Bifidobacterium animalis reshapes the bile acid pool and prevents neonatal jaundice: a clinical microbiome study from correlation to causation

Abstract

Neonatal jaundice (NJ) affects 60–80% of neonates, yet the underlying microbial mechanisms remain elucidated, despite known links between gut dysbiosis and bilirubin and bile acid (BA) metabolism. Through two-stage shotgun metagenomic-metabolomic analysis of 150 fecal samples from 120 neonates, we identified key taxa linked to bile acid (BA) metabolism in moderate-to-severe NJ. Furthermore, multi-omics integration revealed significant interkingdom correlations among gut phages, bacteria, and BAs. Dysbiosis featured enriched Streptococcus and Escherichia, depleted Bifidobacterium animalis, and group-specific phage signatures. In the independent clinical validation cohort, jaundice intervention normalized the dysbiotic profile, demonstrating significant suppression of pathogenic taxa concomitant with restoration of B. animalis abundance. In vitro, B. animalis subsp. lactis Y103-OTU5 remodeled BA via deconjugation. In a phenylhydrazine hydrochloride (PHZ)-induced murine model of hemolytic jaundice, oral administration of isolated B. animalis subsp. lactis Y103-OTU5 significantly attenuated hyperbilirubinemia and hepatic inflammation, likely via Cyp7a1/Cyp7b1-dependent modulation of BA synthesis and detoxification pathways. Structural equation modeling revealed a tripartite regulatory network: phages indirectly modulated BA through bacterial remodeling, while B. animalis directly regulated BA pathways, positioning it as a potential therapeutic candidate for hemolysis-associated neonatal jaundice. Collectively, these findings reveal a gut phage-bacteria-BA network in NJ, highlighting B. animalis as a therapeutic candidate with dual modulation of BA metabolism and phage-bacteria interactions.

Introduction

Neonatal jaundice (NJ), also known as neonatal hyperbilirubinemia (NH), is a prevalent clinical disorder in newborns. It has been observed that approximately 60% of term infants and 80% of preterm infants exhibit jaundice within the initial week of life¹. Although the majority of cases are of a physiological nature and self-limiting, pathologic jaundice that fails to be recognized and intervened in a timely manner, especially when it progresses to severe jaundice (usually defined as a significant elevation of bilirubin levels that approaches or exceeds the threshold for exchange transfusion), carries a dramatically increased risk. Severe jaundice can lead to acute bilirubin encephalopathy and chronic kernicterus, resulting in irreversible neurotoxic damage including hearing loss, motor deficits (e.g., bradykinesia), eye movement abnormalities, and severe cognitive delays, and even death, which poses a significant burden on the child, his or her family, and the world’s health^2–5. In addition to environmental and clinical factors, genetic variations have also been identified as important risk factors for NH⁶, highlighting the multifactorial nature of NJ pathogenesis. Notably, NJ has been implicated as a potential risk factor for autism spectrum disorder (ASD). In our previous work using machine learning-based approaches, we identified a potential association between NJ and ASD, which appears to involve synergistic alterations in gut microbiome features—including viruses, bile acids (BA), and Group B Streptococcus (GBS)⁷. Moreover, further analysis demonstrated that GBS-related gene signatures alone can effectively distinguish ASD individuals from controls⁸. Current standard treatments for reducing bilirubin levels and preventing neurological damage include phototherapy and exchange transfusion^9,10, the latter being a life-saving intervention in cases of severe jaundice. However, some cases of severe jaundice respond poorly to conventional phototherapy or even exchange transfusion, and their underlying etiology remains unclear⁵, suggesting the involvement of more complex and as-yet-undefined pathogenic mechanisms.

The gut microbiota plays a pivotal role in bilirubin metabolism^11,12. In recent years, the role of the gut microbiome as a metabolic regulator has been revealed, and a large number of studies have confirmed its close relationship with NJ^12,13. For instance, Bacteroides and Clostridium have been observed to facilitate bilirubin reabsorption through the action of β-glucuronidase¹⁴. Clinical studies have also found that moderate supplementation with probiotics such as Bifidobacterium may indirectly alleviate jaundice by competitively inhibiting pathogenic bacterial colonization or modulating BA metabolism^15,16. Infection is a significant etiologic factor in NJ. GBS colonization is a significant etiological agent of neonatal infection¹⁷. This condition is characterized by its ability to transmit from mother to offspring during the birth process, thereby disrupting the process of intestinal flora colonization in the neonate. The resulting imbalance in the intestinal microbiota can be evidenced by a decrease in beneficial bacterial species, such as Lactobacillus and Bifidobacterium and a relative increase in the proportion of certain conditionally pathogenic bacteria (e.g., Streptococcus spp. and Enterococcus spp.)^18,19, which may affect bilirubin metabolism and the development of NJ. Consequently, it is imperative to closely monitor infants born to mothers with GBS infection during pregnancy for the early recognition and treatment of jaundice.

BA have been identified as pivotal bridging and signaling molecules in the microbial regulation of bilirubin metabolism. BA has been recognized as a significant endocrine signaling molecule, playing a crucial role in the regulation of glucose, lipid, and energy homeostasis, as well as inflammation²⁰. Growing evidence highlights the intricate interplay between gut microbiota dysbiosis and BA metabolic disorders in the pathogenesis of various diseases. One study demonstrated that Lactobacillus acidophilus modulates BA homeostasis to ameliorate cholestatic liver injury²¹. Matthias Laudes et al. found type 2 diabetes associated with gut microbe-related BA metabolism²². Gut microbiota regulates platelet activation and atherothrombosis through BA signaling²³, and microbiome-derived BA enhance antitumor immunity²⁴. Of particular significance is the direct modulatory effect of BA on bilirubin metabolism. For instance, ursodeoxycholic acid (UDCA), a secondary bile acid, has been shown to inhibit bilirubin production via the farnesol X receptor (FXR)²⁵.

Our previous study provided direct evidence for the “gut microbiota-BA-NJ” pathway, it indicated that BA has the potential to serve as a marker for NJ, and that gut bacteria is associated with BA abnormalities²⁶. Furthermore, we discovered that gut bacteria affect serum total bilirubin (TBIL) levels by influencing BA metabolism through a machine learning-causal inference approach²⁶. This suggests that BA may act directly on the NJ process by affecting TBIL levels. Disruption of BA homeostasis may exacerbate the condition by altering the enterohepatic circulation of bilirubin. The extant literature has been predominantly focused on the gut bacteria, thereby overlooking the potential regulatory function of the gut virome, particularly phage. Phages may regulate bacterial function by lysis of host bacteria or horizontal gene transfer, which in turn affects metabolic pathways²⁷. The complex interactions between gut bacteria, phages and BA metabolic pathways in moderate-to-severe NJ are unknown. And although it is confirmed that the bacteria can influence bilirubin via BA, exactly which key bacterial species or functional modules are responsible for regulating BA metabolism, which in turn determines the severity of jaundice, also needs to be further investigated.

To address these knowledge gaps, a multi-omics was performed on neonates with NJ, combining metagenomic sequencing, targeted BA metabolomics, and culturomics. The objective of this study was to characterize the structure of gut bacterial and viral communities associated with moderate-to-severe NJ. In addition, the study sought to identify key bile acid metabolites associated with moderate-to-severe NJ. Finally, the study explored which key gut microorganisms mediate BA metabolism in order to influence the severity of NJ. This study offers a comprehensive understanding of the multidimensional interactions of NJ, which may lead to the discovery of new biomarkers and therapeutic targets for the early intervention of the disease. By elucidating these interactions, our findings may offer novel insights into the pathogenesis of NJ and facilitate the development of targeted therapeutic strategies, such as modulation of the microbiota or BA supplementation, to mitigate the risk and severity of NJ.

Results

This study integrated clinical research, bacterial isolation/cultivation, and animal experiments (Fig. 1A). The clinical research was divided into a discovery phase and a validation phase. Through bacterial isolation and cultivation, key bacteria associated with jaundice were identified. Subsequently, animal experiments were conducted on mice with phenylhydrazine hydrochloride (PHZ)-induced jaundice. In the clinical study, we identified differential bacteria and key bacterial species (including B. animalis) associated with moderate-to-severe jaundice. Subsequent analyses further revealed distinct gut viromes and differential BA. We further investigated gut microbiota-BA interactions in NJ, demonstrating B. animalis’ role in remodeling the BA pool and reducing hyperbilirubinemia levels in a PHZ-induced murine model (Fig. 1B). Collectively, these results demonstrate the potential of B. animalis to reshape the BA pool and highlight its therapeutic promise against hemolysis-associated jaundice and liver injury.

Fig. 1. Study design and main results.

A This study is primarily divided into three parts: clinical cohort investigation, isolation and culture of key bacterial taxa, and validation in animal models. B Integrated multi-omics analyses of gut bacteria, phages, and BA metabolites (Figs. 2–5) identified B. animalis as a key taxon associated with NJ. Through in vitro and in vivo functional experiments, B. animalis subsp. lactis Y103-OUT5 was shown to remodel BA metabolism and alleviate hyperbilirubinemia and liver injury (Figs. 6–7). Causal inference analyses further revealed the role of a gut “phage–bacteria–BA” axis in neonatal jaundice (Fig. 8). The diagram was created using Microsoft PowerPoint and BioGDP.com⁶⁹.

The clinical component comprised discovery and validation phases, through which we identified differential bacteria, including B. animalis associated with moderate-to-severe neonatal jaundice (NJ). Bacterial isolation techniques confirmed key jaundice-associated strains. Building on these findings, we conducted animal studies using phenylhydrazine hydrochloride (PHZ)-induced jaundice in mice.

Clinical characteristics of enrolled subjects

This study employed a two-phase design encompassing 120 full-term neonates, including 90 in the exploratory phase and 30 longitudinally monitored cases in the validation phase (Fig. 2A). In the exploratory phase, participants were stratified into three groups based on initial serum TBIL levels and maternal GBS colonization status: moderate-to-severe jaundice group (NJ group, n = 30), mild jaundice with maternal GBS positivity (GBS group, n = 30), and mild or non-jaundice controls (HC group, n = 30). The validation cohort comprised 14 jaundiced neonates (NJ-T group) and 16 mild-jaundice neonates born to GBS-positive mothers (GBS-T group). Fecal samples from all participants underwent metagenomic sequencing and targeted BA metabolomic profiling. Comparative analyses examined variations in gut microbial communities and BA metabolites across groups, with particular emphasis on bacterial-phage-BA interactions in NJ cases through multi-omics integration. None of the infants received BA binders or BA medications during the study.

Fig. 2. Disrupted gut microbial structure and composition in moderate-to-severe NJ.

A Cohort study design. B Clinical characteristics of participants in GBS, NJ, and HC groups. C α-diversity of gut microbiota across groups (Shannon index). D The NMDS analysis of intergroup differences. E, F Similarity analyses, E Intergroup comparison in discovery phase; F Pre- vs. post-treatment in GBS validation group. G Influence of clinical confounders on β-diversity (*p < 0.05, **p < 0.01). H Heatmap of signature bacterial abundances across Community State Types (CSTs). Rows: bacterial taxa; Columns: samples grouped by CST. I Genus-level relative abundance in discovery phase groups. J Relative abundance of gut microbial at the species level before and after NJ treatment in the validation phase.

At the exploratory phase, baseline demographic characteristics including gestational age, birth weight, head circumference, birth length, and gender distribution were comparable across all three infant groups (all p-values > 0.05), confirming balanced population enrollment for subsequent analyses. However, the cesarean section rate was significantly higher in the GBS group compared to the NJ group (16.7% vs. 10.0%, p < 0.001), and 40% of neonates in the GBS group received antibiotic treatment after birth (vs. 6.7% in the NJ group, p = 0.001). Detailed statistical results are provided in Supplementary Table 1. Other clinical covariates are visualized in Fig. 2B.

Gut microbiota dysbiosis of moderate-to-severe NJ

Alterations in gut microbiota diversity and community state types

To investigate the characteristics of gut microbiota associated with moderate-to-severe NJ, we conducted metagenomic analysis on all collected fecal samples from study participants. Regarding gut microbial diversity, the plateau phase of rarefaction curve indicated that our sample size was sufficient to comprehensively capture microbial diversity (Supplementary Fig. 1A). The results revealed statistically significant differences in α-diversity among the three groups (p = 0.00022), with both NJ and GBS groups showing higher diversity than the HC group (Fig. 2C). Considering potential confounding factors, we performed stratified analyses using the Kruskal–Wallis test. In antibiotic-exposed subgroups, no significant α-diversity differences were observed (p = 0.77), whereas significant differences emerged in antibiotic-naive subgroups (p = 0.00045) (Supplementary Fig. 1B). Similarly, α-diversity metrics remained statistically comparable (p = 0.061) among infants with maternal pregnancy complications, whereas a significant difference was observed in those without maternal pregnancy complications (p = 0.005) (Supplementary Fig. 1C). Non-metric multidimensional scaling (NMDS) analysis demonstrated significant clustering patterns among groups (Fig. 2D), with pregnancy complications significantly influencing β-diversity (Supplementary Fig. 1D). The analysis of similarities (Anosim) test further confirmed differences in the gut microbiota among the three groups (R = 0.0744, p = 0.002) (Fig. 2E). In the validation phase, significant microbiota alterations were observed in GBS group before and after treatment (p = 0.003) (Fig. 2F). Non-parametric multivariate analysis of variance (Adonis) evaluating clinical confounders revealed that total bile acids (TBA), TBIL, white blood cell count (WBC), hemoglobin (HB), and birth weight accounted for substantial proportions of β-diversity variation, with TBA and TBIL showing particularly significant intergroup differences (p < 0.01) (Fig. 2G).

To elucidate the clustering patterns of gut microbiota structure, we performed classical multidimensional scaling (MDS) and non-metric multidimensional scaling (NMDS) based on Bray-Curtis dissimilarity matrices. Gap statistical analysis determined five clusters as optimal for explaining microbial variance (Supplementary Fig. 1E). MDS analysis revealed significant inter-sample dissimilarities (Supplementary Fig. 1F), while NMDS further stratified samples into five distinct community state types (CST 1–5) across NMDS 1–2 and 3–4 planes, with significant compositional differences between clusters (Supplementary Fig. 1G). CST1 primarily occurred in GBS and HC groups, characterized by Achromobacter insuavis. CST2 predominantly featured NJ group samples driven by Escherichia coli and Streptococcus salivarius with notably reduced Bifidobacterium animalis abundance. CST3 was mainly distributed in the HC group and dominated by Enterococcus faecium. CST4 was primarily structured by Escherichia coli, whereas CST5 was defined by Streptococcus salivarius predominance (Fig. 2H).

Analysis of gut microbial composition revealed distinct taxonomic patterns across groups. At the genus level, both GBS and NJ groups were predominantly characterized by Streptococcus (25.7% in GBS vs. 22.2% in NJ) and Escherichia (27.7% in GBS vs. 20.5% in NJ) (Supplementary Table 2). The HC group was mainly colonized by Enterococcus, Streptococcus, and Achromobacter genera (Fig. 2I). Following jaundice treatment in the validation cohort, we observed substantial microbial shifts: pathogenic taxa including Escherichia coli, Clostridium perfringens, Bacteroides fragilis, and Streptococcus vestibularis significantly decreased, while beneficial species such as Bifidobacterium longum, Streptococcus salivarius, and Phocaeicola vulgatus markedly increased (Fig. 2J). Concurrently, the GBS intervention group demonstrated elevated levels of putative probiotics Streptococcus salivarius and Lactococcus lactis post-treatment (Supplementary Fig. 1G).

Key bacteria associated with moderate-to-severe NJ

To delineate gut microbial features in moderate-to-severe NJ, we integrated microbiome multivariate association with linear models (MaAsLin2) across discovery and validation cohorts. The ASV-based taxonomic tree confirmed B. animalis within phylum Actinomycetota and genus Bifidobacterium (Fig. 3A). Chord diagrams and Sankey plots revealed distinct core microbiota distributions between GBS and NJ groups (Fig. 3B, C).

Fig. 3. Characteristic bacterial markers and metabolic function profiles associated with moderate-to-severe NJ.

A Taxonomic distribution of differential ASV markers (MaAsLin2 analysis, FDR < 0.05) from phylum to species. B Chord diagram illustrating the distribution and abundance of key bacterial taxa in the GBS and NJ groups. C Sankey diagram tracing the distribution and abundance of key bacteria between the GBS and NJ groups. The width of the flow lines reflects the relative abundance of each bacterium across groups. D Relative abundance differences of bacterial species in the GBS and NJ groups compared to the HC group, plotted against statistical significance. The x-axis represents the coefficient of relative abundance difference for bacterial species between groups (positive: positively correlated; negative: negatively correlated). The y-axis shows the −log10-transformed p-value, with larger values indicating greater significance. Green dots (Neg) indicate negative correlations; red dots (Pos) indicate positive correlations; gray dots (NS) indicate non-significance. E Relative abundance of intestinal bacteria at the species level before and after NJ treatment during the validation phase (left panel), and relative abundance differences of bacterial species before and after GBS progression (right panel). F, G Bubble plots displaying indicator species selected during the exploration phase (F) and validation phase (G). Bubble size represents average abundance, and color indicates the significance of associations determined by MaAsLin2 analysis (true or false). H The distribution of microbial metabolic pathways in different groups. I Functional association analysis based on generalized microbial metabolic model was used to analyze the correlation and significance between different metabolic pathways and groups.

Compared to the HC group, both NJ and GBS groups exhibited significant reductions in eight bacterial species. The GBS group specifically lacked Enterococcus faecalis and Klebsiella quasipneumoniae, whereas the NJ group demonstrated significantly reduced abundance of Devosia equisanguinis, Delftia acidovorans, Pedobacter nutrimenti, A. insuavis, Halomonas daqingensis, and most notably, B. animalis (Fig. 3D). Concurrently, 11 species increased in abundance (7 predominant in GBS, 4 in NJ) (Fig. 3D). Therapeutic interventions triggered characteristic shifts: NJ treatment significantly elevated probiotics including B. animalis (most pronounced change, p < 0.001), L. lactis, Lactococcus raffinolactis, and Streptococcus infantarius, while GBS treatment increased B. animalis and Streptococcus lactarius (Fig. 3E).

Indicator species analysis further revealed discovery-phase HC group was characterized by Bifidobacteria species, such as B. pseudocatenulatum, B. longum, and B. bifidum, NJ group by opportunistic pathogens (Staphylococcus haemolyticus, Ralstonia insidiosa), and GBS group by enrichment of Streptococcus equinus (Fig. 3F). Validation confirmed B. animalis as the key post-treatment indicator in NJ, while S. lactarius and Rothia mucilaginosa emerged as therapeutic indicators in GBS (Fig. 3G). Collectively, these findings reveal dynamic dysbiosis patterns and microbial signatures of therapeutic response in NJ. To improve the understanding of gut microbial functions in the NJ and GBS groups, metabolic pathways that were differentially enriched between the groups were identified (Fig. 3H, I).

To evaluate and control for the potential confounding effects of the aforementioned microbiota-disease associations, we systematically analyzed key clinical variables—including antibiotic exposure, delivery mode, pregnancy complications, and infection status—using the metadeconfoundR approach. The results demonstrated that antibiotic exposure exerted a decisive influence on gut microbiota structure. Notably, after controlling for antibiotic effects, the genus Ralstonia replaced Enterococcus as a key taxon associated with clinical phenotypes (Supplementary Fig. 2A). Delivery mode also showed significant confounding effects on several taxon-phenotype associations (Supplementary Fig. 2B). In contrast, the modulating effects of pregnancy complications and infection status appeared relatively limited (Supplementary Fig. 2C, D). Importantly, after adjusting for all the aforementioned confounders, core findings such as the association between Phocaeicola and TBIL remained highly stable (deconfounded, p < 0.01), and the association between Streptococcus and TBIL was not substantially altered (Supplementary Fig. 2A–D). These findings further support the reliability of these microbial features as robust biological markers associated with neonatal jaundice.

Gut virome-keystone bacteria interactions

This study systematically analyzed the characteristics of gut virome associated with moderate-to-severe NJ and their interactions with bacteria. Although there was no significant difference in the α-diversity of the virome (p = 0.58) (Fig. 4A), NMDS analysis revealed significant clustering of the enteric viral communities among the three groups (Fig. 4B). The Anosim analysis provided further robust confirmation of the significant differences among groups (R = 0.1624, p = 0.001) (Fig. 4C). Clinical factor analysis showed that TBA, TBIL, and birth weight were key factors influencing viral β-diversity (p < 0.01) (Supplementary Fig. 3A). In terms of viral composition, the HC group was dominated by Clostridium phage phiCD6356, Streptococcus phage 5093, and Planktothrix phage PaV LD, with the proportion of Clostridium phage phiCD6356 in the HC group (24.2%) being significantly higher than that in the NJ (3.3%) and GBS groups (6.7%) (Supplementary Fig. 2B, Supplementary Table 3). Notably, the total proportion of Streptococcus phages in the GBS group (26.1%) was significantly higher than that in the NJ (14.1%) and HC groups (14.3%) (Supplementary Fig. 3B, Supplementary Table 3). Through community structure analysis, we identified five characteristic viral CSTs with significant differences in composition between different clusters (Supplementary Fig. 3C, D). Specifically, CST5, which was dominant in the GBS group, was characterized by Streptococcus phages; CST4, dominant in the NJ group, was characterized by Enterobacteria phage HK629; and CST3, dominant in the HC group, was characterized by Clostridium phage phiCD6356 (Fig. 4D).

Fig. 4. Alterations in phage abundance associated with “keystone” bacteria correlate with moderate-to-severe NJ.

A Gut virome α-diversity across cohorts. The Y-axis represents the Shannon index, where higher values indicate greater species evenness and biodiversity. B NMDS (Non-metric Multidimensional Scaling) ordination of β-diversity in gut viromes. “Yes” denotes maternal gestational complications; “No” indicates absence thereof. C Similarity analysis of gut viromes among three cohorts (HC, NJ, GBS) using ANOSIM. D Heatmap depicting relative abundance of signature viral species across CST clusters (CST1-CST5). Branch colors denote cohort assignments. E Phylogenetic tree of gut viruses reconstructed from viral contigs. Branch lengths reflect evolutionary divergence; node colors indicate cohort origins. F Procrustes analysis demonstrating co-variation between gut bacteriophages and bacterial communities. Subplots display principal coordinate (PCo) distributions for paired cohort comparisons (HC vs. NJ; HC vs. GBS; NJ vs. GBS). G Spearman correlation analysis of the degree and significance of association between gut phages and various host bacterial genera.

To identify phages associated with moderate-to-severe NJ, we analyzed the bacteriophages from the exploratory stage using Maaslin2 and multilevel pattern analysis. Similar to gut bacteria, bacteriophages have their own evolutionary processes (Fig. 4E). There was a potential coexistence pattern of bacteriophages between the GBS and NJ groups (Supplementary Fig. 4A). The association Sankey diagram visually displays the distribution and relative abundance of key phages in the GBS and NJ groups (Supplementary Fig. 4B). To investigate the influence of clinical variables on key phages, we analyzed the relationships between the top 50 phages in terms of relative abundance and clinical variables (Supplementary Fig. 4C). Compared with the HC group, 23 phages showed decreased abundance in the disease groups (9 in the GBS group and 14 in the NJ group), while 80 phages showed increased abundance (30 in the GBS group and 50 in the NJ group) (Supplementary Fig. 4D). We also found that Yersinia phage L-413C was both a differential phage and an indicator phage for the NJ group, while the GBS group was characterized by Lactobacillus phage phiPYB5 and Enterococcus phage EFC-1 (Supplementary Fig. 4E).

In terms of enteric phage-bacteria interactions, Procrustes analysis revealed significant correlations between enteric bacteria and phages in all groups (Fig. 4F). Spearman correlation analysis further identified negative correlations between D. equisanguinis and Streptococcus phage SMP, Caldibacillus debilis and Pseudomonas phage LKA5, P. nutrimenti and Streptococcus phage SMP, as well as Leifsonia aquatica and Streptococcus phage SMP, suggesting potential lytic effects. Conversely, positive correlations were observed between A. insuavis and Pseudomonas phage LKA5, Enterococcus faecium and Clostridium phage phiCD6356, P. nutrimenti and Pseudomonas phage LKA5, as well as Leifsonia aquatica and Gordonia phage Lucky10, indicating possible symbiotic relationships between these phages and their corresponding bacteria (Fig. 4G).

Gut BA metabolism alteration of moderate-to-severe NJ

Exploratory phase clinical data revealed significantly elevated serum TBA levels in the NJ group compared to GBS and HC groups (Supplementary Table 1). Given the established link between BA metabolism and gut microbiota²⁶, we performed targeted metabolomic analysis of 43 BA in discovery and validation cohorts using liquid chromatography-mass spectrometry (LC-MS). In the discovery phase, distinct clustering patterns in BA profiles were observed across groups, demonstrating significant intergroup metabolic divergence (Fig. 5A). The NJ group exhibited a notably higher secondary-to-primary BA ratio (2.20 vs. 1.63 in GBS and 1.92 in HC) and elevated free-to-conjugated BA ratio (2.88 vs. 0.99 in GBS and 1.53 in HC) (Supplementary Table 4).

Fig. 5. Alterations in gut BA associated with NJ.

A OPLS-DA analysis revealed significant differences in BA between NJ and HC, GBS and HC, and GBS and NJ. B Screening of differential BA metabolites. The OPLS-DA results for pairwise comparisons among HC, NJ, and GBS groups are displayed with variable importance in projection (VIP) scores on the x-axis. Metabolites with VIP > 1 are shown, where higher VIP scores indicate greater discriminatory power for group separation. C Bar plot quantifying shared and unique differential BA across the three groups (GBS, HC, NJ). D Composition of the top 10 most abundant BA metabolites in each group. E Changes in differential BA metabolites following jaundice treatment in the validation cohort. F Correlation matrix between BA profiles and clinical variables.

To identify differential BA metabolites among the three groups, we performed two-by-two comparisons using orthogonal partial least squares discriminant analysis (OPLS-DA) on the three groups of data in the exploratory phase. This analysis revealed 15, 13, and 18 differential BA metabolites in the NJ, GBS, and HC groups, respectively. taurohyocholic acid (THCA), taurocholic acid (TCA), norcholic acid (NorCA), taurodeoxycholic acid (TDCA), and deoxycholic acid (DCA) contributed to differentiating the NJ group from the HC group. TCA, glycoursodeoxycholic acid (GUDCA), tauroursodeoxycholic acid (TUDCA), and glycochenodeoxycholic acid (GCDCA) contributed to differentiating the GBS group from the HC group. THCA, TCA, TDCA, taurohyodeoxycholic acid (THDCA), and others contributed to differentiating the GBS group from the NJ group (Fig. 5B). A total of 15 differential bile acids were present in the NJ and GBS groups compared to the HC group (Fig. 5C). The OPLS-DA model parameters confirmed their good explanatory and predictive ability (Supplementary Fig. 5).

We analyzed the top ten BA in terms of relative abundance and found that all three groups were dominated by secondary BA. The highest percentages were of lithocholic acid (LCA) and its isomers, iso-lithocholic acid (isoLCA) and allo-lithocholic acid (alloLCA) (Fig. 5D). The total percentage of LCA and its isomers was higher in the NJ group (37.5%) than in the GBS (23.1%) and HC (29.1%) groups (Supplementary Table 5). In the validation phase, post-treatment analysis of NJ infants demonstrated significant reductions in six differential bile acids, including GCDCA, tauro-β-muricholic acid (TβMCA), and GUDCA (Fig. 5E).

In order to further investigate the associations between different intestinal BA metabolites and a variety of clinical factors, a strong association was found between serum TBIL and specific BA using MaAsLin2 analysis. TBIL positively correlated with TDCA, GUDCA, THDCA, and TCA, and significantly negatively correlated with DCA, UDCA, and LCA (Fig. 5F).

Multi-omics reveals interplay among the gut virome, bacteria, and BA

Through comprehensive multi-omics integration, we systematically investigated the interplay between gut microbiota and BA metabolism in NJ. Procrustes analysis showed a strong correlation between gut bacteria and BA among the different groups (Fig. 6A). Furthermore, the utilization of circle plots facilitated the visualization of the correlation between the distribution of the samples across various categorical dimensions, including BA and the bacterial α-diversity (Shannon index) (Fig. 6B). The Sankey diagram visualization analysis further elucidated the relationship between specific BA and the α-diversity of intestinal bacterial. The results demonstrated a significant negative correlation between LCA and its isomers and the bacterial Shannon index, suggesting that the accumulation of this type of secondary BA may inhibit microbial diversity (Fig. 6C). Canonical correspondence analysis (CCA) was utilized to examine the impact of clinical subgroups on the distribution of BA, revealing substantial disparities among the groups (Fig. 6D).

Fig. 6. Gut microbiota-BA interactions in NJ.

A Procrustes analysis of BA metabolites and gut bacterial communities. B Phylogenetic tree (circular layout) displaying gut bacterial taxa, with asterisks indicating species significantly associated with Shannon index or BA levels. C Sankey diagram illustrating correlations between specific BA metabolites and bacterial α-diversity. D Relationship between gut microbiota and BA metabolites based on CCA. The rings indicate the clustering of the samples, and the arrows indicate the influence of environmental variables on the distribution of the samples. E NMDS ordination of gut bacterial communities under differential BA conditions. F Correlation heatmap between core gut bacteria and BA metabolites. Colors indicate the strength of the correlation (red for positive correlation, blue for negative correlation). G NMDS ordination of bacteriophages communities stratified by BA profiles. H Correlation heatmap between core gut phages and BA metabolites. I, J Co-inertia analysis (CoIA) of gut bacteria and BA. I Distribution of key BA metabolites in the CoIA ordination space. Green represents Axis 1, and red represents Axis 2. J Distribution of key gut bacterial species in the CoIA ordination space. K, L Reduced abundance of B. animalis-derived BSH in moderate-to-severe NJ (K). Recovery of B. animalis-derived BSH abundance post-NJ treatment (L).

By NMDS analysis of bacterial communities from the three groups of samples under the influence of specific BA (TCA, DCA) environmental factors, we observed clear spatial distribution differences (Fig. 6E). Specifically, the bacterial communities of GBS group and NJ group showed a unique community drift in the upper right quadrant of the coordinate quadrant, which was closely related to the environmental conditions of high TCA concentration and low DCA concentration, thus forming a significant difference from the HC group (Fig. 6E). This shift along a specific axis suggests that bacterial strains capable of tolerating high TCA concentrations may be present in the GBS and NJ groups compared to the HC group; however, these strains appear to be repulsive to high DCA concentrations. Further Mantel test results, based on Bray-Curtis distance matrix, strongly confirmed the strong correlation between bacterial community structure and BA profile (Supplementary Fig. 6A–C), and specific NJ-associated marker bacteria showed a strong positive correlation with conjugated BA (Fig. 6F). These findings suggest that there may be interactions between BA and aspects of bacterial community structure, and that bacterial communities of different groups respond differently to changes in BA concentration. NMDS and Mantel tests were also performed to explore the differences between bacteriophages and intestinal BA. The results showed that group GBS and group NJ bacteriophages were associated with environments with high TCA concentration and low DCA concentration, thus forming a significant differentiation from group HC (Fig. 6G). Meanwhile, Mantel test results showed a strong correlation between bacteriophages and BA profiles (Supplementary Fig. 6D–F), and the NJ-associated core bacteriophages also showed a strong positive correlation with conjugated BA (Fig. 6H).

To explore the complex interactions among bacteriophages, bacteria and BA, multivariate factor analysis (MFA) was performed. The results showed that there was a correlation among the three (Supplementary Fig. 7A–C). The composition of intestinal bacteria was strongly correlated with bacteriophages (RV = 0.44, p < 0.001), suggesting that there might be a close ecological relationship between virus and bacteria (Supplementary Table 6, Supplementary Fig. 7A–C). There was also a moderate positive correlation between bacterial species and BA (RV = 0.41, p < 0.001), indicating that bacteria may play an important role in BA metabolism (Supplementary Table 6, Supplementary Fig. 7A–C). There was also a moderate correlation between viral composition and BA (RV = 0.33, p < 0.001), suggesting that phages may indirectly affect BA metabolism by affecting bacteria (Supplementary Table 6, Supplementary Fig. 7A–C). These findings reveal the existence of a complex network of interactions among bacteriophages, bacteria, and BA metabolites. During the exploratory phase, we employed co-inertia analysis (CoIA) to investigate the coordinated variations between gut microbiota composition and BA metabolic profiles. An integrated analysis of microbiota composition and BA metabolites was performed by CoIA, and statistical significance of the principal axis was first assessed by eigenvalues with a broken stick model. Analysis in the biplot showed that secondary BA such as NorDCA, glycodehydrocholic acid (GDHCA), and TUDCA showed high loading on Axis 1 and were far from the origin of the coordinates, suggesting their significant contribution to the co-variation of the main axis (Fig. 6I). Meanwhile, the dual-labeling map of the key strain shows that the spatial distribution of Thermobispora bispora on Axis 1 is highly similar to that of GDHCA, indicating a positive correlation between the two, while the spatial distribution of B. animalis on Axis 1 is opposite to that of TUDCA, suggesting a negative correlation (Fig. 6J). The analysis of principal axis loadings further confirmed that the contributions of bacteria and BA were significantly higher on Axis 1 than on Axis 2 (Supplementary Fig. 7D, E). CoIA revealed a significant co-variation between BA metabolites and bacterial communities, where the first and second axes of CoIA explained 91.31% and 4.22% of the total projected inertia, respectively (Supplementary Fig. 7F).

As a key hydrolase in BA metabolism, bile salt hydrolase (BSH) expression levels are closely linked to gut microbiota function. To investigate the BSH gene expression profile in moderate-to-severe NJ patients, we conducted systematic genomic analysis via metagenomic sequencing. In the exploratory phase, compared to HC group, the NJ group exhibited significantly reduced metagenome-inferred BSH gene abundance in 9 bacterial species, including B. animalis subsp. lactis CNCM I-2494 (Fig. 6K). The validation phase revealed that effective treatment significantly restored metagenome-inferred BSH gene abundance of B. animalis subsp. lactis CNCM I-2494 in NJ patients (Fig. 6L), suggesting the potential involvement of this BSH in NJ pathophysiology and recovery, although direct enzymatic activity was not measured in these clinical samples.

B. animalis subsp. lactis Y103-OTU5 reshapes the BA pool and alleviates jaundice

We identified B. animalis as a key discriminatory and indicator species post-NJ treatment, with its BSH gene abundance potentially influencing disease resolution. To clarify its metabolic impact, we isolated B. animalis subsp. lactis Y103-OTU5 from fecal samples, whose 16S rRNA gene sequence showed high similarity (100%) to B. animalis subsp. lactis strain YIT 4121 (Fig. 7A). This strain exhibits broad metabolic remodeling capabilities (Supplementary Fig. 7H). After cultivating the isolate in BSM solid medium under anaerobic conditions at 37 °C, we performed quantitative profiling of 43 BA using LC-MS. The results demonstrated significant reductions (p < 0.01) in TBA, primary BA, and secondary BA following inoculation. Notably, DCA levels among secondary unconjugated BA showed a marked increase (p < 0.01), indicating potential biotransformation mechanisms via deconjugation or dehydroxylation activities. These findings confirm that B. animalis subsp. lactis Y103-OTU5 possesses robust BA metabolic activity, effectively reshaping the BA pool composition (Fig. 7B).

Fig. 7. B. animalis subsp. lactis Y103-OTU5 modulates BA metabolism and related pathways.

A, B Isolation and functional validation of B. animalis. A Isolation of B. animalis via culturomics from fecal samples. B In vitro validation of B. animalis’s metabolite-reshaping capacity. The strain significantly alters the bile acid pool, as evidenced by reduced total TBA and increased DCA levels. C–G Animal experiment on B. animalis intervention in PHZ-induced hemolytic jaundice. C Design of the experiment. D Serum bilirubin level. E Typical appearance of liver. F Representative liver histology by H&E staining (× 200). G Volcano plot of differentially expressed genes (PHZ_B. animalis vs. PHZ groups). H Reactome pathway enrichment analysis of differentially expressed genes. I Schematic diagram of BA biosynthesis and metabolism.

To investigate the effect of B. animalis on hemolytic jaundice, we conducted animal experiments using a PHZ-induced hyperbilirubinemia model (Fig. 7C). Phenylhydrazine hydrochloride (PHZ) induces hemolysis by damaging red blood cells, leading to excessive accumulation of bilirubin in the body that exceeds the metabolic capacity of the liver and intestines. The modeling principle resembles the pathogenesis of hyperbilirubinemia, thus effectively simulating the condition. After PHZ administration, mice exhibited significantly elevated serum TBIL, direct bilirubin (DBIL), and TBA, indicating the successful establishment of a mouse jaundice model (Fig. 7D, Supplementary Fig. 8B). Additionally, the AST/ALT ratio increased, further implying serious liver cell damage (Supplementary Fig. 8D). Fortunately, pretreatment with B. animalis subsp. lactis Y103-OTU5 reversed these adverse trends significantly. The results demonstrated that B. animalis subsp. lactis Y103-OTU5 could alleviate PHZ-induced hemolysis in mice, effectively reducing serum TBIL and indirect bilirubin (IBIL) levels, and exhibiting protective effects against hyperbilirubinemia and liver injury (Fig. 7D).

In this study, we further evaluated the protective effect of B. animalis on liver injury associated with PHZ-induced hyperbilirubinemia using hematoxylin and eosin (H&E) staining. Liver surface morphology was examined in three groups of mice. The control group exhibited a healthy, ruddy, and smooth liver surface, the model group showed a dark black liver color, and the B. animalis group showed a reddish-black liver color, indicating that B. animalis intervention could partially ameliorate liver injury (Fig. 7E). H&E staining results revealed that hepatocytes in the control group were normal, tightly packed, and had a clear and intact cellular structure, with large, round nuclei and orderly arrangement. In contrast, hepatocytes in the PHZ group showed inflammatory infiltration, nuclear pyknosis, widening of intercellular spaces, edema, and loose arrangement. However, compared with the observations in the PHZ group, pretreatment with B. animalis significantly improved these histopathological changes (Fig. 7F).

To elucidate the molecular mechanisms underlying this hepatoprotection, hepatic transcriptome analysis revealed profound alterations in BA metabolism pathways. Specifically, comparative analysis identified 313 differentially expressed genes (DEGs) between PHZ and control groups (Supplementary Fig. 9B), while 393 DEGs (include Cyp7a1, Cyp7b1) were detected between the B. animalis and PHZ groups (Fig. 7G). Reactome enrichment analysis revealed significant activation of the pathway Cytochrome P450-arranged by substrate type (R-MMU-211897, p < 0.005), with coordinated upregulation of core enzymes Cyp7a1 and Cyp7b1 driving enhanced bile acid flux (Fig. 7H, Supplementary Table 8). Consistent with this, KEGG analysis identified pronounced enrichment of Steroid hormone biosynthesis (mmu00140, p < 0.005), where Cyp7a1 and Cyp7b1 serve as nodal points linking bile acid synthesis to steroid metabolism (Supplementary Fig. 9C, Supplementary Table 9). Mechanistically, B. animalis intervention activated the hepatic cytochrome P450 system, driving coordinated upregulation of critical BA-synthetic enzymes. This transcriptional potentiation significantly induced the rate-limiting enzyme Cyp7a1 and elevated Cyp7b1 activity in alternative sterol-hydroxylation pathways, as depicted in the schematic integration (Fig. 7I). Collectively, these results demonstrate that B. animalis attenuates PHZ-induced hepatotoxicity by reprogramming the transcriptional landscape of sterol-hydroxylating CYPs, thereby restoring metabolic homeostasis through amplified bile acid flux.

Gut microbiota-related markers validation in B. animalis treated jaundice mice

To explore the impact of B. animalis on the gut microbiota and bile acid metabolites in mice, we collected fecal samples from the day of sacrifice and performed metagenomic sequencing, targeted BA quantification, and differential analysis. The results showed that compared to the PHZ group, both the Control (BALB) group and the B. animalis prevention group exhibited a significant decrease in 16 gut phage species and a significant increase in 8 phage species, such as Lactococcus phage P087 and Bacillus phage SP10 (Fig. 8A). Compared to the PHZ group, the B. animalis prevention group showed a significant increase in 7 gut bacterial species, such as Paramuribaculum intestinale (Fig. 8B). Meanwhile, both the Control group and the B. animalis prevention group displayed a significant increase in 6 fecal BA metabolites, including DCA, CDCA, and HDCA, compared to the PHZ group (Fig. 8C).

Fig. 8. Causal inference reveals the role of the gut “phage-bacteria-BA” axis through validation of microbiota-related markers.

A Volcano plot of differential viral species in fecal samples of mice from the remaining two groups compared to the PHZ group. B Volcano plot of differential fecal bacteria. C Volcano plot of differential fecal bile acids. D SEM of gut phage-bacteria-BA interactions in the clinical study. E SEM of gut phage-bacteria-BA interactions in animals.

Causal inference reveals the role of the gut “phage-bacteria-BA” axis in jaundice

To comprehensively understand the interactions among gut phages, bacteria, and bile acid metabolism, we employed structural equation modeling (SEM) to analyze interactions among viruses, bacteria, and BA metabolites. The results revealed that certain bacterial taxa act as key mediators within the gut phage-bacteria-BA axis, influencing the development of jaundice by modulating BA metabolism (Fig. 8D, E). The clinical results revealed intricate regulatory dynamics, with gut phages indirectly modulating BA metabolism through bacterial community restructuring (Fig. 8D). B. animalis significantly correlated with key NJ clinical markers, including TBIL and BA levels, and more importantly, directly influenced NJ progression through remodeling of BA metabolism (Fig. 8D). Meanwhile, gut phages were also found to affect jaundice through their impact on BA, specifically DCA and CDCA (Fig. 8D, E). Furthermore, bacteria encoding the BSH gene were shown to modulate serum total bilirubin levels (Fig. 8E). The BA receptors FXR and TGR5 are also associated with NJ (Fig. 8E).

Discussion

The role of intestinal microecology in the pathogenesis of moderate-to-severe NJ remains unclear. Building on our previous finding that implicated BA as clinical markers of NJ²⁶, this study highlights the significance of specific BA-related gut microbes. Clinical cohort analysis revealed gut dysbiosis in NJ infants, marked by dominance of Streptococcus and Escherichia genera and depletion of potentially protective bacteria like B. animalis. Effective treatment restored gut microbiota homeostasis, reducing pathogenic bacteria (including Escherichia coli and Clostridium perfringens) and enriching beneficial species, particularly B. animalis, suggesting its pivotal role in NJ pathogenesis alleviation. Combining clinical and experimental data, we showed that B. animalis remodels the BA pool, effectively reducing hyperbilirubinemia in the PHZ model and associated liver injury. Compared to HC, NJ infants exhibited reduced metagenome-inferred BSH gene abundance in B. animalis, which was restored post-treatment. In vitro assays confirmed its ability to decrease TBA levels and increase DCA production via deconjugation, indicating bidirectional BA metabolism regulation. Additionally, NJ-associated gut viromes were dominated by Streptococcus phages, with significant correlations to bacterial communities. SEM revealed that gut viromes indirectly influence BA metabolism by reshaping microbiota structure. Notably, B. animalis correlated with key NJ clinical biomarkers and directly impacted disease progression through BA modulation. These findings offer novel insights into the microbial-metabolic mechanisms underlying NJ pathogenesis and provide a theoretical basis for developing microbiota-targeted therapies.

Accumulating evidence for the importance of gut microbiota in jaundice. Our clinical cohort analysis revealed profound dysbiosis in moderate-to-severe NJ patients, characterized by increased abundances of Streptococcus and Escherichia genera. Consistent with previous studies, the GBS and other Streptococcus species have the ability to damage red blood cells²⁸, which subsequently leads to abnormal bilirubin metabolism and causes or aggravates jaundice. Concurrently, Escherichia coli within the Escherichia genus secretes excessive β-glucuronidase, hydrolyzing conjugated bilirubin into unconjugated bilirubin (UCB)^29,30. This UCB is reabsorbed via the enterohepatic circulation, elevating serum bilirubin levels and aggravating jaundice. In addition, the microbiota abundance results before and after NJ treatment showed that Escherichia coli and Clostridium perfringens in the gut of NJ patients after treatment were reduced compared with those before treatment, accompanied by the recovery of Bifidobacterium, reinforcing the key role of microbiota regulation in the remission of NJ. Notably, MaAsLin2 analysis further revealed a decrease in the abundance of B. animalis in the NJ group compared with the HC group, and the results of the validation phase demonstrated an increase in the abundance of some key bacteria, including B. animalis, after NJ treatment. This suggests that it may play a protective role by competitive inhibition of pathogenic bacteria colonization or by affecting intestinal metabolites.

Bacteriophages (phages), viruses infecting intestinal bacteria, are abundant in the gut microbiome³¹. Phages interact with bacteria: they lyse bacteria to regulate populations and maintain gut microbiota balance, and also form symbiotic relationships, like temperate phages integrating into bacterial genomes for co-evolution^32–34. A study elucidated the coordinated evolution of gut phages and bacteria during the first four years of infancy, demonstrating synchronous abundance dynamics between phages and their host bacteria³⁵. Our study revealed significant compositional correlations between viral and bacterial communities, with certain phages displaying lytic activity against specific bacteria. For instance, proliferation of Streptococcus phages might indirectly promote niche occupation by beneficial bacteria through lysis of pathogenic streptococci. However, phage-bacteria relationships also involve symbiosis. These findings depict gut phage communities as dynamic, multifaceted regulators with “predator” and “symbiotic partner” attributes in bacterial interactions. Moreover, gut phages influence disease pathogenesis by modulating bacterial metabolism and immune responses. Norman JM et al. demonstrated phage-mediated immune regulation in colitis, linked to Caudovirales phages expansion³⁶. Nakatsu G et al. revealed phage impacts on colorectal cancer through bacterial community restructuring and immune microenvironment modulation³⁷. Notably, fresh sterile fecal filtrates alleviate Clostridioides difficile-infection symptoms, providing proof-of-concept for therapeutic phage applications³⁸. Future studies should explore gut phages’ mechanistic role in neonatal hyperbilirubinemia.

BA metabolism plays a pivotal role in NJ pathogenesis. Our clinical data revealed significantly elevated serum TBA levels in moderate-to-severe NJ infants compared to controls, with targeted metabolomics demonstrating profound disruption of “BA pool” homeostasis characterized by an increased free/conjugated BA ratio. This metabolic disturbance likely reflects the combined effects of impaired hepatocyte function and intrahepatic cholestasis in moderate-to-severe jaundice cases, where damaged liver cells exhibit reduced BA uptake capacity and compromised conjugation efficiency, leading to systemic accumulation of free BA. Notably, secondary BA predominated across all study groups, with LCA and its isomers constituting the major components. LCA is generated through gut microbial 7α-dehydroxylation of chenodeoxycholic acid (CDCA)³⁹, and its characteristic elevation in pathological or cholestatic jaundice⁴⁰ may stem from neonatal gut dysbiosis favoring 7α-dehydroxylating bacteria coupled with developmentally immature hepatic sulfation pathways that impair LCA detoxification. The pathophysiological significance of LCA remains controversial. One study has suggested that it can promote the metastasis of tumor cells⁴¹, but the another recent study has demonstrated that LCA has anti-aging effect⁴². Interestingly, one study has demonstrated that alloLCA can restore BA homeostasis and restore a healthier gut microbiota composition⁴³. While these findings highlight the complex interplay between BA metabolism and NJ, the precise mechanistic relationship warrants further investigation to elucidate potential therapeutic targets in neonatal hyperbilirubinemia management.

Multi-omics is often used to reveal disease markers⁴⁴. This study systematically elucidates the characteristics of microbe-host metabolic interactions in NJ through integrated multi-omics analysis of gut bacteria, phages, and BA metabolites. Both Procrustes and CCA analyses consistently demonstrated significant co-variation between NJ-associated gut bacterial community structure and BA composition, suggesting gut microbiota may participate in NJ pathogenesis through BA metabolism modulation. Notably, Mantel tests revealed strong correlations between NJ-associated bacterial or phage biomarkers and specific BA profiles, indicating two potential microbial regulatory mechanisms: (1) direct enzymatic catalysis of BA deconjugation and modification by bacteria expressing BSH activity; and (2) indirect metabolic pathway modulation through phage-mediated bacterial lysis. Crucially, MFA analysis first established a “phage-bacteria-BA” tripartite interaction framework, revealing phages may indirectly drive BA metabolic remodeling by reshaping bacterial community structure, though direct phage regulation of specific bacterial functions requires experimental validation. SEM identified significant interactions among gut viruses, bacteria, and BA metabolites. While consistent with previous reports on phage-bacteria interactions affecting host metabolism⁴⁵, this study first established a three-tier regulatory framework incorporating phages in NJ, providing novel molecular insights into the “virus-microbiota-BA” axis. Future studies should employ germ-free animal models transplanted with specific virus-bacteria combinations to validate this axis’s causality.

Recent years have witnessed well-documented health benefits of probiotics for gastrointestinal disorders, immune modulation, type 2 diabetes and hypertension. Probiotics also show therapeutic potential for NJ, Bifidobacteria increase defecation frequency, reduce enterohepatic circulation, regulate intestinal pH, and inhibit β-glucuronidase activity to enhance bilirubin excretion^46–48; Lactobacilli upregulate tight junction protein expression in enterocytes, reducing bilirubin leakage and reabsorption^48,49. Early postnatal administration of probiotics like Lactobacillus rhamnosus GG positively influences bilirubin metabolism and may reduce hyperbilirubinemia risk⁵⁰. Tsai et al. demonstrated that phototherapy combined with L. salivarius AP-32 or B. animalis subsp. lactis CP-9 significantly shortened treatment duration and accelerated bilirubin clearance¹⁵.

To clarify B. animalis’ BA regulatory mechanisms in NJ, our in vitro experiments confirmed its “BA pool” remodeling capacity, significantly reducing both primary and secondary BA. Metagenomic analysis revealed post-treatment upregulation of BSH gene expression in B. animalis, consistent with its observed metabolic activity. Considering that B. animalis has been experimentally demonstrated to express functional BSH capable of deconjugating bile acids such as GDCA and TDCA⁵¹, our observations of concurrent shifts in the bile acid profile toward deconjugated forms strongly suggest that BSH-mediated deconjugation is a plausible key mechanism underlying this metabolic remodeling. However, the contribution of other bile acid-modifying enzymes cannot be excluded, and direct evidence of BSH-mediated deconjugation in the neonatal gut environment remains to be established. The DCA elevation observed in vitro may stem from its unique potential 7α-dehydroxylase activity, a feature that warrants further investigation in the context of probiotic screening for neonatal applications. SEM further suggested B. animalis’ pivotal role in NJ through BA modulation. Our animal experiment has verified that B. animalis has the effects of reducing bilirubin levels and improving liver damage. These findings collectively indicate that targeted probiotics like B. animalis may optimize current phototherapy or exchange transfusion strategies. This study has several limitations. First, the clinical component involved a relatively small sample size from a single cohort and lacked data on potential confounding factors (e.g., feeding patterns). These findings should therefore be interpreted cautiously and require validation through larger, multi-ethnic studies with more comprehensive datasets. Second, the mechanistic insights, particularly those derived from the animal model, must be interpreted within a specific context. We employed the PHZ-induced hemolytic jaundice model based on its alignment with the non-cholestatic profile of our clinical cohort and its established utility in studying hyperbilirubinemia and the enterohepatic circulation. While this model effectively supported our investigation into B. animalis and hemolysis-driven hyperbilirubinemia, it does not recapitulate neonatal jaundice arising primarily from cholestasis, hepatic immaturity, or genetic defects (e.g., in UGT1A1). Therefore, the therapeutic potential of B. animalis identified here may be most relevant to hemolysis-associated NJ and necessitates further evaluation in etiology-specific models. Furthermore, the in vitro and in vivo validation stages are considered as preliminary exploratory studies. Additional experimental validation is needed to confirm the roles of gut microbiota and associated BA metabolites. Although we have identified key gut microbial taxa and BA metabolites associated with NJ, their specific effects on NJ pathogenesis and underlying mechanisms remain to be elucidated. The complex interactions between gut phages, microbiota, BA metabolites, and NJ itself warrant further investigation to establish causal relationships. Future studies should focus on mechanistic validation of these preliminary findings through controlled intervention experiments and longitudinal analyses.In summary, we further validate the role of intestinal BA in NJ. By integrating our findings with established evidence of B. animalis BSH activity^51,52, we propose that BA modulation mediated by B. animalis, likely through its BSH function, represents a promising therapeutic target for hemolysis-associated NJ, while phage-microbiota interactions reveal novel regulatory dimensions. We demonstrate that B. animalis reshapes the BA pool, primarily reducing total BA levels while increasing DCA concentration. Furthermore, B. animalis exhibits ameliorative effects on hyperbilirubinemia and liver injury in a PHZ-induced hemolytic jaundice model. However, the precise mechanistic role of the pathways through which B. animalis regulates jaundice require further elucidation. Future research should: (1) validate the general applicability of the “phage-bacteria-bile acid” axis in expanded cohorts; (2) conduct animal studies to further investigate the specific mechanisms through which B. animalis regulates jaundice by remodeling the BA pool; and (3) employ multi-omics integration to dynamically resolve this interactive network.

Methods

Study cohort and sample collection

This study complied with the principles of the Declaration of Helsinki and was approved by the Medical Ethics Committees of Shenzhen Longgang Central Hospital (Approval No. 2024ECPJ083) and Longgang District Maternity & Child Healthcare Hospital (Approval No. LGFYKYXMLL-2024-100). Written informed consent was obtained from the parents of each neonate. Study participants were recruited from the premature infants’ gut microbiota assembly and neurodevelopment (PIGMAN) cohort⁵³ at Shenzhen Longgang Central Hospital and Shenzhen Longgang District Maternity & Child Healthcare Hospital. A total of 120 full-term neonates were enrolled, divided into discovery (n = 90) and validation (n = 30) phases. The discovery phase comprised 30 neonates with moderate-to-severe jaundice (NJ group) and 60 non-moderate/severe jaundice neonates: 30 with mild jaundice born to mothers with late-pregnancy GBS colonization (GBS group) and 30 with mild or no jaundice born to GBS-negative mothers (HC group). The validation phase included 30 neonates receiving longitudinal intervention (14 NJ, 16 GBS). Inclusion criteria required that mothers had no prenatal high-risk factors, no antibiotic use, and no fetal defecation after birth before enrollment. Exclusion criteria included confirmed congenital abnormalities after admission, maternal high-risk characteristics, recent antibiotic exposure, preterm birth, and severe neonatal illnesses. Diagnosis and severity classification of NJ followed the American Academy of Pediatrics’ Clinical Practice Guideline (CPG) for hyperbilirubinemia¹⁰ and the Chinese Medical Association Neonatology Group’s Diagnosis and Treatment Guidelines for Neonatal Hyperbilirubinemia^54,55. Bilirubin level was measured using a Beckman Coulter automatic biochemistry analyzer during their hospitalization (Beckman Coulter, CA, USA). Maternal GBS colonization status during late pregnancy was determined according to the Chinese Expert Consensus on Prevention of Perinatal Group B Streptococcal Disease⁵⁶. The selection criteria for the HC group were to randomly select neonates who were similar to the NJ group in terms of sex, gestational age at birth, and general condition at birth, and who did not meet the diagnostic criteria for neonatal hyperbilirubinemia or whose bilirubin levels in the neonatal period did not meet the criteria for phototherapy.

Researchers wearing medical sterile masks and sterile gloves collected fresh fecal samples within one hour of neonatal defecation using disinfected disposable fecal collection tubes to prevent contamination. All samples were assigned unique identifiers to ensure anonymity. Samples were processed within one hour of collection, aliquoted into 3–5 g portions per tube, and stored at −80 °C.

Human fecal metagenomic sequencing

As described, fecal samples were collected from all 120 full-term neonates enrolled in the study. Genomic DNA was extracted using the TIANGEN Magnetic Soil/Fecal DNA Kit. Extracted DNA was assessed for purity using a NanoDrop spectrophotometer, while fragment integrity and potential RNA/protein contamination were evaluated via 1% agarose gel electrophoresis. DNA concentration was precisely quantified using a Qubit 4.0 Fluorometer. Qualified DNA was mechanically sheared, followed by end repair, 3’-end adenylation, and ligation of MGI platform-specific adapters. PCR amplification was then performed to incorporate dual-index tags, constructing NDA libraries. Libraries were quantified by Qubit, and fragment size distribution (target range: 300–500 bp) was confirmed using capillary electrophoresis to exclude adapter dimer contamination or large fragments. Purified double-stranded DNA libraries were denatured and circularized into single-stranded circular templates. DNA nanoballs (DNBs) were synthesized via Rolling Circle Amplification (RCA). DNBs were then high-density loaded onto microarray chips, and paired-end 150 bp sequencing was performed on the MGI DNBSEQ platform using Combinatorial Probe-Anchor Synthesis (cPAS) technology.

Gut microbiome analysis

Raw reads were quality-filtered with fastp⁵⁷ (v0.22.0) to remove adapters and low-quality bases (Q-score <20). Host DNA was depleted by aligning reads to the GRCh37 human genome using Bowtie2⁵⁸ (v2.4.5). Taxonomic profiling was performed using MetaPhlAn (v4.0.2) with the default species-level database (CHOCOPhlAnSGB_202403 release). Functional annotation was conducted via HUMAnN3 (v3.9) with custom databases: ChocoPhlAn (chocophlan_v31_201901), UniRef90 (uniref90_v201901b), and mapping files (mapping_v201901b). Statistical analyses and visualization were implemented in R (v4.2.3) using the phyloseq and ggplot2 packages.

To assess α-diversity, we calculated the Shannon index, Chao1 richness, and Simpson evenness using the R package vegan (v2.6-4). Group-wise differences were compared using the Kruskal–Wallis test, a non-parametric alternative to one-way ANOVA suitable for non-normally distributed data. To evaluate β-diversity (community structure differences), we constructed a Bray-Curtis distance matrix and performed Principal Coordinates Analysis (PCoA) and Non-Metric Multidimensional Scaling (NMDS) for visualization using phyloseq (v1.42.0). Group separation was tested using PERMANOVA.

Based on the genus and species level data, the vegdist function in R package Vegan is used to calculate the Bray-Curtis distance. Then, based on the obtained Bray-Curtis distance, the anosim function in R package Vegan is further used to calculate the sample similarity matrix. Finally, the stat_boxplot function in ggplot2 of R package is used for comparative analysis and visualization of similarities within and between groups.

A Bray-Curtis dissimilarity matrix was constructed to quantify compositional differences between microbial communities using the vegdist() function from the R package vegan v2.6-4. To determine the optimal number of clusters, the Gap Statistic method was employed: the within-cluster dispersion of the observed data was systematically compared against 500 bootstrap reference datasets simulated under a uniform distribution, with k-means clustering iterated across candidate cluster numbers (k = 1–10). The optimal k value was identified as the inflection point of the Gap curve, where the difference between observed and expected dispersion plateaued. Subsequently, Partitioning Around Medoids (PAM) clustering was performed using the pam() function from the R package “cluster” v2.1.8.1, which partitions samples around robust medoids based on the precomputed Bray-Curtis distance matrix. The biological validity of clusters was assessed via silhouette coefficients, with values exceeding 0.5 indicating robust separation between clusters. To statistically confirm inter-cluster differences, PERMANOVA (Adonis test with 999 permutations) was applied to evaluate the significance of Bray-Curtis dissimilarity across clusters, ensuring the observed partitioning was not attributable to random variation. For dimensionality reduction and visualization, two complementary approaches were implemented: Classical Multidimensional Scaling (MDS/PCoA) using cmdscale() to project data into a 2D space while quantifying variance explained by principal coordinates, and Non-metric Multidimensional Scaling (NMDS) via the metaMDS() function (vegan package) with rigorous optimization parameters (maximum iterations = 100, stress threshold < 0.15). The final NMDS configuration with acceptable stress was selected, and cluster labels derived from PAM were overlaid onto the ordination. To emphasize group separation, 95% confidence ellipses were drawn around cluster centroids using ordiellipse(), visually reinforcing the spatial coherence of clustered samples.

To identify gut microbial taxa exhibiting significant inter-group differences, we first filtered out low-abundance features with mean relative abundances below 0.1% and features absent in more than 20% of samples. The remaining taxa were subjected to differential abundance analysis using MaAsLin2 (v1.16.0), employing total sum scaling (TSS) normalization to account for sequencing depth variability. For the primary analysis, the grouping variable was included as a fixed effect while adjusting for the following neonatal clinical covariates: birth weight, head circumference at birth, gestational age, hemoglobin, TBA, TBIL, and WBC. A generalized linear mixed model (GLMM) with negative binomial distribution was fitted to evaluate group differences. To further address potential confounding effects of other clinically relevant factors—including antibiotic exposure, delivery mode, maternal pregnancy complications, and neonatal infection status—we conducted an additional confounder evaluation using the metadeconfoundR package (v1.0.4). This analysis iteratively assessed each variable for association with both the outcome and microbial features; only taxa-phenotype associations that remained significant after controlling for all identified confounders were considered robust (“deconfounded”). Statistically significant taxa from the primary MaAsLin2 analysis were identified based on a false discovery rate (FDR)-adjusted q-value < 0.05, as implemented by the Benjamini-Hochberg procedure within the software. To ensure biological relevance, we further required an absolute log2 fold change (|log2FC|) exceeding 1.2, which corresponds to a ≥ 2.3-fold differential abundance between groups. This stringent pipeline prioritizes taxa with both robust statistical evidence and substantial ecological impact, minimizing false positives while retaining markers of functional relevance in gut microbiome dynamics.

Viral metagenomic analysis

To systematically characterize the gut virome, we implemented an integrated bioinformatics pipeline combining state-of-the-art tools for viral identification, quality control, and functional annotation. Raw metagenomic sequencing reads were first subjected to quality trimming using fastp⁵⁷ (v0.22.0) to remove low-quality bases (Q-score <20) and adapter sequences. Host-derived DNA was subsequently filtered by aligning reads to the human reference genome (GRCh37) using Bowtie2⁵⁸ (v2.4.5). Filtered reads were assembled de novo with MEGAHIT (v1.2.9) using a k-mer range of 27–127 (step size: 10). Contigs exceeding 1.0 kb were extracted using the seq module of SeqKit v2.3.0, followed by deduplication of redundant sequences at ≥90% identity via BBmap’s dedupe.sh (v38.18, minidentity = 90). VirSorter2⁵⁹ (v2.2.4) as the latest viral identifier by integrating a collection of customized automatic classifiers to improve the classification accuracy of all types of viruses, which was run with settings (min-score 0.95), alongside VIBRANT⁶⁰ (v1.2.1) analysis that preserved contigs classified as medium, high, or complete quality. Afterwards, all putative viral sequences from two viral identifiers were combined on a per-sample basis.

Thereafter, contigs were rigorously filtered through CheckV⁶¹ (v1.0.2), retaining only those with medium, high, or complete quality grades that additionally satisfied (1) with higher number of viral genes than host genes, (2) longer than 3 kbp, (3) the times the viral sequences is represented in the contig less than/equal to one (kmer_freq ≤ 1), and (4) exclusion of all contigs with quality warnings. Taxonomic classification of viral contigs was assigned via vContact2⁶² (v0.11.3) with default RefSeq prokaryotic viral database. For quantification, a salmon⁶³ (v1.4.0) index was built with default parameters (salmon index -p 8 -t $contigs -i contig_index), followed by abundance estimation using salmon quant in meta-mode (--meta). The analysis method was basically consistent with the gut microbiota analysis method.

Targeted BA metabolomics

Fecal BA profiling was performed using a targeted metabolomics approach based on ultra-high performance liquid chromatography coupled with triple quadrupole tandem mass spectrometry (UHPLC-MS/MS). Sample preparation and extraction were conducted as follows. Samples were thawed at 4 °C, and an appropriate weight was transferred to a 2 mL centrifuge tube. After adding 600 μL of ice-cold methanol, the mixture was shaken for 1 min, ground for 5 min, allowed to stand at 4 °C for 30 min, and centrifuged at 12,000 rpm for 10 min. A second extraction was performed by adding another 600 μL of ice-cold methanol to the residue, repeating the shaking, grinding, standing, and centrifugation steps. The two extracts were combined and evaporated to dryness. The residue was reconstituted in 0.2 mL of methanol containing 50 ng/mL of deuterated cholic acid (CA-D4) as the internal standard, centrifuged for 15 min, and the supernatant was collected and dried using a rotary evaporation concentrator (Hunan Herexi Instrument Equipment Co., Ltd.) prior to instrumental analysis.

Chromatographic separation was conducted on a Waters Acquity UPLC system (Waters Corporation, Milford, MA, USA) equipped with an Acquity UPLC BEH C18 column (1.7 µm, 2.1 × 100 mm). The column temperature was maintained at 40 °C, and the injection volume was 5 µL. Mobile phases consisted of water containing 0.05% formic acid (eluent A) and acetonitrile containing 0.05% formic acid (eluent B), with a flow rate of 0.3 mL/min. A 16-min gradient elution was performed as follows: 90% A/10% B for the first 1 min; B increased to 40% over 2 min, to 45% at 5 min, to 60% at 7.5 min, to 65% at 9.5 min, and to 80% at 11.5 min, maintained for 2.5 min, followed by re-equilibration to initial conditions for 2 min. The complete list of the 43 targeted BAs with their retention times, MRM transitions, and optimized mass spectrometry parameters is provided in Supplementary Table 14.

Mass spectrometry detection was performed using an AB SCIEX 5500 QQQ-MS triple quadrupole mass spectrometer (AB SCIEX, Framingham, MA, USA) equipped with an electrospray ionization (ESI) source operating in negative ion mode with multiple reaction monitoring (MRM). The optimized ion source parameters were as follows: curtain gas, 35 arb; collision gas, 9 arb; ion spray voltage, 4500 V; source temperature, 450 °C; ion source gas 1, 45 arb; ion source gas 2, 45 arb. MRM transitions, declustering potentials (DP), collision energies (CE), and cell exit potentials (CXP) for each BA were optimized using Quanoptimizer software (Waters) and are listed in Supplementary Table 14. Peak integration was performed with MultiQuant software, and quantification was achieved based on a single-point internal standard calibration method. Deuterated CA-D4 was used as the internal standard for all 43 BAs at a concentration of 50 ng/mL. The calibration coefficient (f) for each BA was calculated as f = (A_s/A_r) × (m_r/m_s), where A_s and A_r are the peak areas of the internal standard and reference standard, respectively, and m_s and m_r are their corresponding amounts. Quality control (QC) samples were analyzed in quadruplicate; for quantifiable BAs, the relative standard deviation (RSD) ranged from 2.0% to 9.8%, confirming acceptable method precision. Limits of detection (LOD, signal-to-noise ratio = 3) ranged from 0.045 to 0.887 ng/mL, and limits of quantification (LOQ, signal-to-noise ratio = 10) ranged from 0.1485 to 2.9271 ng/mL. Complete MRM parameters, retention times, calibration coefficients, LOD, and LOQ values for all 43 BAs are provided in Supplementary Table 14. Data acquisition and processing were performed using MassLynx 4.1 software (Waters).

To rigorously validate bile acids exhibiting significant inter-group differences, we first applied stringent filtering criteria: metabolites with low abundance (mean relative abundance <0.01%) or sparse prevalence (undetected in >50% of samples) were excluded to eliminate noise from sporadically observed or trace-level compounds. The remaining 43 bile acids were analyzed using both orthogonal partial least squares discriminant analysis (OPLS-DA) and MaAsLin2, with the objective of identifying metabolites contributing to group separation. For OPLS-DA, variable importance in projection (VIP) scores—derived from the first predictive component—were used to rank metabolites by discriminatory power, while MaAsLin2 assessed group effects via generalized linear models with negative binomial distribution (controlling for batch effects or technical covariates). Metabolites meeting the following criteria were selected as potential biomarkers: VIP > 1 (indicating substantial influence on classification), ANOVA p-value < 0.05 (statistical significance), and maximum fold change ≥1 (biological relevance).

BSH gene abundance

BSH gene quantification was performed using a customized protein sequence database. The database was constructed by integrating Spearman ally curated BSH sequences from literature with canonical entries retrieved from UniProt (Universal Protein) (search terms: “bile salt hydrolase” OR “choloylglycine hydrolase”, taxonomic filter: Bacteria), details are provided in the Supplementary Table 7. Sequence deduplication and formatting were implemented using the makedb module in DIAMOND (v2.0.15). Finally, metagenomic sequencing reads were aligned against this reference index using DIAMOND’s blastx module, generating a BSH gene abundance file for each metagenomic sample.

To identify significantly differentially abundant BSH genes between groups, low-abundance BSH genes (mean relative abundance < 0.01%) and low-prevalence genes (undetected in >50% of samples) were initially filtered out. Differential abundance analysis was subsequently performed using MaAsLin2. Data normalization used the default Total Sum Scaling (TSS) method. Significance thresholds were set at a P-value < 0.05 and an absolute effect size (|log2 fold change|) > 0.5.

It should be noted that BSH gene abundance, as inferred from metagenomic sequencing, reflects the genetic potential for bile salt hydrolase activity rather than directly measured enzymatic activity in clinical fecal samples. In the mouse experiment, BSH enzymatic activity was additionally measured using a commercial assay kit (see below).

Multi-omics analysis

Procrustes analysis is a method for assessing the concordance between two sets of data by comparing their shape distributions. Its principle is based on least-squares orthogonal mapping, iteratively identifying a consensus shape and determining the affine transformation for each object shape relative to this consensus using least-squares minimization. Procrustes analysis was completed using the R package vegan, with reference to a previous publication⁶⁴. First, principal component analysis (PCA) was applied separately to each dataset using the rda function for dimensionality reduction. Coordinates along the principal axes (representing linear combinations of the variable sets) were extracted for comparison. The procrustes function was then executed with parameters permutation = how and nperm = 999, yielding the sum of squared deviations statistic (M²) and a P-value derived from 999 permutation tests. Results were visualized using the ggplot2 package in R.

To assess the association between bile acid metabolites and intestinal microbial characteristics (such as Shannon diversity), we conducted a Spearman partial correlation analysis following the method described in previous literature⁶⁵. The partial correlation coefficient was calculated using the pcor.test function in the R package ppcor (v1.0), and birth weight and gestational age were included as covariates. To control for multiple testing, we performed FDR correction using the Benjamini-Hochberg method from the R statistical package (v4.2.3). After FDR correction, correlations with q-values less than 0.05 were considered statistically significant. In visualization, the significant correlations (coefficients positive or negative) with FDR-corrected q-values less than 0.05 were represented as positive or negative correlations, respectively.

Mantel tests were performed using the Mantel_test function from the R package linkET (v0.0.7.4) with default parameters to assess correlations between species-by-environment matrices (based on all or well-grouped variables). Subsequently, correlation coefficient matrices among environmental variables were computed using relevant functions within linkET. Finally, correlations between the matrices were visualized using the qcorrplot function from the same package. It should be noted that no correction for multiple comparisons was initially applied to the p-values derived from the Mantel tests.

For canonical correspondence analysis (CCA), the data at genus level were selected as species composition, and the composition of metabolite markers was selected as environmental composition. The cca function in the R package Vegan was used for CCA to obtain restrictive ranking relationships for species and environmental data. Then, the anova function in the R package Stats (v4.2.3) was used to perform Permutation Multivariate Analysis of Variance (PERMANOVA), calculating p values and R-squared values (R²) for restrictive rankings. Finally, after taking the first two axes and removing the largest deviations, the restrictive ranking relationships were visualized and displayed through the geom_point function in ggplot2 of R package.

MFA was performed using the FactoMineR package in R. The three data tables (bacterial, phage, and BA profiles) were integrated with group weighting. Initial factors were extracted via Principal Component Analysis (PCA), and factor loading matrices were optimized using maximum likelihood estimation. Following Varimax orthogonal rotation, significant common factors were retained based on an eigenvalue threshold. Model reliability and validity were assessed through cross-validation and inter-group consistency testing (RV coefficient matrix) within FactoMineR. Results were visualized using biplots.

To investigate whether altered BA abundance correlates with changes in the gut microbiota, CoIA was employed to identify co-variation between fecal BA metabolites and metagenomic profiles. It was performed on fecal BA and metagenomic gene clusters using the R packages ade4 and omicade4, and the Spearman rank correlation coefficient was used to assess the correlation between the metabolome and metagenome corrected for multiple testing using the R package psych.

For SEM, please refer to the previously published article²⁶. First, based on either gut microbiota genus level data (profile) or gut bile acid data (profile), the vegdist and pcoa functions in the R package Vegan (v2.6-4) were used to count the Bray–Curtis distances and perform principal coordinates analysis (PCoA), respectively, and the resulting first axis (PC1) was used to represent the microbial community or bile acid betadiversity. Second, for the data of PC1 and the variables to be studied, the scale in the R package base (v4.2.3) was used to standardize the data, and a set of structural equations of “Y ~ X” was constructed. The sem function in the R package lavaan (0.6-16) was used to fit the model. The parameter estimates function in the R package lavaan was used to obtain the model’s partial regression coefficients (estimate), the execution intervals, and the p value. The fitMeasures function in the R package lavaan was used to assess the reliability of the model parameters, with p > 0.05 implying the reliability of the model, i.e., the model predictions were not significantly different from the actual observations. Finally, the semPaths function in the R package semPlot (v1.1.6) provided an alternate visual presentation of the structural equation model.

Other analysis performed using methods similar to those previously published²⁶. In order to evaluate the correlation between gut microbiota composition or BA with statistically significant differences between groups and clinical manifestations, the lm function in R software was used to construct a logistic regression model, and the p value and determination coefficient (R-squared) of the logistic regression model were obtained by summing the function. Boxplots and scatter plots were made using the R package beeswarm, and p values for statistically significant differences were obtained by referring to wilcox. Species diversity analysis (Shannon index) was constructed using personalized scripts. The correlation heatmap was completed by the labeledHeatmap function in the R package WGCNA, and the significance was set as the absolute value of Spearman correlation coefficient greater than 0.3 and p value less than 0.05.

B. animalis drives BA pool remodeling

Fecal samples from the HC group were pooled, serially diluted, and plated on Basic Salt Medium (BSM) agar (composition: 10.0 g/L peptone, 8.0 g/L beef extract, 4.0 g/L yeast extract, 20.0 g/L glucose, 2.0 g/L K₂HPO₄, 2.0 g/L ammonium citrate, 5.0 g/L sodium acetate, 0.2 g/L MgSO₄, 0.04 g/L MnSO₄, 1.0 g/L Tween 80, 0.5 g/L L-cysteine hydrochloride, and 50 mg/L mupirocin; pH 5.7 ± 0.2). Quadrant streaking under anaerobic conditions (85% N₂, 10% CO₂, 5% H₂) at 37 °C for 72 h. Following three successive rounds of purification, a single colony was obtained and cryopreserved in glycerol and ceramic bead cryotubes at −80 °C. For bacterial identification, cells were harvested and lysed using a modified phenol-chloroform method. Briefly, 100 μL of bacterial culture was treated with lysozyme (50 mg/mL) at 37 °C for 30 min, followed by proteinase K (10 mg/mL) and SDS (1% final concentration) at 55 °C for 2 h. Genomic DNA was extracted with phenol:chloroform:isoamyl alcohol (25:24:1), precipitated with ice-cold isopropanol, washed with 70% ethanol, and resuspended in TE buffer. DNA quality was assessed by 1% agarose gel electrophoresis using a DL9000 marker. The full-length 16S rRNA gene (V1–V9 regions) was amplified using primers 8F (5′-AGAGTTTGATCCTGGCTCAG-3′) and 1492R (5′-TACGGYTACCTTGTTAYGACTT-3′). PCR reactions (50 μL) contained Phusion High-Fidelity DNA Polymerase (NEB), 10 μM primers, and 5–50 ng template DNA. Thermocycling conditions were as follows: initial denaturation at 94 °C for 3 min; 27 cycles of 94 °C for 30 s, 55 °C for 30 s, and 72 °C for 1.5 min; and a final extension at 72 °C for 5 min. Amplification products were verified by 1% agarose gel electrophoresis with a DL2000 marker. Purified PCR products were sequenced bidirectionally on an ABI 3730xl platform (Applied Biosystems). Raw sequences were quality-filtered using Seqtk v1.2 and assembled with MOTHUR v1.39.5. The consensus sequence was aligned against the NCBI 16S rRNA database using BLASTn (excluding uncultured and environmental sequences), confirming 100% identity to B. animalis subsp. lactis strain YIT 4121 (accession NR_040867.1). Phylogenetic analysis was performed using MUSCLE v3.28.0 for sequence alignment and MEGA v10.1.8 with the neighbor-joining method (1000 bootstrap replicates). The isolated strain was designated B. animalis subsp. lactis Y103-OTU5.

The previously isolated and purified B. animalis subsp. lactis Y103-OTU5 underwent resuscitation and cultivation. The strain was anaerobically cultured in BSM medium (85% N₂, 10% CO₂, 5% H₂) at 37 °C for 72 h. Following incubation, the culture was centrifuged (10,000 × g, 10 min, 4 °C) to pellet bacterial cells, and the supernatant was collected. Sterile BSM medium (uninoculated) was processed identically as a blank control. To confirm strain identity, post-culture bacterial cells underwent genomic DNA extraction, followed by PCR amplification and sequencing of the 16S rRNA V1–V9 hypervariable regions. BAs in both bacterial supernatants and blank controls were analyzed using the same LC-MS/MS method described in “Targeted BA metabolomics” above, including the identical 43-BA panel, MRM parameters, internal standard (CA-D4, 50 ng/mL), and QC procedures (see Supplementary Table 14 for complete parameters and LOD/LOQ values). Briefly, supernatants were filtered (0.22 μm), diluted 1:10 in methanol, and spiked with CA-D4. Three technical replicates were analyzed per sample. BA concentrations were normalized to blank controls to exclude background interference. Fold changes were calculated to assess bacterial metabolic activity. Statistical significance (p < 0.05) was determined using Student’s t-test.

Animals

BALB/c mice (approximately 21 days old, half male and half female) were commercially sourced from Beijing Beiyo Biotechnology Co., Ltd., China. Mice were fed standard chow in a temperature-controlled room (22 °C) with a humidity of 55% and a 12-h light and dark cycle. The experiment was conducted after 7 days of adaptive feeding. Clean standard feed was provided throughout. With reference to previous studies^66–68 on phenylhydrazine hydrochloride (PHZ)-induced hyperbilirubinemia via intraperitoneal injection, BALB/c mice were randomly divided into 3 groups: Control, PHZ model (PHZ), and B. animalis subsp. lactis Y103-OTU5 intervention (B. animalis_ PHZ). For three consecutive weeks before and after model induction, mice received daily oral gavage of 200 μL B. animalis suspension (5 × 10⁹ CFU/mL). On the final 2 days of gavage, PHZ (50 mg/kg/day) was administered intraperitoneally. The control group received an equal volume of sterile saline by gavage. At the end of the experiment, mice were humanely euthanized 48 h after the last PHZ injection. Anesthesia was first induced using inhaled 1–4% isoflurane. Following the establishment of deep anesthesia (characterized by the loss of the pedal withdrawal reflex and steady, slowed respiration), euthanasia was performed by exsanguination to ensure the rapid and complete cessation of vital functions. Fecal, plasma, liver, and colon samples were collected for subsequent analysis. Throughout the procedure, all operations were conducted in accordance with the principle of minimizing animal suffering. All animal experimental procedures were performed in accordance with the guidelines of the National Institutes of Health for the use of laboratory animals and were approved by the Institutional Animal Care and Use Committee (IACUC) (Approval No. B202505-40).

Mouse fecal metagenomic sequencing and BA profiling

Methods for metagenomic DNA extraction, library preparation, and sequencing from mouse fecal samples were identical to those used for human samples, as detailed in the section “Human fecal metagenomic sequencing”. Similarly, targeted BAs metabolomics on mouse feces was performed as described for human sample.

Measurement of BSH activity in mouse feces

The activity of BSH in fecal samples was determined using a commercial assay kit (Nanjing Mofan Biotechnology Co., Ltd., China). Briefly, weighed fecal samples were homogenized on ice with pre-cooled extraction buffer and then centrifuged at 4 °C for 10 min. The resulting supernatant was collected for analysis. In this assay, BSH catalyzes the hydrolysis of conjugated bile acids, releasing free amino acids. These amino acids then react with ninhydrin reagent in a boiling water bath to form a blue-purple product with a characteristic absorption peak at 570 nm. The enzymatic reaction was conducted at 37 °C for 30 min. A control tube was set up in parallel to account for any background interference. A standard curve was generated simultaneously using known concentrations of glycine for quantitative calculation. One unit of BSH activity was defined as the amount of enzyme required to produce 1 μmol of free amino acid per minute per gram of fecal sample.

Mouse liver transcriptome sequencing

Total RNA was isolated from mouse liver tissues using QIAzol® Reagent with DNase I digestion, followed by integrity verification (RIN ≥ 8.0, Agilent 2100 Bioanalyzer). Strand-specific libraries were constructed through mRNA enrichment with Oligo dT beads, fragmentation, and cDNA synthesis using random hexamers for first-strand generation and dUTP incorporation for second-strand orientation. After end repair, A-tailing, adapter ligation, and PCR amplification, library quality was assessed via Qubit™ quantification and size distribution analysis. Paired-end sequencing (150 bp, 40 million reads/sample) was performed on the Illumina NovaSeq 6000 platform using Sequencing-by-Synthesis technology.

Raw sequencing data underwent quality control with fastp to remove adapter-contaminated reads, poly-N sequences, and low-quality bases (Q20/Q30 thresholds). Clean reads were aligned to the Mus musculus reference genome (GRCm38) using HISAT2, followed by gene-level quantification via featureCounts with FPKM normalization. Differential expression analysis was conducted using DESeq2 for groups with biological replicates, applying Benjamini-Hochberg false discovery rate correction (significance thresholds: |log₂FC| ≥ 1, padj ≤ 0.05). Functional interpretation included GO/KEGG/Reactome enrichment of differentially expressed genes (DEGs) through clusterProfiler and pathway-level characterization via Gene Set Enrichment Analysis (GSEA).

Biochemical parameter measurement

Freshly collected whole venous blood samples were allowed to clot at room temperature for 30 minutes in the dark to prevent bilirubin photo-oxidation, and then centrifuged at 3000 rpm for 20 minutes. The serum supernatants were carefully separated and collected. All serum biochemical parameters, including alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), TBA, TBIL, and DBIL, were measured using a fully automated biochemical analyzer (BS-240 VET, Shenzhen Mindray Animal Medical Technology Co., Ltd., China) in a single run. All assays were performed with Mindray original reagents (Shenzhen Mindray Animal Medical Technology Co., Ltd., China) strictly according to the manufacturer’s standard operating procedures.

Histological assessment of liver

The liver tissues were fixed in 10% neutral buffered formalin for 48 h. After dehydration, the tissues were embedded in paraffin and sliced (5μm). The sections were dewaxed in xylene and rehydrated using ethanol. Hematoxylin and eosin (H&E) and Masson staining were performed for histological examination.

Quantitative real-time PCR

Total RNA was isolated from mouse colon tissues using TRIpure Reagent (Aidlab, China) according to the manufacturer’s instructions. The concentration and purity of the extracted RNA were verified spectrophotometrically. First-strand cDNA was synthesized from total RNA using the Hifair® II 1st Strand cDNA Synthesis SuperMix (gDNA digester plus) (Yeasen Biotechnology, China). Quantitative real-time PCR was performed using Hieff UNICON® qPCR SYBR Green Master Mix (Yeasen Biotechnology, China) on a BIOER FQD-48A(A4) real-time PCR system. The thermal cycling protocol was as follows: initial denaturation at 95 °C for 5 min, followed by 40 cycles of 95 °C for 10 s and 60 °C for 30 s. The primer sequences for the target genes (Tgr5, Fxr) and the reference gene (Gapdh) are listed in Supplementary Table 10. The relative mRNA expression levels were normalized to Gapdh and calculated using the comparative 2^–ΔΔCt method.

Statistical analysis

All statistical analyses were performed in R (v4.2.3). Group comparisons for continuous variables utilized the Mann–Whitney U test (two groups) or Kruskal–Wallis test (≥ three groups) for non-normally distributed data. Visualizations were generated using ggplot2 (v3.5.2), with significance denoted as follows: *p < 0.05, **p < 0.01, ***p < 0.001. Continuous data are presented as median (interquartile range) or mean ± standard deviation, as appropriate, while categorical data are reported as count and proportion (%).

Author contributions

Xianhong Chen: Data curation, investigation, formal analysis, visualization, writing—original draft, and writing—review & editing. Cheng Chen: Data curation, investigation, formal analysis, visualization, and writing—original draft. Peng Zhang: Data curation, investigation, formal analysis, visualization, and writing—original draft. Xiaohong OuYang: Data curation, investigation, and formal analysis. Haiyan Ma: Methodology. Wanling Chen: Methodology and investigation. Tingting Li: Methodology and investigation. Jing Han: Investigation. Yanli Wang: Investigation. Huijuan Wang: Investigation. Qiujing Zhou: Investigation. Guoqiang Cheng: Investigation. Wei Zhou: Investigation. Zhangbin Yu: Investigation. Wenhao Zhou: Supervision, project administration, and writing—review & editing. Mingbang Wang: Conceptualization, supervision, funding acquisition, project administration, resources, and writing—review & editing. Shujuan Zeng: Conceptualization, supervision, funding acquisition, project administration, resources, and writing—review & editing.

Data availability

The data that support the findings of this study have been deposited into CNSA with accession number CNP0007919.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41522-026-01057-w.