Integrated analysis of human-mouse gut microbiota in RSV infection based on machine learning Authors: You Duan, Chen Guo, Liang Xie, Hanmin Liu, Yang Liu Published: 2026 ABSTRACT BackgroundRespiratory syncytial virus (RSV) is a leading cause of lower respiratory tract infections in children, but effective treatment options remain limited. The gut-lung axis, which highlights the role of gut microbiota in regulating respiratory immunity, provides new opportunities for developing probiotic-based therapies. However, existing studies on RSV-associated gut microbiota are often small-scale and lack systematic integration. To address this gap, we conducted a comprehensive machine learning-based analysis of gut microbiota data from RSV-infected children and mice, integrating five public datasets comprising 319 samples (154 children, 165 mice).ResultsPediatric samples were divided into control, infected, and recovery groups, while mouse samples included control and infected groups. Microbial diversity analysis revealed RSV infection disrupted gut microbiota structure in children, with reduced α-diversity in the infected group and significant β-diversity differences among groups (P < 0.001). Mice exhibited higher α-diversity than children, with distinct dominant taxa: Bifidobacteriaceae and Escherichia-Shigella prevailed in children, whereas Lachnospiraceae and Ligilactobacillus dominated in mice. Using 13 machine learning algorithms, we developed disease-prediction models at the family and genus levels, achieving superior performance in pediatric data (maximum AUC = 0.952) compared to mouse data. Cross-species analysis identified 62 family-level and 54 genus-level high-importance taxa (e.g., Bacteroidaceae, Bifidobacteriaceae, Romboutsia) shared between the two host species, accounting for 36%-60% of feature microbiota. Functional profiles showed significant remodeling during recovery, characterized by the loss of native functions such as D-arabinitol 4-dehydrogenase but acquisition of novel metabolic capabilities, including fructan biosynthesis pathways. Notably, Bacteroidaceae contributed extensively to differential functions, particularly short-chain fatty acid metabolism via propionyl-CoA carboxylase, highlighting its role in gut-lung immune regulation.ConclusionThese findings provide the first cross-species machine learning analysis of RSV-associated, though not necessarily RSV-specific, gut microbiota, offering insights into gut-lung axis mechanisms and identifying potential targets for probiotic-based interventions in RSV management.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12866-026-05228-9. Introduction Respiratory syncytial virus (RSV) is the most common cause of lower respiratory tract infections in children and a leading contributor to pediatric hospitalization and mortality. Notably, RSV infection accounts for approximately 1 in 50 deaths among children under 5 years of age [1]. Current therapeutic options remain severely limited, with palivizumab serving as the only reliable prophylactic agent—yet its application is restricted to high-risk subpopulations. Given the substantial socioeconomic and healthcare burdens imposed by RSV on families and society, there is an urgent need for mechanistic studies to identify novel therapeutic targets. Emerging evidence indicates that gut microbiota regulates respiratory immune responses via the “gut-lung axis”, providing a theoretical foundation for probiotic-based interventions [2]. During acute infection, host gut microbial composition undergoes significant alterations due to inflammatory responses, dietary changes, antibiotic exposure, and other factors [3]. Conversely, beneficial gut microbiota can mitigate infection through mechanisms such as promoting type I interferon secretion and CD8+ T-cell proliferation in the lungs [4–6]. Notably, probiotic supplementation in neonatal mice demonstrated protection against RSV through the microbiota-alveolar-macrophage axis [7]. However, the development of probiotics for RSV relies heavily on gut microbiota studies in humans and mice [8, 9], which are often hindered by profound interspecies differences in microbial composition [10, 11]. These differences limit the clinical translation of candidate biomarkers, underscoring the need for systematic cross-species comparisons to identify conserved microbial targets. Machine learning offers a powerful alternative to traditional linear statistical methods for microbiome research. It excels at modeling nonlinear relationships and efficiently handles the high-dimensional, non-normally distributed microbiome data [12, 13]. By capturing complex associations between microbial profiles and disease phenotypes, machine learning enables the construction of high-accuracy predictive models. Furthermore, feature extraction from these models facilitates the identification of core microbiota linked to disease pathogenesis, providing critical insights for mechanistic exploration and clinical target development. In this study, we systematically screened the SRA database and integrated five eligible projects comprising 319 samples (154 children, 165 mice; Fig. 1 and Tables S1–S2) [8, 9]. By combining microbial abundance profiling, diversity analyses, and 13 machine learning algorithms, we identified 62 and 54 key microbial taxa at the family and genus levels, respectively. These taxa, ranked as top predictive features in machine learning models for both species, represent critical candidates for mechanistic studies and probiotic development. Functional analysis identified classical and novel pathways related to the gut-lung axis, which were closely linked to the key microbial taxa. Our findings provide the first cross-species machine learning analysis of RSV-associated, though not necessarily RSV-specific, gut microbiota, paving the way for innovative interventions targeting the gut-lung axis. Fig. 1Scheme for data sources and research protocols Scheme for data sources and research protocols Materials and methods Data collection and quality control Through a search of the SRA database, we identified and collected data from five suitable projects (PRJNA669372, PRJNA627257, PRJEB21782, PRJNA579491 [Children], PRJNA978360 [Children]) [8, 9]. Samples with incomplete/missing clinicopathological information or insufficient sequencing depth were excluded. A total of 319 samples were ultimately collected, comprising 154 samples from children and 165 samples from mice (Supplementary Table S1, Supplementary Table S2). The downloaded data included both single-end and paired-end reads. These were preprocessed separately, and representative sequences were generated. The qiime demux method in QIIME2 was used to generate data quality summaries. Based on the data quality, the dada2 denoising method was employed to remove noise and generate representative sequences. Parameters for paired-end data: --p-trunc-len-f 220, --p-trunc-len-r 160, --p-trim-left-f 0, --p-trim-left-r 10; Parameters for single-end data: --p-trim-left 10, --p-trunc-len 180 [14]. Taxonomic classification and community composition analysis The merged data were rarefied (subsampled) to an even sequencing depth of 10,000 reads per sample using the rarefy command. Taxonomic annotation of the representative sequences was performed using the SILVA database (version 138.99). QIIME2 was used for subsequent analyses including visualization of microbial abundance profiles, calculation of α-diversity metrics (Shannon entropy), calculation of β-diversity metrics, and principal coordinates analysis. Statistical analysis Differences in α-diversity among groups were assessed using the Kruskal–Wallis test to evaluate global differences, followed by post-hoc pairwise comparisons conducted via Dunn’s test with Bonferroni adjustment. Significance testing for differences in β-diversity between groups was performed using Permutational Multivariate Analysis of Variance (PERMANOVA). Batch effect assessment To assess potential batch effects across projects, Principal Component Analysis (PCA) was performed using the prcomp method in R. Confounding factors were evaluated using PCA for global pattern visualization, followed by the Kruskal–Wallis test for between-subgroup comparisons. Machine learning model construction and feature selection An in-house Python script was used to randomly split the dataset samples into a training set (70%) and a validation set (30%). Thirteen classical machine learning algorithms were implemented using various R packages: randomForest (4.7–1.1) for Random Forest [15]; caTools (1.18.2) for Generalized Linear Model (GLM) [16]; xgboost (1.7.5.1) for XGBoost and set learning_rate = 0.3, max_depth = 6 to reduce the risk of overfitting while maintaining the complexity of the model and the accuracy of prediction [17]; lightgbm (3.3.5) for LightGBM [18]; rpart (4.1.19) for SuperPC [19]; catboost (1.2) for CatBoost with number of iterations set to 30 [20]. Given the moderate size of our dataset, this conservative choice was made to ensure computational efficiency without compromising model performance; NeuralNetTools (1.5.3) and neuralnet were used for Neural network, with a 30 × 30 hidden layer to provide sufficient capacity to capture complex patterns in the data while avoiding excessive computational burden [21]; glmnet (4.1-7) for Lasso, Elastic network, and Ridge [22]. Model performance was evaluated using the pROC package (v1.18.4) [23]. Receiver Operating Characteristic (ROC) analysis was performed, and the Area Under the Curve (AUC) value was calculated to assess the performance of the constructed models and the selected features. The contribution of each feature to the model’s classification performance was calculated and the top 30 features from each model were selected as key signature features. The core signature microbiota were defined as the overlapping key signature features shared between children and mice. Functional analysis Functional prediction of the microbiota was performed using PICRUSt2 combined with the KEGG, EC, and MetaCyc databases [24]. Differential functional abundance analysis between groups was conducted with the ggpicrust2 package, applying the ALDEx2 Welch’s statistical test and adjusted using the FDR method [25, 26]. Visualization of functional contributions was performed using the ggpicrust2 and ggplot2 R packages, as well as in-house scripts. Results Distinct gut microbiota structures between species After standardized preprocessing, α-diversity (Shannon entropy) analysis revealed significantly higher overall microbial diversity in mice than in children (P < 0.05, Kruskal-Wallis test). Among mice, no significant intergroup differences were observed, though controls showed slightly higher diversity than infected groups. In children, the infected group exhibited significantly lower diversity than the recovery group (P < 0.05), with infected samples consistently showing the lowest diversity (Fig. 2A; Supplementary Table S3). Fig. 2Microbiota diversity analysis of children and mice datasets. A α-diversity analysis, reflecting within-sample microbial richness and evenness; B β-diversity analysis using Bray-Curtis distances, showing the compositional differences of all samples relative to the children control group; C β-diversity analysis using Bray-Curtis distances, showing the compositional differences of all samples relative to the mouse control group; D Distribution of dominant microbiota (Top 10 genera) in children samples; E Distribution of dominant microbiota (Top 10 genera) in mouse samples Microbiota diversity analysis of children and mice datasets. A α-diversity analysis, reflecting within-sample microbial richness and evenness; B β-diversity analysis using Bray-Curtis distances, showing the compositional differences of all samples relative to the children control group; C β-diversity analysis using Bray-Curtis distances, showing the compositional differences of all samples relative to the mouse control group; D Distribution of dominant microbiota (Top 10 genera) in children samples; E Distribution of dominant microbiota (Top 10 genera) in mouse samples β-diversity analysis (Bray-Curtis distance) demonstrated significant intergroup dissimilarities (PERMANOVA, P < 0.001). In children, the mean distance between infected and control group exceeded 0.8, while the mean distance between recovery and control group was greater than 0.9 (both P < 0.001, Fig. 2B; Supplementary Table S4), suggesting persistent dysbiosis even during recovery. For mice, the mean distance between infected and control group was greater than 0.7 (P < 0.001, Fig. 2C; Supplementary Table S4), reflecting similar but less pronounced microbial shifts. Taxonomic analysis revealed distinct dominant taxa between children and mice. At the genus level, Bifidobacterium and Escherichia Shigella were predominated in children, while these taxa were scarce in mice. Conversely, Muribaculaceae, Lachnospiraceae, Lachnospiraceae_NK4A136_group, and Ligilactobacillus were dominated mouse microbiota but were rare in children (Figs. 2D–E). These findings highlight the necessity for species-specific modeling due to the marked compositional differences between human and mouse gut microbiota. Successful construction of RSV prediction models Principal Component Analysis (PCA) confirmed significant interspecies divergence (Supplementary Figure S1) but no batch effects within groups (Supplementary Figures S2–S3). We therefore constructed separate disease-prediction models using 13 machine learning algorithms at taxonomic family and genus levels, validated via independent validation sets (Fig. 1). Performance of prediction models in children data For children’s data, five models based on family-level data achieved AUC > 0.7. Among these, random_forest, xgboost, and lightgbm demonstrated the best performance, with AUC values ranging from 0.8 to 0.9 (Supplementary Figure S4). At the genus level, nine models achieved > 0.7, with catboost and random_forest performing the best, achieving AUC values of 0.952 and 0.937, respectively (Fig. 3A). Overall, prediction models constructed using children genus data exhibited better performance. Fig. 3Disease prediction modeling using genus-level microbiota data with 13 machine learning algorithms. A Area under the curve (AUC) ranking of prediction models for children data, showing the performance of 10 successful algorithms (3 algorithms failed to build valid models); B Precision of top-performing prediction models for children data (AUC > 0.8) across experimental groups; C Predictive efficacy of children-specific signature microbiota (Top 10 genus-level features); D AUC ranking of prediction models for mouse data, showing the performance of all 13 algorithms; E Predictive efficacy of mouse-specific signature microbiota (Top 10 genus-level features) Disease prediction modeling using genus-level microbiota data with 13 machine learning algorithms. A Area under the curve (AUC) ranking of prediction models for children data, showing the performance of 10 successful algorithms (3 algorithms failed to build valid models); B Precision of top-performing prediction models for children data (AUC > 0.8) across experimental groups; C Predictive efficacy of children-specific signature microbiota (Top 10 genus-level features); D AUC ranking of prediction models for mouse data, showing the performance of all 13 algorithms; E Predictive efficacy of mouse-specific signature microbiota (Top 10 genus-level features) The children data included three groups: control, infected, and recovery. The recovery group consisted of fecal specimens collected 15 to 150 days post-discharge [9]. Most models showed strong predictive power for the infected group, but poor accuracy for the control group (Fig. 3B). We extracted the top 30 important features from each prediction model and tested the predictive ability of single features. The results showed that the predictive power of single features was significantly lower than that of the overall prediction model. Nevertheless, seven bacteria still had AUC values higher than 0.7, indicating they still possess certain predictive capabilities (Fig. 3C). Performance of prediction models in mouse data For mouse data, seven models based on family-level data had AUC values exceeding 0.7, with three models (lasso, random_forest, and adaboost) having AUC values exceeding 0.8 (Supplementary Figure S5). At the genus level, eight models had AUC values exceeding 0.8, including the three tools that performed well at the family level (Fig. 3D). Similar to the children data, genus-level models outperformed family-level models. The mouse data were divided into two groups: control and infected. We also tested the individual classification ability of important microbiota. The results showed that the classification ability of the top 10 important microbial communities was all above 0.6, but none of them exceeded 0.7 (Fig. 3E). This result indicates that prediction in mouse data relies more on the combined effect of microbial communities rather than individual taxa. Compared with mouse data, models performed better on data from children. When the decomposed overall performance on the pediatric dataset into three binary classification tasks, infected vs. control, infected vs. recovery, and recovery vs. controls, the models performed comparably on the infected vs. control task for both children and mouse data at both family and genus levels (Supplementary Figure S6, S7, Supplementary Table S5). The predictive models consistently yielded the highest performance for the children infected vs. recovery classification, regardless of taxonomic level (family or genus). In particular, at the genus level, performance on this task significantly exceeded that of the infected vs. control comparison in mouse data. Integrated analysis identifies 62 core signature microbiota (family level) shared between children and mouse Given the substantial compositional differences in gut microbiota between children and mice, we performed an integrated analysis at family and genus levels to investigate whether RSV infection induces similar microbial changes, focusing on significant signature microbiota defined by their predictive importance in machine learning models. At family level, 62 significant microbial taxa were shared between children and mice, representing 58% of the significant taxa in children and 60% in mice (Fig. 4A, Supplementary Table S6). Notably, the top 5 microbiota ranked by AUC value in children were also significant in mice. Fig. 4Analysis of shared family-level signature microbiota between children and mice. A Venn diagram showing the overlap of genus-level signature microbiota identified from infected versus control group comparisons in children and mice, respectively; B Predictive performance (AUC values) for distinguishing children experimental groups based on the relative abundance of Bacteroidaceae alone; C AUC values for predicting children experimental groups based on the relative abundance of Bifidobacteriaceae alone; D Relative abundance distribution of Bacteroidaceae and Bifidobacteriaceae in children samples; E Relative abundance distribution of Bacteroidaceae and Bifidobacteriaceae in mouse samples Analysis of shared family-level signature microbiota between children and mice. A Venn diagram showing the overlap of genus-level signature microbiota identified from infected versus control group comparisons in children and mice, respectively; B Predictive performance (AUC values) for distinguishing children experimental groups based on the relative abundance of Bacteroidaceae alone; C AUC values for predicting children experimental groups based on the relative abundance of Bifidobacteriaceae alone; D Relative abundance distribution of Bacteroidaceae and Bifidobacteriaceae in children samples; E Relative abundance distribution of Bacteroidaceae and Bifidobacteriaceae in mouse samples Bacteroidaceae exhibited the highest AUC value in children (AUC = 0.88) and most effectively distinguished the recovery group from the control group (Fig. 4B). In children, Bacteroidaceae showed relatively high abundance in most control samples, variable abundance in some infected samples, and was nearly absent in the recovery group (Fig. 4D). In mouse, Bacteroidaceae was present in some control samples with high abundance but showed significantly lower abundance in the infected group compared to controls (Fig. 4E). Bifidobacteriaceae ranked second in children (AUC = 0.88) and similarly demonstrated high efficiency in distinguishing the recovery and control groups (Fig. 4C). Its distribution pattern was opposite to that of Bacteroidaceae: Bifidobacteriaceae was more abundant and prevalent in the recovery group, less abundant and prevalent in the control group, and showed intermediate levels in the infected group (Fig. 4D). However, Bifidobacteriaceae was nearly absent in mouse samples (Fig. 4E). Integrated analysis identifies 54 core signature microbiota (genus level) shared between children and mouse At genus level, 54 significant microbial taxa were shared between children and mouse, accounting for 36% of significant taxa in children and 40% in mice (Fig. 5A, Supplementary Table S7). Among the top 5 microbiota by AUC value in children, only two genera Romboutsia and Bacteroides were also significant in mice. Fig. 5Analysis of shared genus-level signature microbiota between children and mice. A Venn diagram showing the overlap of genus-level signature microbiota identified from infected versus control group comparisons in children and mice, respectively; B AUC values for predicting children groups based on the relative abundance of Romboutsia alone; C AUC values for predicting children groups based on the relative abundance of Bacteroides alone; D Relative abundance distribution of Romboutsia and Bacteroides in children samples; E Relative abundance distribution of Romboutsia and Bacteroides in mouse samples Analysis of shared genus-level signature microbiota between children and mice. A Venn diagram showing the overlap of genus-level signature microbiota identified from infected versus control group comparisons in children and mice, respectively; B AUC values for predicting children groups based on the relative abundance of Romboutsia alone; C AUC values for predicting children groups based on the relative abundance of Bacteroides alone; D Relative abundance distribution of Romboutsia and Bacteroides in children samples; E Relative abundance distribution of Romboutsia and Bacteroides in mouse samples Romboutsia ranked third in AUC value among children. It effectively discriminated recovery samples but poorly distinguished control and infected samples (Fig. 5B). Romboutsia was almost exclusively detected in the recovery group of children (Fig. 5D). Interestingly, in mouse, Romboutsia appeared in some control samples but was extremely rare in the infected group (Fig. 5E). Notably, Romboutsia belongs to the family Peptostreptococcaceae, which had an overall low AUC value (0.589), indicating limited individual predictive power. Bacteroides ranked fourth in children. As the sole genus identified within the family Bacteroidaceae in this study, its predictive efficiency closely mirrored that of Bacteroidaceae (Fig. 5C, D and E). Functional heterogeneity revealed by multi-database analysis Compared to 16S rRNA sequencing, metagenomic sequencing enables the analysis of microbial community functions by providing access to the entire gene repertoire of the microbiota. Leveraging databases of species-function relationships established through extensive metagenomic sequencing analyses, we can infer the functions of target microbial communities based on the taxonomic profiles obtained from 16S rRNA sequencing, by matching against these databases and considering phylogenetic relationships [24]. We predicted the functions of the microbiota by using PICRUSt2 with three databases: EC, KEGG, and MetaCyc. At the genus level, we identified differentially abundant functions and pathways across three comparison groups: children infected vs. control (CI_VS_CC), children recovery vs. control (CR_VS_CC), and mice infected vs. control (MI_VS_MC). Based on differences in the abundances of predicted functions/pathways across groups, we identified a total of 1835, 265, and 373 differential functions/pathways in the EC, KEGG, and MetaCyc databases, respectively (Fig. 6A). Among the three databases, the EC database yielded the highest number of differentially abundant functions, with 647, 933, and 255 identified in the child infected group vs. child control group (CI_VS_CC), child recovery vs. child control (CR_VS_CC), and mice infected vs. mice control (MI_VS_MC) comparisons, respectively. The KEGG database yielded the fewest differential pathways, with 118, 122, and 25 identified in the aforementioned three comparisons. Fig. 6Cross-species comparative functional analysis of gut microbiota following RSV infection and recovery (genus level). A Counts of significantly differentially abundant genes/pathways identified from three comparisons: Child Infected vs. Child Control (CI_VS_CC), Child Recovery vs. Child Control (CR_VS_CC), and Mouse Infected vs. Mouse Control; B Venn plot for top 10 significantly differentially abundant ECs identified from the three comparisons; C Top 10 significantly differentially abundant ECs identified in the comparison of Child Recovery group versus Child Control group. D Contributions of Bacteroides to the top 10 significantly differentially abundant ECs for each of the three comparisons, where contributions are calculated as fractions of contributed samples with ggpicrust predicted contribution value exceeding 0.05. Columns with all zero values were omitted from the display; E Venn plot for top 10 significantly differentially abundant KEGG pathways identified from the three comparisons; F Contributions of Bacteroides to the top 10 significantly differentially abundant ECs for each of the three comparisons; G Venn plot for top 10 significantly differentially abundant MetaCyc pathways identified from the three comparisons; H Top 10 significantly differentially abundant MetaCyc pathways identified in the comparison of Child Infected group versus Child Recover group; I The Fraction of samples containing taxa that carry genes associated with MetaCyc PWY-822 (fructan biosynthesis) Cross-species comparative functional analysis of gut microbiota following RSV infection and recovery (genus level). A Counts of significantly differentially abundant genes/pathways identified from three comparisons: Child Infected vs. Child Control (CI_VS_CC), Child Recovery vs. Child Control (CR_VS_CC), and Mouse Infected vs. Mouse Control; B Venn plot for top 10 significantly differentially abundant ECs identified from the three comparisons; C Top 10 significantly differentially abundant ECs identified in the comparison of Child Recovery group versus Child Control group. D Contributions of Bacteroides to the top 10 significantly differentially abundant ECs for each of the three comparisons, where contributions are calculated as fractions of contributed samples with ggpicrust predicted contribution value exceeding 0.05. Columns with all zero values were omitted from the display; E Venn plot for top 10 significantly differentially abundant KEGG pathways identified from the three comparisons; F Contributions of Bacteroides to the top 10 significantly differentially abundant ECs for each of the three comparisons; G Venn plot for top 10 significantly differentially abundant MetaCyc pathways identified from the three comparisons; H Top 10 significantly differentially abundant MetaCyc pathways identified in the comparison of Child Infected group versus Child Recover group; I The Fraction of samples containing taxa that carry genes associated with MetaCyc PWY-822 (fructan biosynthesis) The EC database reveals extensive and shared differential enzymatic functions in pediatric groups Among all differentially abundant ECs, the CI_VS_CC and CR_VS_CC comparisons shared 549 differential enzymes, whereas the MI_VS_MC comparison shared only 134 differentially abundant enzymes with the two children-related comparisons (Supplementary Figure S8). Among the top 10 EC functions ranked by adjusted P value in each comparison group, two functions were common between the CI_VS_CC and CR_VS_CC comparisons, while no functions were shared between the mouse and children comparisons (Fig. 6B). These two shared functions were D-arabinitol 4-dehydrogenase (1.1.1.11) and L-threonate 2-dehydrogenase (1.1.1.411). The Escherichia-Shigella genus was the most prominent contributor to these two functions, detected in over 90% of children in the control group and over 50% in the infected children group (Supplementary Figure S9, S10). Overall, the abundances of these two functions were suppressed in the child infected group and were entirely absent in the child recovery group (Fig. 6C, Supplementary Figure S11). Notably, the gut microbiota of the recovery group exhibited a substantial number of new functions that were absent in the control group for children, such as Lycopene beta-cyclase (5.5.1.19), Proton-exporting ATPase (3.6.3.6), 4-oxalomesaconate tautomerase (4.2.1.83), Zeaxanthin glucosyltransferase (5.3.2.8), Glucose 1-dehydrogenase (FAD, quinone) (1.1.5.9), 4-oxalmesaconate hydratase (4.2.1.83), and Levansucrase (2.4.1.10) (Fig. 6C). Romboutsia and Bacteroides are two genera with strong discriminatory power between groups. Among them, Romboutsia exhibits a relatively low contribution to the aforementioned differentially abundant ECs, while Bacteroides shows a relatively higher contribution to the differential ECs (Fig. 6D, Supplementary Figure S12). However, Bacteroides shows no contribution to the differential ECs shared between the CI_VS_CC and CR_VS_CC comparisons. Among the 8 unique differential functions of CR_VS_CC, Bacteroides only contributes to Lycopene beta-cyclase, with an overall low contribution degree. Bacteroides contributes to several differential functions unique to CI_VS_CC and MI_VS_MR (4 and 5 functions, respectively). Furthermore, in the child control and child infected groups, it is extensively involved in functions including Propionyl-CoA carboxylase (6.4.1.3), beta-1,2-mannosidase (3.2.1.197), Formyltetrahydrofolate deformylase (3.5.1.10), 4-hydroxy-4-methyl-2-oxoglutarate aldolase (4.1.2.27), Alpha-glucuronidase (3.2.1.139), and Aspartate 4-decarboxylase (4.1.1.12) (Fig. 6D). KEGG pathway analysis highlights core metabolic perturbations Among all differentially abundant KEGG pathways, the CI_VS_CC and CR_VS_CC comparisons shared 91 differential pathways, whereas the MI_VS_MC comparison shared only 15 differential pathways with both children comparisons (Supplementary Figure S13). Among the top 10 differential KEGG pathways, 5 pathways are found significantly differential in both CI_VS_CC and CR_VS_CC groups. Three of these pathways are related to metabolism, namely Peptidoglycan biosynthesis (ko00550), Glycolysis / Gluconeogenesis (ko00010), and Chloroalkane and chloroalkene degradation (ko00625). The others are Alzheimer disease (ko05010) and Phosphatidylinositol signaling system (ko04070) (Fig. 6E). In these pathways, the microbial abundance in the child infected and child recovery group is higher than that in the child control group (Supplementary Figure S14, S15). Since each KEGG pathway involves a large number of genes and enzymes, the number of microbial taxa contributing to the aforementioned pathways exceeded 100 species each (Supplementary Figures S16-S20). Among the identified core signature microbiota, both Romboutsia and Bacteroides contributed to the differentially abundant KEGG pathways. Overall, Romboutsia showed a low contribution to differential functions, with a relatively higher contribution observed in the child recover group to pathways Glycerophospholipid metabolism (ko00564) and Aminobenzoate degradation (ko00627) (Supplementary Figure S21). Bacteroides is extensively involved in the differential pathways, where it shows a higher abundance in child control group to pathways such as DNA replication (ko03030), Meiosis - yeast (ko04113), Apoptosis (ko04210), Steroid hormone biosynthesis (ko00140), and Aminobenzoate degradation (ko00627) (Fig. 6F). MetaCyc analysis identifies fructan biosynthesis as a key pathway in post-infection recovery Among all differentially abundant MetaCyc pathways, the CI_VS_CC and CR_VS_CC comparisons shared 119 differential pathways, whereas the MI_VS_MC comparison shared only 22 differential pathways with both children comparisons (Supplementary Figure S22). In the MetaCyc analysis, only one pathway, fructan biosynthesis (PWY-822), was common to the top 10 significant results from both the CI_VS_CC and CR_VS_CC comparisons (Fig. 6G). The abundance of fructan biosynthesis was very low in the child control group, while it was relatively higher in both the child infected and child recovery groups (Fig. 6H, Supplementary Figure S23). Additionally, two other pathways—gallate degradation II (GALLATE-DEGRADATION-I-PWY) and gallate degradation I (GALLATE-DEGRADATION-II-PWY)—were identified from the CR_VS_CC comparison, and they exhibited a similar abundance pattern. Streptococcus was the genus with the highest contribution to fructan biosynthesis, detected in over 25% of infected group samples and over 90% of recover group samples (Fig. 6I). Among the identified core signature microbiota, Romboutsia showed low contribution to the aforementioned differentially abundant MetaCyc pathways, while Bacteroides showed a relatively higher contribution to the differential pathways (Supplementary Figure S24, S25). Bacteroides showed the highest contribution to 2-oxobutanoate degradation I (PWY-5130) in the child control group, with over 80% of samples contributing to this function. Categorizing the differential pathways by their source of significance, the pathways associated with Bacteroides were predominantly from the CI_VS_CC comparison (6 out of 8). Across all 8 pathways, Bacteroides exhibits the highest contribution in the child control group, while its contribution is lower in the child infected group and almost absent in the child recovery group (Supplementary Figure S25). Discussion This study systematically analyzed the gut microbiota dynamics during RSV infection in children and mice, revealing significant interspecies differences and shared microbial signatures. Our findings provide insights into how RSV infection disrupts gut microbial structure and highlight the potential of microbiota-based biomarkers and therapeutic interventions. By integrating five public datasets comprising 319 samples (154 from children, 165 from mice), we observed that the overall α-diversity of gut microbiota in mice was higher than in children, consistent with prior reports of greater microbial diversity in rodents compared to humans [27]. In both children and mice, the RSV infected group exhibited lower α-diversity than the corresponding control group, though the difference was not statistically significant. This suggests that RSV infection partially disrupts the gut microbiota structure in both hosts. Notably, the recovery group in children showed significantly higher α-diversity than the infected group, indicating partial restoration of microbial structure during convalescence. β-diversity analysis (Bray-Curtis distance) further confirmed significant structural differences in microbiota between the RSV infected and control groups in children (mean distance > 0.8). The distance between the recovery and control groups expanded further (> 0.9; PERMANOVA, P < 0.001). This suggests that while microbial richness may recover over time, the overall community structure remains disrupted or reprogrammed to an alternative stable state, potentially contributing to long-term health consequences such as increased susceptibility to respiratory diseases such as asthma [28]. In mice, although the distance between infected and control groups was smaller (< 0.7), it remained statistically significant (P < 0.001), indicating similar but less pronounced microbial shifts. At the taxonomic level, Bifidobacteriaceae and Escherichia-Shigella were identified as key microbial signatures dominating in children (collectively accounting for approximately 35% of the total abundance), while Lachnospiraceae and Ligilactobacillus were enriched in mice (accounting for over 40%). Despite these structural disparities, based on the principles of functional redundancy and the core metagenome, we posit that the core gut microbiota performs conserved functions across hosts [29, 30]. Therefore, to reconcile compositional divergence with functional similarity, we adopted a strategy of initially developing models separately for each species, followed by cross-species validation of the identified microbial signatures. Compared to traditional methods, machine learning techniques enable more in-depth analysis and mining of omics data and demonstrate immense potential for uncovering latent relationships within high-dimensional datasets. However, different machine learning algorithms possess distinct characteristics, with varying strengths and limitations that define their applicable scopes. When dealing with complex biomedical omics data, a single machine learning algorithm often falls short in fully capturing the intricate and multifaceted features of such datasets. To address this challenge, we employed 13 machine learning models representing mainstream algorithms to construct RSV infection prediction models for children and mice, respectively. Through model evaluation, feature microbiota analysis, and cross-species validation, we identified a set of relatively reliable signature microbial taxa. Among all prediction models of children, 5 out of 10 successfully built models based on family data and 9 out of 10 models based on genus data achieved AUC > 0.7. For mice 7 out of 13 family-level data and 8 out of 13 genus-level models exceeded AUC 0.7. Overall, models based on genus data outperformed models based on family data in both children and mice, underscoring the importance of higher taxonomic resolution for biomarker discovery. Additionally, predictive models showed their strongest performance in distinguishing children infected and recovery data, likely due to larger intergroup microbiota differences in children recovery group, as evidenced by both α-diversity and β-diversity analysis. Despite substantial compositional differences between mouse and human gut microbiota, shared microbial responses to RSV infection may reveal core mechanisms. We identified “shared signature microbiota” by intersecting the taxa with high predictive importance from both species. Surprisingly, at family level, the shared taxa accounted for 58% of children’s and 60% of mice’s core microbiota; at genus level, these proportions dropped to 36% and 40%, respectively. Among all shared microbiota, Bacteroides (the sole genus identified within Bacteroidaceae) is a dominant commensal in healthy human guts [2]. Here, Bacteroides/Bacteroidaceae abundance alone strongly discriminated controls from the recovery group. It was highly abundant in children’s controls, variably present in the RSV infected group, nearly absent in the recovery group, and significantly reduced in infected mice group compared to controls. This pattern suggests a protective role. Notably, the changes in Bacteroidaceae abundance following infection displayed distinct pathogen-specific dynamics: its abundance was decreased in community-acquired pneumonia, but increased in patients with COVID-19 [31, 32]. Intriguingly, mouse studies show Bacteroides thetaiotaomicron protects against Candida albicans colonization via immune activation (HIF1α/LL-37) [33]. Its depletion in post-RSV recovery may weaken immune protection, potentially increasing susceptibility to long-term asthma [28]. Bifidobacteriaceae, identified as another key shared taxon, appears to play an important role in children. However, it cannot serve as a reliable biomarker due to its low abundance in mice. Its levels increased post-infection and were significantly elevated during the recovery phase in children. Interestingly, this pattern appears to be virus-specific: while its abundance increased in children with community-acquired pneumonia, it decreased following influenza A virus infection, underscoring the pathogen-specific dynamics of microbial community responses [31, 34]. Romboutsia, a low-abundance genus in humans, was specifically enriched in the recovery group of children and detected in some mice controls but rare in infected groups—hinting at cross-species protective potential. Meta-analyses have linked reduced Romboutsia abundance to childhood depression and cardiovascular diseases [35–37], further supporting its protective capacity. The abundance of Romboutsia was detected reduced in both COVID-19 and influenza A infection, which also indicates the virus-specific dynamics [38]. The signature microbiota identified here represent valuable resources for developing anti-RSV probiotic therapies. For instance, Bifidobacteriaceae, a well-commercialized and safe probiotic, exhibits significant abundance changes both in this study and during influenza infections. Future work should focus on designing clinical trials to elucidate its specific role in combating viral infections. Based on the EC, KEGG, and MetaCyc databases, we performed functional prediction and differential abundance analysis of the identified bacteria. The results of Venn analysis showed that among the three sets of comparisons—child infected group vs. child control group, child recovery group vs. child control group, and mouse infected group vs. mouse control group—the first two comparisons exhibited more consistent “differential functions/pathways,” while the “differential functions/pathways” of the latter had less overlap with the former two. In contrast to the high consistency of “shared core signature microbiota” between humans and mice, the microbiota functions/pathways showed high heterogeneity between RSV-infected children and the RSV-infected mouse model. We speculate that this phenomenon can be explained by the following two reasons: First, function is an outcome of microbial community changes. A single bacterial taxon can correspond to hundreds of functions; therefore, changes in microbiota will exhibit an amplified effect in terms of function. Second, during the identification of differential functions, we used conventional statistical methods to calculate P-values based on the differences in the abundance of microbiota associated with a specific function between groups. In contrast, the identification of the “core” microbiota involved multiple machine learning models. These non-linear methods may better capture the essence of changes in certain cases. Among the identified differential functions/pathways, we noted that the core microbiota Bacteroides was involved in multiple differential functions. Specifically, it exhibited the highest contribution to Propionyl-CoA carboxylase (EC: 6.4.1.3) in the child control group, while its contributions were relatively lower in the child infected and child recover groups. This function is directly involved in propionate metabolism. Propionic acid is a type of short-chain fatty acid (SCFAs), and numerous studies have demonstrated that SCFAs can regulate pulmonary immune function through the gut-lung axis, thereby participating in the immune processes of infections including influenza and RSV [39–41]. Beta diversity analysis revealed that the intestinal microbiota during the recovery process still showed significant differences from that of the control group. Functionally, this was reflected in the loss of certain functions present in the control group, such as EC D-arabinitol 4-dehydrogenase (EC: 1.1.1.11) and L-threonate 2-dehydrogenase (EC: 1.1.1.411); along with the emergence of numerous new functions, including Lycopene beta-cyclase (EC: 5.5.1.19), Proton-exporting ATPase (EC: 3.6.3.6), 4-oxalomesaconate tautomerase (EC: 5.3.2.8), Zeaxanthin glucosyltransferase (EC: 5.3.2.8), Glucose 1-dehydrogenase (FAD, quinone) (EC: 1.1.5.9), 4-oxalmesaconate hydratase (EC: 4.2.1.83), and Levansucrase (EC: 2.4.1.10). Among these, Levansucrase (EC 2.4.1.10) uses sucrose as a substrate, catalyzing its decomposition into glucose and elongating fructan chains. Interestingly, the differential functions identified in both KEGG and MetaCyc database, Glycolysis/Gluconeogenesis (ko00010) and fructan biosynthesis (PWY-822), are associated with fructan enzymes, and both pathways were highly expressed in the recovery group. As reported, fructo-oligosaccharides (FOS), a type of prebiotic, can promote the proliferation of Bifidobacterium [42]. It indicated that the activation of the fructo-oligosaccharide pathway is associated with the higher proportion of Bifidobacterium in the recovery group. In addition to infection, the structure of the gut microbiota in children is influenced by a variety of factors. These factors include technical biases, such as sequencing platform and DNA extraction methods; population differences, such as age, disease severity, duration of hospitalization/post-discharge time, dietary habits, and sex; and experimental/drug interventions, such as the use of antibiotics or probiotics. Although we collected some of this metadata, due to the limited sample size, it was difficult for this integrative study to perform stratified analyses or machine learning-based multi-class modeling to account for these confounding factors (Supplementary Table S2). To address this limitation, we used PCA to globally assess the impact of these factors and further compared the abundance of core microbiota in subgroups. All five included projects were sequenced on the Illumina MiSeq platform, but each project used a different DNA extraction method. The effect of DNA extraction methods was confounded with the project batch effect and could not be independently assessed (Supplementary Table S2). Some studies recorded information on disease severity, age, and time post-discharge in children, but these variables showed no significant effect on the identified core microbiota (Supplementary Figures S26–S34). This may be attributed to the small number of relevant samples. Only 5 children were included in the severe group, and only 9 cases were collected at 10 days after discharge. The collected mouse samples included three interventions: Ovalbumin, PBS, and Streptomycin. Among these, the Streptomycin-treated group showed a clear separation from the untreated group in the PCA. However, possibly due to its small sample size (n = 8), there was no significant difference in the abundance of core microbiota between the Streptomycin-treated and untreated groups (Supplementary Figures S35–S37). The Ovalbumin and PBS treatment groups did not show significant separation from the untreated group in the PCA. Furthermore, there were no differences in the abundance of core microbiota (at the genus level) between the Ovalbumin/PBS treatment groups and the untreated group. Notably, mouse models are relatively easy to manipulate and ethically feasible, so they often involve relatively strong interventions. These interventions are further intertwined with laboratory-specific protocols, experimental objectives, and study designs, making their effects difficult to fully eliminate during multi-source data integration. Although statistical analyses suggest that the three treatments in mice have a limited impact on the core microbiota, their influence cannot be completely ruled out. Therefore, these factors should still be adequately considered in future studies. Despite its contributions, this study has limitations. The functional analysis was based on predictions from 16S rRNA data, which are not as reliable as those obtained directly from metagenomic or metatranscriptomic data. The relatively small sample size and the inclusion of only publicly available datasets may limit the generalizability of our findings. Additionally, critical variables such as RSV infection severity, host age, and environmental factors were not fully controlled for. Particularly regarding the recovery group (samples collected 15–150 days post-discharge), the small sample size restricted our ability to conduct further subgroup stratification and subsequent discussions. From a biological perspective, the persistent differences observed in the recovery group might reflect a “long-term microbiota reprogramming” induced by early-life RSV infection. However, interpreting these findings is complicated by the fact that children under 5 years of age—the primary demographic for severe RSV—undergo drastic natural microbiota restructuring due to evolving dietary habits. Thus, the observed β-diversity differences may represent a superimposed effect of both viral disturbance and normal physiological maturation, and disentangling these factors will require carefully age-matched control cohorts in future studies. Given the lack of an internal control group with other acute respiratory viral infections (e.g., SARS-CoV-2, influenza) in the present cohort, the observed gut microbiota changes may reflect general hallmarks of acute infection rather than RSV-specific signatures. For example, although published evidence implies that fluctuations in Bacteroidaceae and Bifidobacteriaceae abundance may follow virus-specific trends, the datasets incorporated herein and previous reports are not rigorously comparable, and confounding variables such as age and dietary status cannot be entirely ruled out. Thus, the RSV specificity of these microbial candidates warrants further verification and confirmation. Future studies should aim to expand sample sizes, incorporate direct controls for other acute respiratory viral infections, include more diverse cohorts and well-matched cohorts (e.g., age-matched control groups), record a more comprehensive set of potential confounding (e.g., dietary) and clinical variables, incorporate more types of sequencing data (e.g., metagenomics and metatranscriptomics), and conduct longitudinal analyses to better understand the temporal dynamics of RSV-induced microbiota changes. Methodologically, state-of-the-art machine learning frameworks and ensemble learning models can be adopted in future work to construct clinically translatable predictive models [43, 44]. Furthermore, clinical trials are needed to validate the therapeutic potential of identified taxa, such as Bifidobacteriaceae, in mitigating RSV-associated dysbiosis and improving recovery outcomes. Conclusion This study provides a detailed analysis of gut microbiota dynamics during RSV infection in children and mice, uncovering significant interspecies differences and shared microbial signatures. RSV infection was found to disrupt gut microbial α-diversity in both children and mice, with partial restoration of microbial richness during recovery in children, though β-diversity analysis indicated persistent dysbiosis. Machine learning-based disease prediction models demonstrated high accuracy, particularly with genus-level data, highlighting the importance of high-resolution taxonomic analysis for identifying microbial biomarkers. Cross-species analysis identified 62 family-level and 54 genus-level shared taxa with high predictive importance, including key microbial signatures such as Bacteroidaceae and Bifidobacteriaceae, which hold promise as targets for microbiota-based therapeutic interventions. The functional profiles of the gut microbiota exhibited marked heterogeneity between children and mouse models and underwent remodeling during the recovery phase. Specifically, the recovery-phase microbiota lost certain native functions, such as D-arabinitol 4-dehydrogenase, while acquiring a range of novel functions, including multiple enzymes and pathways related to fructan metabolism. Among the core microbiota, Bacteroidaceae was involved in numerous differential functions, including the metabolism of short-chain fatty acids (SCFAs)—a classic gut-derived mediator known to regulate respiratory immunity. Future research should focus on expanding sample sizes and incorporating more diverse cohorts, including controls for other viral infections to delineate RSV-specific alterations, thereby enhancing the generalizability of findings. Longitudinal studies are needed to evaluate the long-term effects of RSV on gut microbiota and its implications for respiratory and overall health. Additionally, clinical trials should be designed to assess the efficacy of probiotics, such as Bifidobacteriaceae, in mitigating RSV-induced dysbiosis, promoting recovery, and potentially preventing long-term complications. Supplementary Information Supplementary Material 1. Supplementary Material 1. Supplementary Material 2. Supplementary Material 2.