitiesMetagenomic data mining, protein sequence understandingNLP, natural language processing; PCA, principal component analysis; t-SNE, t-distributed stochastic neighbour embedding; UMAP, uniform manifold approximation and projection. Summary of key AI-related terms and methods used in this review NLP, natural language processing; PCA, principal component analysis; t-SNE, t-distributed stochastic neighbour embedding; UMAP, uniform manifold approximation and projection. From traditional methods to AI approaches in microbiome research Traditionally, culture-based methods have been the gold standard for studying micro-organisms, providing experimental validation and enabling landmark discoveries. However, replicating the complex and diverse conditions of natural microbial ecosystems in vitro remains challenging, leaving many human-associated microbes uncultured and poorly characterised. Metagenomics provides a culture-independent alternative by directly sequencing DNA from microbial communities. As a consequence, the advent of high-throughput sequencing technologies has led to an exponential accumulation of microbiome data, transforming the field from hypothesis-driven experimentation to large-scale data-driven analysis. Managing and interpreting these complex, high-dimensional data sets necessitates the development of advanced analytical and computational strategies.21 To facilitate data accessibility, raw sequencing data are routinely deposited in public repositories such as the Sequence Read Archive22 and European Nucleotide Archive.23 Curated and reprocessed data sets have been further organised into dedicated resources like GMrepo24 and MAMI,25 thereby improving the interpretability and reuse of large-scale data sets (figure 1b). However, traditional bioinformatics tools often struggle to derive meaningful insights from these heterogeneous data sets. The limitation has accelerated the adoption of AI, particularly ML, which encompasses unsupervised, semisupervised and supervised techniques. ML models have demonstrated remarkable capabilities in integrating, processing and extracting key biological insights, primarily from taxonomic abundance and functional annotation tables, significantly improving analytical efficiency.26 Among these, DL leverages deep neural networks (DNNs) to perform automated feature extraction from large-scale raw sequencing data, enabling end-to-end learning and reducing reliance on manual feature selection (figure 1b). As microbiome research advances, our understanding of microbial genomes, functions and ecological roles has deepened, giving rise to comprehensive, annotated databases such as Genome Taxonomy Database27 and Kyoto Encyclopedia of Genes and Genomes.28 These resources provide essential references for characterising unknown species and functional elements. The availability of labelled microbiome data has further enabled DL models to learn complex patterns and uncover associations between microbial communities and host physiology. Despite their superior predictive performance, DL models usually require extensive labelled data sets for training, suffer from limited interpretability and may be prone to overfitting, especially when applied across diverse populations or environments.29 Recently, LLMs, initially developed for natural language processing (NLP),30 have emerged as powerful tools for biological sequence modelling and knowledge extraction.31 Their adaptation to life sciences is facilitated by structural parallels between language and biological sequences.32 For instance, ESM2 (esm-650M), pretrained on the UniRef50 data set comprising over 50 million protein sequences,33 has effectively learnt evolutionary representations, achieving high-accuracy protein structure prediction with reduced inference time compared with AlphaFold2.34 Notably, ESM2 has been applied to predict the structures of over 617 million metagenomic proteins, forming the ESM Metagenomic Atlas (https://esmatlas.com/), which provides large-scale structural insights into previously uncharacterised microbial proteins.33 Beyond structure, LLM-based approaches have also facilitated functional and regulatory inference in microbial genomes. The gLM model integrates protein sequence embeddings with genomic context to predict gene functions and transcriptional regulatory elements in uncultured microbes, thus bridging protein-level language models and genome-scale inference.35 In parallel, generative LLMs have shown promise for microbiome-centric synthetic biology. For instance, Evo, a large-scale autoregressive model, enables the de novo generation of proteins guided by fitness constraints and evolutionary plausibility,36 while megaDNA, a lighter-weight transformer-based model, can generate new bacteriophage genome fragments.37 These innovations reflect a broader shift towards integrating LLMs into core microbiome analysis pipelines, extending beyond prediction to generation, annotation and design. In the era of data-intensive and knowledge-driven biological discovery, AI-based methodologies offer scalable, versatile and increasingly interpretable frameworks for tackling complex challenges in microbiome research. These approaches enable robust metagenomic data denoising,38 refined taxonomic profiling39 and accurate MAG binning.40 LLMs, in particular, offer unique advantages in deciphering large-scale, unlabelled data sets41 and extracting valuable insights from extensive biological knowledge bases.42 By bridging structured omics data with unstructured biological knowledge, LLMs are reshaping our understanding of microbial communities and host interactions. Their integration promises deeper mechanistic insights and supports precision microbiome-based interventions. Microbiome data preparation and preprocessing To address the challenge of traditional bioinformatics in handling complex and heterogeneous microbiome data, AI-based approaches are increasingly adopted for data processing, representation and interpretation, offering more efficient and scalable solutions for large-scale integrative analysis (table 2). For instance, a critical first step in meta-analyses is integrating data sets across diverse cohorts or studies. However, differences in experimental conditions, including sampling procedures, time points and geographical locations, introduce systematic batch effects that can significantly skew data analysis43 (figure 2). These unavoidable biases have prompted the development of computational methods aimed at minimising their impact. Statistical and ML approaches, such as ConQuR44 and PLSDA-batch,45 have been proposed to address these issues. Additionally, DL methods like transfer learning are employed to mitigate regional variations in microbiome-based diagnostics. Wang et al proposed a ML framework that integrates neural networks and transfer learning to reduce regional effects, improving cross-regional microbial diagnosis.46 In addition, recent AI-driven methods have enhanced species-level resolution in microbiome profiling. For example, sylph, a model based on zero-inflated Poisson statistics, corrects for average nucleotide identity bias under low sequencing coverage, thereby improving the accuracy of taxonomic assignment.39 Table 2Summary of representative AI methods in human microbiome researchMethod/toolModel architectureApplication scenarioKey featuresPerformance metric/outcomeStrengths and limitationssylph39Zero-inflated Poisson statisticsHigh-resolution taxonomic profilingCorrects average nucleotide identity biasEfficient profiling; faster (>10-fold) than Kraken2Accurate under low coverage; limited to database completenessMDL4Microbiome49Multimodal DL with latent featuresDisease (IBD, T2D, CRC, etc) classification from gut metagenomesIntermediate integration of omics dataAccuracy: IBD 0.98, T2D 0.76, cirrhosis 0.84, CRC 0.97High classification accuracy; may require complex preprocessing and integration set-upphyLoSTM53CNN and LSTMLongitudinal microbiome analysis for disease (food allergy, preterm delivery) predictionCaptures spatial and temporal dynamicsSimulated AUC: 0.897; Real data AUCs: 0.762 (DIABIMMUNE) and 0.713 (DiGiulio)Modelling temporal microbiome shifts; requires dense time-series dataPh-CNN61CNN with phylogeny-based custom layersMetagenomic data classificationIncorporates phylogenetic structure via patristic distancesOutperformed SVM, RF and MLP on synthetic and gut microbiota data (healthy and IBD samples)Integrates phylogenetic structure; requires tree-based distance computation and tested on small cohortDeepMicrobes54CNN with one-hot and k-mer embeddingsTaxonomic classification from raw sequences (especially for uncultured or novel species)Combines sparse and local feature embeddings for end-to-end classificationOutperformed traditional tools in genus/species classificationEnables detection of uncultured species of IBD; requires large, high-quality training dataGMEmbeddings56Word embeddings (GloVe) + PCACross-study host phenotype prediction from 16S rRNA dataEmbedding-based transformation of ASV tablesImproved generalisability across six data sets involving three phenotypesEnhances cross-cohort generalisation; requires alignment and embedding stepsPopPhy-CNN60CNN and phylogenetic tree embeddingHost phenotype predictionConverts phylogenetic tree and taxa abundance into 2D matrix as CNN inputOutperformed classical ML methods in nine data sets and robust in multiclass tasksIncorporates phylogenetic structure for better interpretability and performance; limited by input requirementsMEGMA62Unsupervised manifold embedding and ConvNet (AggMapNet)Disease prediction and biomarker identificationEmbeds and clusters microbial profiles into 2D multichannel mapsIdentified biomarkers with 84% literature support across 74+ studiesRobust biomarker discovery; requires a two-step pipeline and may depend on embedding qualityGraphMB65GNN integrating contig features and assembly graph structureLong-read metagenomic binningLearns contig embeddings via graph structureYielded on average 17.5% more high-quality bins versus other binnersCaptures assembly graph; depends on quality of assembly graphDeepMicro67Autoencoders and ML classifiersFeature extraction for disease (IBD, T2D, obesity, etc) predictionLearns robust low-dimensional embeddingsAchieved higher AUC (IBD 0.955, obesity 0.659, cirrhosis 0.940)Effectively addresses high dimensionality and sparsity; requires hyperparameter tuningMetaNN77Fully connected neural networkHost disease (eg, IBD) classificationApplies a novel data augmentation technique to combat overfittingOutperformed several ML models (eg, SVM, RF, other NNs) on both synthetic and real microbiome data setsHandles synthetic and real data; less interpretableRead2Pheno58CNN and RNN and attention mechanismsRead-level phenotype prediction from 16S rRNA readsLearns biologically meaningful representationsRead-level host phenotype prediction comparable to conventional methodsEnd-to-end learning without manual feature extraction; demands large training dataAI, artificial intelligence; ASV, amplicon sequence variant; AUC, area under curve; CNN, convolutional neural network; cNODE, compositional neural ordinary differential equation; CRC, colorectal cancer; CRISPR, clustered regularly interspaced short palindromic repeats; DL, deep learning; FDA, Food and Drug Administration; FGeneBERT, a function-driven pretrained gene language model for metagenomics; GNN, graph neural network; LSTM, long short-term memory network; MEGMA, microbial embedding, grouping, and mapping algorithm; ML, machine learning; MLP, multilayer perceptron; mNODE, metabolomic profile predictor using neural ordinary differential equations; NN, neural network; PCA, principal component analysis; ProtBERT, a BERT-based model pretrained on a large corpus of protein sequences; RF, random forest; RNN, recurrent neural network; SVM, support vector machine; T2D, type 2 diabetes ; VALENCIA, vaginal community state type nearest centroid classifier. Summary of representative AI methods in human microbiome research AI, artificial intelligence; ASV, amplicon sequence variant; AUC, area under curve; CNN, convolutional neural network; cNODE, compositional neural ordinary differential equation; CRC, colorectal cancer; CRISPR, clustered regularly interspaced short palindromic repeats; DL, deep learning; FDA, Food and Drug Administration; FGeneBERT, a function-driven pretrained gene language model for metagenomics; GNN, graph neural network; LSTM, long short-term memory network; MEGMA, microbial embedding, grouping, and mapping algorithm; ML, machine learning; MLP, multilayer perceptron; mNODE, metabolomic profile predictor using neural ordinary differential equations; NN, neural network; PCA, principal component analysis; ProtBERT, a BERT-based model pretrained on a large corpus of protein sequences; RF, random forest; RNN, recurrent neural network; SVM, support vector machine; T2D, type 2 diabetes ; VALENCIA, vaginal community state type nearest centroid classifier. Artificial intelligence (AI)-driven analytical workflow for microbiome data. Step 1: Microbiome data preparation and preprocessing. Multi-omics data sets are integrated to mitigate batch/regional biases using direct/indirect combination or model fusion. Time-series models (RNNs, LSTMs) analyse dynamic patterns for host status prediction. Step 2: Feature extraction and representation learning. Embedding techniques evolve from sparse to dense/static and context-aware models. Deep learning and ML transform raw data into biologically interpretable features for tasks like microbial interaction modelling and biomarker discovery. Step 3: Data analysis and interpretation. Downstream analyses combine unsupervised/supervised learning with advanced AI models. Interpretability techniques identify key microbial factors, clarifying their roles in host health. ANN, artificial neural network; BERT, Bidirectional Encoder Representations from Transformers; LASSO, least absolute shrinkage and selection operator; LSTMs, long short-term memory networks; ML, machine learning; PCA, principal component analysis; PCoA, principal coordinate analysis; RNN, recurrent neural networks; TF-IDF, term frequency-inverse document frequency; t-SNE, t-distributed stochastic neighbour embedding; UMAP, uniform manifold approximation and projection. When dealing with multi-omics data associated with the microbiome, integrating heterogeneous data sets for downstream analysis is crucial. Several ML-based methods have been developed for this purpose, which can be broadly categorised into different integration strategies (figure 2). One common approach, early data integration, merges multi-omics data prior to model training, typically by concatenating different features into a unified matrix. For instance, a study using a gradient boosting regression model successfully integrated blood biomarkers, dietary records, anthropometric data, physical activity and gut microbiota to predict postprandial glucose responses to real-life meals.47 Another study investigating IBD performed joint analyses of metabolomic and metagenomic data, identifying disease-associated microbial species and metabolite features, and achieving accurate patient classification using random forest models.48 The study demonstrated the potential of integrating multiple omics layers for robust microbiome-based disease modelling. The second strategy, intermediate integration, transforms omics data into latent representations before feature integration. MDL4Microbiome, a DL-based approach for disease prediction, demonstrated high accuracy by leveraging multimodal data from metagenomic sequence.49 Similarly, Garcia-Jimenez et al employed a deep autoencoder (DAE)-based architecture to extract latent representations from microbial community profiles and successfully predict microbiome composition using environmental variables such as plant age, temperature and precipitation.50 This approach highlights how latent feature learning can facilitate microbiome modelling, particularly under data-scarce or low-resource conditions. Most microbiome studies rely on cross-sectional data, limiting their ability to capture dynamic host–microbiome interactions, particularly during disease progression or physiological changes. In contrast, time-series data preserve sequential and historical patterns, providing richer insights and more accurate predictions (figure 2). RNNs and LSTM networks are well suited for such tasks due to their ability to model long-term dependencies. Chen et al proposed an RNN-based framework to predict host status using longitudinal microbiome data, demonstrating the advantages of RNN approaches.51 Metwally et al used an LSTM model to predict food allergies based on infant gut microbiome profiles across 3 years.52 Sharma et al introduced phyLoSTM, integrating CNNs and LSTMs to enhance prediction accuracy in longitudinal studies.53 These advancements underscore the effectiveness of RNNs and LSTMs in capturing key features from temporal microbiome data, surpassing traditional statistical and ML-based models. Beyond RNNs and LSTMs, Temporal Convolutional Networks (TCNs) offer an effective alternative for modelling microbiome time-series data. By using dilated causal convolutions, TCNs enable parallel processing while preserving temporal order, allowing efficient modelling of long-range dependencies without the vanishing gradient problem. While their use in microbiome research is still nascent, TCNs offer a promising direction for capturing temporal microbial patterns and predicting dynamic host–microbiome interactions. Microbiome data characterisation and feature extraction Traditional ML models in microbiome research have primarily relied on normalised taxonomic abundance tables, which, while useful, often fail to capture the high-resolution patterns inherent in raw sequencing data. In contrast, DL models enable richer feature extraction through advanced embedding and representation techniques. Inspired by NLP models, word embedding techniques have been adapted to encode sequencing reads into high-dimensional representations (figure 2). One-hot encoding, a sparse embedding method, was used in DeepMicrobes, which combined one-hot and k-mer embeddings to model gut microbiome sequences. This approach enabled the discovery of novel microbiome signatures in IBD, outperforming existing tools in species and genus identification.54 To enhance the efficiency of microbiome sequence modelling and embedding, dense embedding methods such as Word2Vec and GloVe have been introduced (figure 2). Woloszynek et al developed a pretrained Skip-Gram Word2Vec model, trained on over two million full-length 16S rRNA amplicon sequences to create k-mer embeddings.55 The GMEmbeddings R package, incorporating GloVe and principal component analysis (PCA) embeddings, was designed for predictive modelling of 16S rRNA V4 gut microbiome sequencing data, contextualising microbial occurrences while reducing dimensionality.56 Contextualised representations have further enhanced metagenomic analyses. For instance, Metagenome2Vec, a self-supervised representation learning model, captures both global structural properties and local contextualised features, aiding in pathogen detection from clinical samples.57 Zhao et al introduced Read2Pheno, an attention-based DL network that encodes sequences into dense and biologically meaningful representations, providing insights into microbial functions.58 Similarly, FGeneBERT models gene sequence-function relationships using contextualised metagenomic embeddings and a context-aware tokeniser.59 Beyond sequence embeddings, various structural encoding strategies, including tree, network and graph-based embeddings, have been integrated into DL frameworks. PopPhy-CNN incorporates phylogenetic structures to predict host phenotypes by representing phylogenetic trees as 2D matrices populated with microbial relative abundances.60 The Ph-CNN model follows a similar approach, leveraging patristic distances to embed phylogenetic trees in Euclidean space.61 Meanwhile, MEGMA, an unsupervised embedding and clustering algorithm, transforms metagenomic data into individualised multichannel 2D microbiome representations.62 Furthermore, graph neural networks (GNNs) offer powerful means to model the complex topological structure of microbiome data. By encoding microbial taxa or genes as nodes, and their interactions (eg, co-occurrence, phylogeny, functional similarity) as edges, GNNs extract context-aware node embeddings through message passing. Notably, GraphSAGE introduces inductive learning by aggregating information from a node’s neighbourhood,63 enabling generalisation to unseen microbial nodes or genes. For instance, Guo et al used a GNN-based framework to identify disease-associated taxa and genes, leveraging topological features to improve classification performance and biological interpretability.64 Similarly, GraphMB applies GraphSAGE to metagenomic binning by modelling contigs as nodes linked via sequence similarity, improving bin quality and robustness.65 These examples illustrate how GNNs can be flexibly adapted to capture microbial ecology, enhance genome reconstruction and reveal latent functional modules. Feature extraction and selection remain crucial for microbiome analysis (figure 2). Random forest classifiers are widely employed to identify microbiome signatures across different groups.66 DL-based autoencoders offer an alternative for feature selection. The DeepMicro framework integrates various deep representation learning techniques, such as shallow autoencoder, DAE, variational autoencoder (VAE) and convolutional autoencoder, to derive robust low-dimensional embeddings, enabling accurate disease predictions across five data sets and outperforming conventional approaches.67 The extracted features enhance classification and prediction tasks, accelerating model training and improving performance in downstream analyses. Microbiome data analysis and interpretation The compositionality, high dimensionality and sparsity of metagenomic data pose significant challenges for biomarker selection, differential analysis and community visualisation. These challenges are compounded by the combination of relatively small sample sizes in microbiome studies, which exacerbates the curse of dimensionality and limits the effectiveness of traditional statistical methods. Consequently, AI-based approaches are increasingly employed to overcome these limitations, offering enhanced scalability, robustness and interpretability in complex microbiome analysis. Dimensionality reduction is a common strategy for selecting abundant taxa and visualising community composition, as taxonomic abundance tables often contain many zero values.68 Common techniques include PCA, principal coordinate analysis (PCoA), t-distributed stochastic neighbour embedding (t-SNE),69 and uniform manifold approximation and projection (UMAP),70 all of which extract key features from high-dimensional data to facilitate visualisation (figure 2). While PCA and PCoA offer greater interpretability, t-SNE and UMAP are better at capturing non-linear structures, although they are less explainable. Odrzywolek et al employed a language model alongside these methods to visualise microbiome protein space, enabling alignment-free protein analysis.71 Clustering methods are also widely used to stratify microbiota into similar groups (figure 2). Hierarchical clustering, one of the simplest and most common approaches, categorises microbiota by computing pairwise distances between samples. Ravel et al introduced the concept of community state types to classify vaginal microbial communities based on composition.72 K-means clustering has also been employed, such as in Ahlqvist et al’s study, which identified five subgroups of adult-onset diabetes.73 Additionally, partitioning around medoid clustering with Jensen–Shannon distance is commonly used for enterotype analysis.74 Classification methods play a key role in linking the microbiome to host phenotypes (figure 2). VALENCIA, for instance, uses nearest centroid classification to analyse vaginal microbial communities,75 while Zhou et al trained the random forest model on gut microbiota data to identify universal IBD biomarkers and predict response to infliximab.76 Similarly, MetaNN, a neural network-based approach, achieved superior classification accuracy for host phenotype prediction using both synthetic and real metagenomic data.77 Regression models are used when variables are continuous (figure 2). Zeevi et al developed a gradient boosting regression model to predict postprandial glucose responses based on dietary and gut microbiome data.47 Read2Pheno, an attention-based deep network, encodes sequencing reads into dense, biologically meaningful representations for read-level phenotypical prediction.58 Among these approaches, dimensionality reduction and clustering are typically used for unsupervised learning on unlabelled data, while classification and regression models require well-labelled data sets for supervised learning. AI-driven microbiome insights across biological scales Beyond method development, AI plays a vital role in translating complex microbiome data into biologically meaningful insights. The selection of appropriate computational strategies not only facilitates data interpretation, but also uncovers functional patterns and mechanistic relationships within the microbiome. Increasingly, AI enables multiscale analysis, from modelling microbial community interactions to identifying keystone taxa and functional biomolecules, highlighting its transformative impact across hierarchical levels in microbiome research (figure 3). Multiscale artificial intelligence (AI) applications in microbiome data mining. (A) Community-level analysis. AI models integrate labelled multi-omics data to predict community compositions and metabolic functions. Computational frameworks link microbial consortia dynamics to host physiological outcomes and enable systematic modelling of community–host interactions through parameter optimisation. (B) Species-level analysis. AI-based models can identify keystone species and biomarkers in health-disease transitions. (C) Molecular-level analysis. Deep learning and generative AI (eg, LLMs) decode molecular mechanisms, including functional annotation and molecule generation. CDI, Clostridium difficile infection; DL, deep learning; LLM, large language models; ML, machine learning; SCFA, short-chain fatty acids. At the community level, the non-linear dynamics within microbial ecosystems necessitate mathematical models to quantitatively link community structure with function. Traditional approaches, such as Lotka–Volterra models,78 have been widely used for studying microbial interactions. For example, Stein et al extended the generalised Lotka–Volterra model to incorporate external perturbations, successfully predicting microbiota dynamics in antibiotic-treated mice and identifying alternative stable states.79 Similarly, LIMITS further applied sparse linear regression with bootstrap aggregation to infer discrete-time Lotka–Volterra models.80 Recently, DL-based approaches have been adopted to explore microbial community composition and function (figure 3a). To enable accurate species-level profiling, models like sylph leverage statistical-AI hybrid approaches to correct for sequencing biases, thereby supporting community-level ecological inference.39 cNODE, a data-driven DL-based model, predicts community compositions from species assemblages,81 while its extension, mNODE, improves metabolomic profiling predictions for microbial communities.82 Additionally, MelonnPan, based on elastic net regularisation, predicts microbial metabolic profiles and identifies key predictive taxonomic or functional features.83 Despite their predictive power, many AI models, including advanced frameworks such as cNODE and mNODE, are inherently limited by the observational nature of microbiome data, which poses substantial challenges for inferring causal rather than merely correlational relationships. To address this limitation, researchers are increasingly combining AI models with causal inference frameworks, such as those based on structural equation modelling84 or network analysis.85 These efforts are still evolving; they represent critical steps towards transforming AI from predictive tools to mechanistic inference engines capable of guiding rational microbiome applications. At the species level, AI methods provide insights into species functionality and their effects on host health (figure 3b). Raman et al introduced the ‘Ecogroup’ framework, which combines statistical covariance with matrix decomposition to identify co-fluctuating microbial networks, providing insight into microbiota development in malnourished Bangladeshi children.86 Ensemble quotient optimisation was developed to discover stable functional microbial groups despite shifts in species composition.87 AI has also facilitated microbial biomarker discovery in disease states such as obesity and Clostridium difficile infection. In one study, an ML framework identified microbiome biomarkers in obesity, uncovering the role of Megamonas rupellensis in lipid absorption.88 Furthermore, AI-driven approaches have enabled the identification of keystone species, micro-organisms that play critical roles within microbial communities. A DL-based framework has been proposed to learn microbial community assembly rules from habitat-specific microbiome samples.89 At the molecular level, AI models leverage large-scale microbiome data for functional annotation, gene mining and elucidating regulatory mechanisms90 (figure 3c). Early applications of DL in metagenomics focused on sequence identification. For example, Borgman et al developed a rapid sequence identification method with single-nucleotide resolution, demonstrating its potential for phenotypical prediction.91 DL has also facilitated the discovery of functional molecules, such as antimicrobial peptides (AMPs). An NLP-based model identified candidate AMPs from the human gut microbiome, many of which exhibited antimicrobial activity.92 Expanding on this, Santos-Junior et al curated AMPSphere, a catalogue of 863 498 non-redundant AMPs,93 further mining AMPs from human microbiomes.94 In antibiotic discovery, Wong et al identified a novel structural class of antibiotics using explainable DL,95 while another study trained neural networks on growth inhibition data sets to predict structurally novel antibiotics effective against Acinetobacter baumannii.96 Given the challenges of annotating viral genomes in metagenomic samples, a protein language model (PLM)-based method was introduced to enhance prokaryotic viral protein annotation by leveraging pretrained PLM representations.97 LucaOne, a foundational model integrating nucleic acid and protein data sets, effectively learns molecular biology principles and outperforms existing models in various downstream tasks.41 Beyond functional molecule identification, generative AI is emerging to design and optimise novel, easily synthesisable antibiotics.98 Multiscale AI methods in synthetic microbiology Recent advances in synthetic microbiology have enabled precise manipulation of microbial functions and the rational engineering of microbial communities for diverse biotechnological applications99 (figure 4). Central to this field is the integration of multi-omics data sets and biological knowledge, which presents a significant challenge for traditional experimental and computational frameworks. These approaches often fall short in navigating the vast design space and achieving timely, efficient outcomes.100 The emergence of DL and generative AI has begun to reshape synthetic microbiology by offering scalable solutions for synthetic ecology, engineered species and molecular design.101 Among these advances, transformer-based models have shown particular promise in modelling microbial genomic sequences and predicting functional elements in synthetic genomes.102 Originally developed within an encoder-decoder framework, the transformer architecture has evolved into powerful variants such as BERT (Bidirectional Encoder Representations from Transformers) and GPT (generative pretrained transformer).103,105 BERT-based models capture long-range dependencies in genomic sequences, enhancing the identification of regulatory elements, coding regions and structural variants.35 Conversely, GPT-like models can generate novel genome sequences that conform to biological constraints, thereby enabling de novo genome synthesis.106 These AI-driven strategies support synthetic genome optimisation, improve metabolic pathway engineering and accelerate microbial strain development tailored to specific applications. Artificial intelligence (AI) in multiscale synthetic microbiome. At the methodological level, ML algorithms are employed to design target microbial communities using species-level data, while LLMs analyse genomic sequences to extract latent features. These approaches facilitate synthetic consortia design, species editing/generation and molecular engineering of gene elements. At the application level, these technologies can be applied in diverse fields, including clinical interventions (eg, FMT therapy, probiotic intervention), microbial engineering and molecular design. BERT, Bidirectional Encoder Representations from Transformers; FMT, faecal microbiota transplantation; GPT, generative pretrained transformer; LLM, large language models; ML, machine learning. In both natural ecosystems and human microbiomes, micro-organisms exist within highly interactive communities that exhibit intricate interspecific dynamics and functional synergies.107 Traditional approaches to community design rely on labour-intensive, trial-and-error experiments with numerous strain combinations. AI methods address this inefficiency by leveraging high-throughput omics data and curated knowledge to predict microbial interactions, assess community stability and optimise collective function.108 ML models can rapidly identify strain combinations with complementary metabolic capabilities,109 while DL models trained on synthetic and natural microbiome data sets can infer cooperative or antagonistic relationships.110 Reinforcement learning algorithms further enhance community design by iteratively simulating and refining ecological configurations. At the individual strain level, genome editing technologies such as CRISPR have revolutionised microbial engineering.111 However, identifying optimal gene modifications within the vast combinatorial space of mutations and pathway designs remains a key bottleneck.112 AI-driven generative models, including DNNs and generative adversarial networks (GANs), facilitate large-scale prediction of gene editing outcomes and metabolic feasibility.113 These models integrate multi-omics data with literature-based knowledge to evaluate gene knockouts, overexpression strategies and multigene constructs, thus offering decision support for complex metabolic engineering tasks.114 At the molecular level, synthetic biology encompasses the design of genetic components (eg, promoters, terminators, ribosome binding sites),115,117 proteins, enzymes118 and metabolic modules.119 AI plays a critical role in accelerating this design process.120 LLMs enable the rapid generation of novel proteins and enzyme variants with optimised functionality under defined conditions.121 Deep generative models, such as GANs and VAEs, facilitate the exploration of novel sequences and structures for synthetic intermediates and small-molecule therapeutics.122 Furthermore, AI-based prediction of intermolecular interactions, including binding affinity and structural docking, significantly enhances the screening of enzyme substrates and protein-ligand interactions,123 thereby streamlining experimental validation and reducing the time from design to deployment.124 While AI-driven approaches show great promise in synthetic microbiology, several critical challenges persist. First, the experimental validation of in silico designs remains resource-intensive and often fails due to unforeseen regulatory crosstalk, host–microbe incompatibilities or emergent ecological behaviours. Second, the generalisability of trained models across different microbial environments and hosts remains limited, particularly when models are trained on biased or sparse data sets. This hinders their application across diverse microbial environments and hosts. Finally, the rapid development of synthetic microbiome engineering raises ethical and biosafety concerns,125 including ecological disruption, horizontal gene transfer and dual-use risks. Establishing regulatory frameworks and ethical principles is essential for the safe translation of AI-designed microbial systems into real-world applications.126 AI-enabled microbiome-based diagnostics and therapeutic strategies Microbiome-based therapeutic strategies have emerged as pivotal components of precision medicine, offering innovative avenues for disease management and prevention.127 However, the dynamic heterogeneity of the microbiome, unresolved causality in host–microbe interactions, and systemic physiological differences between animal models and humans collectively contribute to a substantial translational gap between basic discoveries and clinical application.128 Bridging this gap requires a systems-level approach that integrates data-driven hypothesis generation with experimental validation, enabling a continuum from statistical association to causal inference and from laboratory findings to clinical implementation.129 By harnessing large-scale, high-dimensional data sets, AI can uncover complex, non-linear patterns, aiding in biomarker discovery,130 modelling host–microbiome interactions and predicting therapeutic responses. Recent advances have applied ML and DL to enable early diagnosis, personalised interventions, prognostic modelling, and tailored dietary or probiotic therapies. Accurate disease classification and diagnosis represent one of the most well-established and clinically relevant applications of AI in microbiome research. Over the past decade, ML and DL models have demonstrated remarkable performance in identifying diagnostic biomarkers and supporting clinical decision-making131 (figure 5). These methods are particularly effective in conditions such as type 2 diabetes (T2D), IBD and colorectal cancer (CRC), where microbiome alterations are closely linked to disease onset and progression. For example, LaPierre et al employed various ML models to conduct a meta-analysis of microbiome data sets on T2D and obesity, revealing robust patterns across cohorts.132 Gou et al further developed an interpretable ML framework to identify T2D-associated microbiome features across three Chinese cohorts.133 Tataru et al proposed a predictive model trained on microbiome property data to improve IBD classification,134 while Pope et al developed a deep language model trained on large-scale unlabelled microbiome data sets for IBD prediction.135 In CRC research, interpretable ML models have been employed to identify microbiome-derived features relevant to disease classification, shedding light on molecular mechanisms underlying microbiome alterations during CRC progression.136 Beyond academic and experimental research, several AI-powered diagnostic platforms have begun to enter real-world use. BiomeSense (https://www.biomesense.com/) is a notable example, offering GutLab, a home-based, continuous microbiome monitoring system designed to provide real-time insights into gut health and disease risk through AI-driven analysis of longitudinal microbiome data. This system supports convenient, non-invasive diagnostics and is particularly suited for chronic disease monitoring and preventive care. Another company, Viome (https://www.viome.com/), has developed AI-based microbiome testing and wellness recommendation platforms, focusing on personalised nutrition and metabolic health. Artificial intelligence (AI) in microbiome-based therapy. In predictive diagnosis, various models (ML, CNNs and RNNs) analyse microbiome data to predict diseases. In therapeutic intervention, AI optimises treatment strategies, such as donor selection for FMT and phage engineering for targeted treatments. In disease prognosis, the Cox-net model, applied to time-series microbiome data, forecasts patient prognosis, aiding in the evaluation of long-term treatment outcomes. In dietary intervention, k-mer-based methods (eg, SVMs) and LLMs are employed for probiotic identification, while tree-based models analyse postprandial responses in large cohorts to offer personalised dietary recommendations. AUROC, area under the receiver operating characteristic curve; CNN, convolutional neural network; FMT, faecal microbiota transplantation; LASSO, least absolute shrinkage and selection operator; LLMs, large language models; LM, language model; LSTM, long short-term memory network; ML, machine learning; RNN, recurrent neural network; SVM, support vector machine. AI has also shown significant promise in microbiome-based interventions, particularly in faecal microbiota transplantation (FMT) and phage therapy, both of which represent innovative interventions for microbiome modulation and the treatment of microbiota-related disorders (figure 5). FMT, a key strategy for restoring gut microbiota balance, relies heavily on donor microbiota selection. AI models integrate large-scale microbiome data to assess microbiome compatibility between donors and recipients, optimising donor selection and microbial consortia to enhance treatment efficacy while minimising risks.137 For instance, AI-assisted modelling can predict the therapeutic potential of donor microbiota in disease-specific scenarios, thereby enabling personalised FMT protocols.138 Recent clinical advancements have further demonstrated the feasibility and efficacy of such strategies.139 A phase III clinical trial investigating faecal microbiota, live-jslm (RBL) for the treatment of recurrent Clostridioides difficile infection reported that the intervention not only restored microbial diversity but also achieved durable clinical remission through successful clonal engraftment of key microbial populations.140 In parallel, phage therapy, a highly specific antibacterial strategy targeting multidrug-resistant pathogens,141 has benefited greatly from the integration of AI. AI-assisted phage engineering optimises binding affinity and bactericidal efficacy, while simulation-based evaluations provide predictive insights into therapeutic outcomes, thereby enhancing clinical applications.142 Phagos (https://phagos.org/), a biotechnology company, has developed an AI-driven platform to streamline phage-host matching, reducing the traditional reliance on labour-intensive trial-and-error assays. Crucially, a recent multinational, multicentre study involving 100 cases of personalised bacteriophage therapy has provided strong clinical evidence for the real-world efficacy of phage therapy.143 This study reported high rates of infection resolution in complex, antibiotic-resistant cases, with personalised phage cocktails tailored to pathogen profiles using integrated computational and clinical workflows. Beyond diagnosis and intervention, AI plays a crucial role in the prognostic evaluation of microbiome-based therapies (figure 5). For example, the Cox-net model has been widely applied in microbiome therapy studies as a survival analysis tool integrating regularisation techniques.144 By incorporating patient microbiome profiles and clinical parameters, AI models can predict survival risk and recurrence probability post-treatment, guiding personalised therapeutic adjustments.145 This data-driven approach helps elucidate the long-term impacts of microbiome interventions and their underlying mechanisms. Moreover, emerging commercial platforms are integrating microbiome-based prognostic modelling into clinical decision support. For example, Viome has expanded beyond diagnostics to include AI-driven personalised health assessments, including longitudinal risk scoring for metabolic and gastrointestinal disorders based on microbiome changes. These developments suggest that AI not only enhances short-term diagnostic and intervention strategies but also plays a central role in forecasting clinical trajectories and in evaluating the long-term impact of microbiome-based therapeutics. Dietary interventions, including probiotic supplementation, represent a critical dimension of microbiome-based therapies due to their direct influence on microbial composition and metabolic function (figure 5). The gut microbiota plays a central role in host metabolism, and dietary patterns profoundly affect microbiome composition and metabolic responses.146 AI-driven integrative analyses of dietary, microbiome and clinical data sets have enabled the identification of personalised nutrition strategies and targeted probiotics capable of altering disease trajectories, with promising results in conditions such as prediabetes147 and obesity.148 To accelerate the discovery and functional characterisation of probiotics, computational tools have been developed that leverage AI models for large-scale microbial screening and predictive modelling. For example, Sun et al developed iProbiotics, a probiotic mining platform that leverages k-mer features within a support vector machine framework to enable rapid identification from whole-genome sequences.149 Meanwhile, Wu et al introduced metaProbiotics, a computational approach that applies language models to probiotic identification.150 A real-world example of AI-powered probiotic development is Pendulum Therapeutics (https://pendulumlife.com/), which integrates next-generation sequencing, functional genomics and AI-based screening pipelines to develop novel therapeutic probiotics targeting chronic diseases. Its flagship product, Pendulum Glucose Control, is the first clinically validated probiotic formulation shown to reduce blood glucose spikes in individuals with T2D. This formulation includes a proprietary blend of Akkermansia muciniphila, Clostridium butyricum and other butyrate-producing bacteria, which restore metabolic homoeostasis by enhancing gut-derived butyrate synthesis.151 While AI-driven microbiome-based therapies are rapidly evolving, their clinical translation is still emerging. Several companies, including Pendulum Therapeutics and BiomeSense, have advanced AI-developed microbiome therapeutics into early phase human trials. At the same time, regulatory agencies such as the US FDA have issued guidance for AI/ML-based Software as a Medical Device (https://www.fda.gov/medical-devices/digital-health-center-excellence/software-medical-device-samd), which will likely impact future approval pathways for microbiome-AI tools. In brief, the rigorous evaluation of efficacy, reproducibility and safety remains essential for real-world adoption and clinical trust. Conclusion and perspectives The rapid advancement of computational technologies has established AI as an indispensable tool in human microbiome research, reshaping the way we analyse, interpret and use high-dimensional biological data. Traditional approaches based on association inference and conventional bioinformatics pipelines often struggle to cope with the complex scale of modern microbiome data sets. In contrast, AI, particularly ML and DL, offers powerful tools capable of capturing non-linear relationships, integrating diverse data modalities and uncovering subtle biological patterns that might otherwise remain hidden. However, despite these advancements, critical challenges persist.152 Overfitting is a major concern, particularly when models are trained on sparse, noisy or non-representative data sets,21 leading to poor generalisability across populations or environments.46 Reproducibility also suffers from inconsistent preprocessing and a lack of standardised benchmarks. While traditional bioinformatics methods have scalability limits, they still offer value in contexts that require mechanistic transparency or hypothesis testing. Moreover, the ‘black box’ nature of many AI models impedes biological interpretation and clinical trust. To address this, explainable AI techniques, such as LIME,153 SHAP,154 DeepLIFT155 and attention mechanisms,156 are increasingly employed to enhance interpretability and align predictions with biological relevance. Equally important is the governance of microbiome data.157 Given its intrinsic link to human health, behaviour and even identity, microbiome data raise unique privacy concerns.158 Federated learning, a decentralised approach where models are trained across multiple institutions without sharing raw data, has emerged as a promising solution. Such strategies not only preserve individual privacy but also enable large-scale, collaborative learning across geographically and institutionally diverse cohorts. Moving forward, it will be essential to develop comprehensive data-sharing frameworks that balance accessibility with ethical integrity. Translating AI models into clinical practice brings its own set of challenges.159 These include stringent regulatory requirements, the need for rigorous validation in diverse populations and integration into existing diagnostic workflows. Limited model interpretability further hampers clinician trust, particularly in critical tasks like risk prediction or therapeutic decision-making. Addressing these barriers requires developing clinically relevant, transparent and user-centred AI systems. To advance the field, several concrete steps are needed. First, the establishment of community-wide benchmarks and curated annotated data sets would facilitate more robust model evaluation and comparison. Second, interdisciplinary collaboration between microbiologists, data scientists, clinicians and ethicists is crucial to ensure responsible model development and deployment. Third, greater emphasis should be placed on designing interpretable, generalisable and resource-efficient AI models that can operate effectively across diverse populations and ecosystems.160 In summary, the convergence of AI with single-cell sequencing, multi-omics integration and knowledge-driven inference models will unlock new layers of microbiome biology. Advances in transfer learning and language-based pretraining hold promise for making AI models more data-efficient and biologically grounded. Ultimately, a sustainable and transparent AI ecosystem, built on ethical, technical and collaborative foundations, will accelerate the translation of microbiome science into precision diagnostics, personalised therapies and preventive health strategies. Supplementary material Footnotes Funding: This work was supported by grants from the National Key R&D Program of China (2021YFA1301000, 2022YFA1303900) and the National Natural Science Foundation of China (32025009). Provenance and peer review: Not commissioned; externally peer reviewed. Patient consent for publication: Not applicable. Ethics approval: Not applicable. References Associated Data Supplementary Materials