Identification of functional bacterial-viral pneumotypes associated with airway inflammation and all-cause mortality in critically ill patients

Abstract

Alterations in lung virome and bacteriome compositions have been associated with hospital-acquired pneumonia in critically ill patients, but data on lung microbiome functions underlying progression to severe complications remain unclear. Here, we analyse 184 endotracheal aspirates (ETAs) collected in 94 intubated critically ill patients enrolled in a clinical trial (NCT04793568). Combining viral-metagenomics with metatranscriptomics analyses, we identify a high-risk functional pneumotype observed before the onset of hospital-acquired pneumonia and associated with lung inflammation and mortality. This unfavourable pneumotype is characterised by enrichment in virulent Streptococcus-bacteriophages and temperate Klebsiella-bacteriophages, decreased transcriptional activity of commensal taxa such as Streptococcus and Alloprevotella, and increased transcriptional activity of Klebsiella pneumoniae. Machine-learning models define two-viral- and four-bacterial-factor signatures predicting unfavourable pneumotype (AUC 0.8 and 0.7, respectively). Validation in an independent cohort (NCT02003196) of 117 ICU patients (239 ETAs) demonstrates the robustness of the association of these signatures with the risk of mortality. Causal inference identifies virulent bacteriophages associated with Streptococcus and host genes including IL18R1, ADM and PTX3, as key regulators of severe lung dysbiosis linked to mortality in critically ill patients.

Here, the authors identify high-risk viral-bacterial functional pneumotypes associated with all-cause mortality in critically ill patients, which are characterized by expansion of virulent bacteriophages, depletion of health-associated bacteria, and lung inflammation.

Introduction

Critically ill patients are particularly vulnerable to hospital-acquired pneumonia (HAP), which is one of the most common causes of nosocomial infections in the intensive care units (ICU)¹. Despite advances in guideline-based management and antibiotic optimization, HAP remains associated with a high risk of mortality and treatment failure, with ~30% of cases progressing to acute respiratory distress syndrome (ARDS) and mortality rates reaching up to 40%^2–5. Recent data suggest that, in addition to the presence of a pathogen, the disruption of healthy host-respiratory microbiome interactions plays a central role in HAP severity and inter-individual heterogeneity in the response to treatment^6,7. Several studies have identified an association between lung microbiome dysbiosis and respiratory complications. Notably, the reduction in the healthy lung microbiome core (i.e., the microbial taxa typically found in healthy lung samples) was associated with HAP and acute respiratory failure⁸. This “healthy” bacterial core likely contributes to lung homeostasis^8,9. Using viral metagenomics, we have recently demonstrated significant shifts in the lung viral community composition of intubated critically ill patients, suggesting a contribution of the lung virome to HAP pathogenesis and treatment outcomes¹⁰. Historically, pneumotypes have been defined as conserved combinations of microorganisms based on bacterial genus-level profiles from 16S rRNA data and a pneumotype enriched in supraglottic-associated taxa was linked to heightened subclinical lung inflammation in asymptomatic individuals¹¹. More recently, a pneumotype enriched in oral-associated taxa at the time of diagnosis was associated with successful pneumonia therapy in patients with HAP and CAP¹². In this work, using clinical data from our previous clinical trial¹³, we extend this framework by defining functional pneumotypes, integrating the active bacteriome and the virome to capture the microbiome heterogeneity among critically ill patients before HAP diagnosis. Two functional pneumotypes were identified as distinct meta-clusters through integrated analysis of viral metagenomics and metatranscriptomic data from endotracheal aspirates (ETAs) and associated with clinical outcomes.

Results

Patients characteristics and multi-omic sequencing approach

We analysed ETAs from ICU patients enrolled in the multicentre, placebo-controlled, randomised clinical trial (PREV-HAP study) testing interferon-gamma-1b (IFN-γ) for the prevention of HAP (inclusion across 11 European centres, April-October 2021). Adults (18–80 years), under invasive mechanical ventilation, with one or more acute organ failure were included within the first 48 h of ICU hospitalisation, randomised to interferon-γ−1b or placebo (100 microg every 48 h for 9 days), and followed clinically for 3 months. ETAs were collected at study enrollment day 0 (immediately before the first treatment injection) and at days 3 and 7.

Based on sample availability, we analysed 184 ETAs from 94 randomised patients enrolled. The integrated dataset comprises viral taxa and functions from metatranscriptomics, as well as bacterial taxa, functions, and predicted metabolic pathways. It also includes viral composition and functions from viral metagenomics (Fig. 1, Supplementary Data 1–8 and Supplementary Fig. 1–3). Samples included in downstream metatranscriptomic and viral metagenomic analyses were selected based on volume sample availability and sequencing quality criteria. Hence, among the 184 ETA samples collected in the PREV-HAP cohort (Fig. 1), 122 had sufficient volume for metatranscriptomic analysis. Of these, 116/122 passed quality control based on the detection of ≥10 MS2 phage reads, an internal control threshold (Supplementary Fig. 4), and were interpretable for virome and bacteriome analyses; 115 were also suitable for host-response analysis (threshold of 5 Log10 of human coding reads). For viral metagenomics, 133/184 samples passed sequencing quality control and were retained for downstream analysis (Supplementary Fig. 4).

Fig. 1. Flowchart representation of sample handling and analytical workflow using metatranscriptomics and viral metagenomics approaches.

More information about sample sequencing and filtering in each omic is provided in Supplementary Fig 4.

Clinical characteristics did not differ between patients with and without HAP (Supplementary Table S1). The causes of ICU admission were 48% trauma, 37% surgical, and 15% medical. HAP developed in 45 patients (48%) at a median of 4 days after inclusion (interquartile range [IQR] 3–6). Out of the 94 patients, 22 (23%) died by day 90. During ICU stay, 74% received corticosteroids, 97% antibiotics and 49% interferon-γ−1b (Supplementary Table S1). Of the patients who developed HAP, 26% had received corticosteroids, 53% antibiotics, and 56% interferon-γ−1b prior to their HAP diagnosis (Supplementary Tables S2–S3 and Supplementary Fig. 4–6). While antibiotics had a major effect on the lung microbiome (Supplementary Fig. 7), interferon and corticosteroid treatments were found to have very limited effect on lung microbiome functions (Supplementary Fig. 8–11). All treatments were added as independent covariates in the subsequent host, bacterial and viral signature analyses.

Prior to HAP cluster derivation using metatranscriptomic and viral metagenomic data

To investigate the heterogeneity in the lung microbiome functions underlying progression to HAP, we selected 64 ETA samples collected from 39 patients before HAP diagnosis (defined as redefined Day 0, with earlier samples indexed relative to this reference point; Supplementary Fig. 4, 12). To identify distinct lung meta-clusters (so-called functional pneumotypes), we applied a three-step unsupervised analytical framework. First, we used MEFISTO (Method for the Functional Integration of Spatial and Temporal Omics Data) on viral metagenomics and metatranscriptomics to reduce the dimensionality of the selected features over time relative to HAP onset. Core microbiome (viral and bacterial taxa and their associated functions) found in at least 50% of all samples (n = 64) collected before HAP onset were retained (Supplementary Fig 13). Second, unsupervised hierarchical clustering was subsequently applied to the first two resulting MEFISTO factors capturing the most variance to define discrete community structures. Model fit metrics indicated that a two-cluster solution (k = 2) best represented the data across the core features (Fig. 2a, Supplementary Fig. 13–15 and Supplementary Data 9–10). Third, because patients contributed multiple longitudinal samples before HAP onset, cluster assignment was determined by the cluster of the sample collected closest to HAP onset. Patients generally maintained consistent cluster assignments across their longitudinal samples, with only 4/17 transitions between days 0–3 and 2/7 transitions between days 3–7 (Supplementary Fig. 16). Three transitions were from pneumotype 2 to 1, and three from 1 to 2. This series of analyses demonstrated that two main functional pneumotypes of virome-microbiome status can be observed in ICU patients who later developed HAP. However, non-HAP patients did not exhibit distinct or shared microbial community structures that could be classified as pneumotypes (Supplementary information).

Fig. 2. Identification of high-risk and inflammatory pre-HAP functional pneumotype prior to HAP onset.

a pre-HAP clusters identified using unsupervised clustering of MEFISTO factors (n = 64 samples). The x- and y-axes present factors 1 and 2. b Kaplan-Meier curves for survival probability between pre-HAP pneumotypes (log-rank p < 0.05) (n = 39 samples, one per patient). c Forest plot showing the relative risk for 90-day mortality between clinical variables stratified by pneumotype (Global log-rank p = 0.009) (n = 39 samples, one per patient). Squares indicate the estimated relative risk, and horizontal lines indicate the corresponding 95% confidence intervals (CIs). d Volcano plot showing the differentially gene expression in each pre-HAP pneumotype (n = 38 samples). Differential expression was assessed using the edgeR quasi-likelihood F-test (two-sided test). Genes with raw p < 0.05 were considered significant, without adjustment for multiple comparisons. e Gene set enrichment analysis of the upregulated differentially expressed genes in pneumotypes (All 15 DEGs were selected) (n = 38 samples). GO enrichment was assessed using a one-sided hypergeometric over-representation test. p values were adjusted for multiple comparisons using the Benjamini-Hochberg method, with an adjusted p value threshold of 0.05.

To investigate the clinical significance of the identified pre-HAP pneumotypes, we compared the mortality rates between the two pneumotypes. No major difference in baseline clinical characteristics was observed between the two pneumotypes (Supplementary Table S2). However, Kaplan-Meier survival analyses using study enrollment as the starting point showed different survival probabilities for pneumotypes during ICU follow-up (Fig. 2b and Supplementary Fig. 17). Cox proportional hazards models were then employed to quantify these disparities as relative risks over time. Compared with pneumotype 2, pneumotype 1 was associated with a lower hazard of all-cause mortality (hazard ratio [HR] = 0.2, 95% CI [0.04–0.58], log-rank p = 0.00153, Fig. 2b) and ahigher probability of successful early extubation (HR = 3.39, 95% CI [1.34–8.56], p = 0.01, Supplementary Fig. 17). Association of pneumotype 2 with all-cause mortality remained statistically significant after multivariate analysis taking into account major baseline risk factors of death (RR 9.60, 95%CI 1.14–80.8, Fig. 2c).

To investigate differences in the human immune response across pneumotypes, we performed a negative binomial generalized linear model analysis (NB-GLM) (Supplementary Data 11). After adjusting for interferon-γ−1b, corticosteroid and antibiotic treatment, HAP onset day, patient identity, and duration of mechanical ventilation, we identified 15 differentially expressed human genes (DEGs) (Fig. 2d and Supplementary Table S4). Using gene set enrichment analysis of these 15 DEGs, pneumotype 2 exhibited elevated expression of pathways associated with immune and inflammatory responses, including positive regulation of tumor necrosis factor superfamily cytokine production and of T-helper 1 type immune response and response to interleukin-18. In contrast, pneumotype 1 was associated with response to prostaglandin, sensory perception of temperature stimulus and detection of stimulus involved in sensory perception of pain (Fig. 2e and Supplementary Data 12). In silico deconvolution analysis showed no significant differences in estimated cellular composition across pneumotypes (Supplementary Fig. 18).

These findings showed that the investigation of meta-clusters including of viral and bacterial composition and transcriptomic activity enables the identification of two functional pneumotypes associated with robust clinical outcomes in critically ill patients.

Transcriptional bacterial signature and score in proinflammatory high-risk pneumotype

To determine how microbiome composition drove differences between the pneumotypes at ICU hospitalisation, we compared transcriptionally active bacteria, gene expression, functional pathways and diversity between groups using metatranscriptomic data. Patients with the pneumotype 1 (hereafter referred to as the low-risk pneumotype) exhibited greater β-diversity of transcriptionally active taxa across patients (Wilcoxon test, p = 4.88 ×  10⁻¹¹; Fig. 3a, b, Supplementary Data 13 and Supplementary Fig. 19) than those with pneumotype 2 (high-risk pneumotype). This profile was also shown when analyzing the lung mycobiome in pneumotypes where fungi were more abundant in the low-risk pneumotype (Supplementary Fig. 20). The relative transcriptional activity of taxa constituting the healthy lung core microbiome previously described as a proxy for a healthy respiratory microbiome, including Veillonella, Streptococcus, Prevotella, Haemophilus, and Fusobacterium⁸, was significantly reduced in the high-risk pneumotype 2 compared to the low-risk group (34% vs. 52%, respectively; Wilcoxon test, p = 0.04; Fig. 3c and Supplementary Fig. 21). This trend was shown also in the RPKM-normalized abundance profile (Wilcoxon test, p = 0.01; Supplementary Fig. 21). We next identified a panel of 22 differentially transcriptionally active bacterial species between low- and high-risk pneumotypes (Fig. 3d and Supplementary Table S5). Bacteria with higher transcriptional activity in the high-risk pneumotype included Klebsiella pneumoniae, Prevotella corporis, Paracoccus_unclassified and Haemophilus_unclassified, while Alloprevotella_unclassified, Prevotella_unclassified and Streptococcus_unclassified were more transcriptionally active in the low risk pneumotype. Notably, Klebsiella, Haemophilus, Staphylococcus and Streptococcus were among the bacteria that were consistently identified by culture at HAP onset with high concordance between culture-based and sequencing-based detection (Supplementary Table S2).

Fig. 3. Characterization of the transcriptionally active bacteria in functional pneumotypes.

a boxplots showing the median of beta diversity (Bray-Curtis dissimilarity) between pneumotypes (n = 42 samples). Data are presented through box plots, showcasing the 25th to 75th percentiles along with the median, while the whiskers extend to the minimum and maximum observed values. Statistical significance (p = 4.881e-11) was calculated by the two-sided Wilcoxon test (**** for p < 0.0001). b PCoA of Bray-Curtis distance between bacterial species characterizing the clustering between samples of pneumotypes (n = 42 samples). We assessed differences in community composition between groups using PERMANOVA test. c Relative abundance of the healthy lung core metatranscriptome core (Veillonella, Streptococcus, Prevotella, Haemophilus and Fusobacterium) in pneumotypes (two-sided Wilcoxon test, p = 0.04) (n = 42 samples). d Top features of the differentially transcribed (DT) bacteria at species level in pneumotypes. Statistical significance for each species was assessed using a non-parametric two-sided Wilcoxon test. Fold changes were presented as the difference (delta) in mean relative abundance between pneumotypes (n = 42 samples). e Heatmap showing the expression of bacterial proteins across functional pneumotypes (n = 42 samples). f Generalized linear model (GLM) represents the bacterial differentially transcriptional functions in pneumotypes. A function-wise logistic regression model was employed to assess the association between function expression levels and pneumotype membership using two-sided Wald test. Functions with p < 0.05 were considered statistically significant, and the direction of association was determined by the sign of the regression coefficient. In addition, a permutational multivariate analysis of variance (PERMANOVA) based on Bray-Curtis dissimilarities was performed (p = 0.017 and R2 = 0.08) (n = 42 samples). g FFTree-based staging scheme using differentially transcribed bacteria signature to predict the high risk pneumotype (n = 42 samples). h smooth ROC of the high risk pneumotype classification applying a tenfold nested cross-validation scheme with ten repetitions (n = 42 samples). The black line represents the mean ROC curve, and the shaded area represents 95% confidence interval, grey lines represent ROC curves from individual folds. i, Kaplan-Meier curves for survival probability between predicted pneumotypes for all patients (n = 42 samples). The survival differences between clusters were assessed using the Kaplan-Meier method and compared using a two-sided log-rank test. Hazard ratios and 95% confidence intervals were estimated using a Cox proportional hazards regression model. 17/22 patients have contributed for more one sample in (g–i).

The overall expression of bacterial protein-coding genes was higher in the low-risk pneumotype (Wilcoxon test, p < 0.00001; Fig. 3e and Supplementary Fig. 22), which showed a stronger COG (Clusters of Orthologous Groups) functional profile. Eight differentially expressed bacterial functions were significantly enriched in the low-risk pneumotype, including cell division, defense mechanisms and coenzyme transport and metabolism (multivariate permutation test, p = 0.017, R² = 0.08; Fig. 3f). These functions were observed more frequently in Streptococcus pneumoniae, Alloprevotella_unclassified, Prevotella_unclassified, Streptococcus_unclassified, Haemophilus_unclassified, Parvimonas_unclassified, and Enterobacteriaceae_unclassified (Supplementary Fig. 23).

Similarly, a pathway-level analysis revealed increased transcription of metabolic pathways in the low-risk group, with 25 upregulated pathways versus ten in the high-risk group (Fisher’s exact test, p = 0.0007; Supplementary Table S6 and Supplementary Fig. 24).

Finally, to create a simple and interpretable predictive tool for pneumotype classification, we applied a stacked machine learning approach that combined spectral clustering with a fast-and-frugal decision tree (FFT) model on before HAP samples. The resulting four-factor decision tree robustly discriminated high-risk from low-risk pneumotypes (accuracy= 80%, sensitivity = 100%, specificity = 70%; Fig. 3g). The high-risk pneumotype was defined by lower relative transcriptional activity of Alloprevotella_unclassified (<0.23%) and Prevotellaceae_unclassified (≤0.29%), as well as higher relative expression of Paracoccus_unclassified (>2.21%) and Klebsiella pneumoniae (>2.31%). The model’s predictive accuracy, assessed via tenfold cross-validation, achieved a mean area under the curve (AUC) of 0.8 (Fig. 3h). To assess the clinical utility of such biomarkers in patients regardless of future HAP status, we used the entire cohort, selecting all samples within 0 to 4 days after study enrollment. Kaplan-Meier analyses showed reduced survival probabilities in individuals classified by the XGBoost model as high-risk pneumotype based on the four bacterial features (Fig. 3i and Supplementary Fig. 25). Cox proportional hazards models further quantified this association, showing that the high-risk pneumotype was associated with an increased mortality risk compared with the low-risk group (HR = 3.9, 95% CI: 1.1–14, p = 0.025). Taken together, this series of analyses demonstrated that the pre-HAP high-risk pneumotype was characterized by a reduction in the transcriptional activity of the healthy core microbiome, with several pathobionts showing high transcriptional activity.

Bacteriophage patterns and score in pre-HAP high-risk patients

We next evaluated the virome composition and function across pneumotypes using viral metagenomic. Bacteriophages were found to be the most abundant viral entities in the two pneumotypes, followed by giant viruses (Fig. 4a, Supplementary Fig. 26 and Supplementary Table S7). Patients classified as high-risk pneumotype exhibited significantly lower Bray-Curtis dissimilarity than those assigned to the low-risk pneumotype (Wilcoxon test, p = 4.73×10⁻¹¹; Fig. 4b, c, Supplementary Data 14 and Supplementary Fig. 19). We subsequently identified 69 bacteriophage genera with significant differences in relative abundance between the low- and high-risk pneumotypes (Fig. 4d and Supplementary Table S8). Functional profiling of the virome revealed distinct viral activity patterns across pneumotypes. Notably, a substantial proportion (61%) of the bacteriophages that were more abundant in the high-risk pneumotype were predicted to display a virulent lifestyle (Supplementary Table S8). Using bacteriophage metagenomics data, we observed that ten viral functions assigned were significantly abundant in the high-risk pneumotype, the top ones being the transport of chemicals molecules/nutrients, and the viral adhesion and invasion (GLM with multivariate permutation test, p = 0.005, R² = 0.07; Fig. 4e). In addition, we identified using bacteriophage metatranscriptomics data, that the macromolecule modification, regulation of gene expression and the host-related functions (stress-associated response) were more transcriptionally expressed in high risk pneumotype (p = 0.001, R² = 0.14; Fig. 4f).

Fig. 4. Characterization of the viral signatures in functional pneumotypes.

a Bubble plot showing the composition of the whole virome in functional pneumotypes (n = 44 samples). The size of the bubbles in the plot represents the RPKM-normalized abundance (in log₁₀RPKM) mediane of the entire virome in each pneumotype. Circle area is proportional to viral type mediane abundance (log10 RPKM), with reference circle sizes representing log10(RPKM) values ranging from 2.9 (smallest circle) to 5.4 (largest circle). b boxplots showing the median of beta diversity (Bray-Curtis dissimilarity) between pneumotypes (p = 4.734e-11). Data are presented through box plots, showcasing the 25th to 75th percentiles along with the median, while the whiskers extend to the minimum and maximum observed values. Statistical significance was calculated by the two-sided Wilcoxon test (**** for p < 0.0001) (n = 44 samples). c PCoA of Bray-Curtis distance between viral species characterizing the clustering between samples of pneumotypes (n = 44 samples). The differences in community composition between groups were assessed using PERMANOVA test, based on Bray-Curtis dissimilarities and 9,999 permutations. d, Top features of the differentially abundant (DA) bacteriophage at genus level in pneumotypes. Statistical significance for each species was assessed using a non-parametric two-sided Wilcoxon test. Fold changes were presented as the difference (delta) in mean relative abundance between pneumotypes (n = 44 samples). e, f Generalized linear model (GLM) represents the differentially abundant (e, n = 44 samples) and differentially expressed (f, n = 42 samples) functions in pneumotypes. A function-wise logistic regression model was employed to assess the association between function expression levels and pneumotype membership using two-sided Wald test. Functions with p < 0.05 were considered statistically significant, and the direction of association was determined by the sign of the regression coefficient. *, ** and *** for p < 0.05, p < 0.01 and p < 0.001, respectively. Besides, Permanova based on Bray-Curtis dissimilarities was computed (p = 0.005-R2 = 0.07 in e and p = 0.001-R2 = 0.14 in (f). g FFTree-based staging scheme using differentially abundant bacteriophage signature to predict the high risk pneumotype (n = 44 samples). h smooth ROC of the high risk pneumotype classification applying a tenfold nested cross-validation scheme with ten repetitions (n = 44 samples). The black line represents the mean ROC curve, and the shaded area represents 95% confidence interval, grey lines represent ROC curves from individual folds. i Kaplan-Meier curves for survival probability between predicted pneumotypes for all patients (n = 44 samples). The survival differences between clusters were assessed using the Kaplan–Meier method and compared using a two-sided log-rank test. Hazard ratios and 95% confidence intervals were estimated using a Cox proportional hazards regression model. 12/31 patients have contributed for more one sample in (g–i.).

Importantly, the effect sizes of MEFISTO features identified that pneumotype separation (n = 64) was primarily driven by viral functional features, with lesser contributions from viral taxonomy and bacterial functions and taxonomy (Supplementary Fig. 15 and Supplementary Data 9-10).

A two-feature decision model using FFT achieved robust discrimination between high- and low-risk pneumotypes (accuracy =  90%, sensitivity = 94%, specificity = 84%; Fig. 4g). The high-risk signature was primarily driven by elevated relative abundances of Rhizobium bacteriophage, Cuauhnhuacivirus (>0.06%), whose lifestyle is unknown, and Lactobacillus bacteriophage, Lidleenavirus (>0.09%), which is predicted to be virulent. Model performance, evaluated through tenfold cross-validation, yielded a mean AUC of 0.9 (Fig. 4h). Using XGBoost with the two bacteriophage features, we analyzed the entire cohort independently of HAP status using samples collected within 0–4 days after study enrollment. Kaplan-Meier analyses showed reduced survival probabilities among individuals predicted as high-risk, compared with the low-risk group (Fig. 4i and Supplementary Fig. 25). Cox proportional hazards modeling further demonstrated that the predicted high-risk group was associated with increased mortality risk (HR = 4.8, 95% CI: 1.7–14, p = 0.0012). These findings show that the high-risk pneumotype harbors a virome signature reflecting lung viral convergence, characterized by the expansion and increased abundance of virulent bacteriophages.

Validation of the functional pneumotype scores in an independent cohort

Finally, we aimed to validate the accuracy of these scores in an independently recruited cohort of patients admitted in the Nantes University Hospital ICU for severe brain injury and who required invasive mechanical ventilation (IBIS cohort). We analyzed 239 ETAs samples collected on days 0, 3 and/or 7 after study enrollment (n = 117 patients, Supplementary Data 15). The HAP rate was 51%; with a median time of four days (IQR = 3-6 days) after being admitted to the ICU. The characteristics of the study population are described in Supplementary Table S9. The PREV-HAP and IBIS cohorts enrolled critically-ill patients in different years and different clinical centers, with different inclusion and exclusion criteria but similar sample processing and sequencing. In the entire patient cohort during the first 0 to 3 days following study enrollment, patients with high-risk pneumotype classification based on the independent two viral, or four bacterial, and on the six combined FFT features had lower survival probabilities than those classified as low-risk. Hence, Cox proportional hazards modelling therefore yielded an HRs of 2 (95% CI 0.96–4), 3.25 (95% CI 1.1–9.5), and 2.4 (95% CI 1.1–4.9), respectively (Fig. 5a–c and Supplementary Fig. 25). We found that patients classified using the combined FFT signature overlapped more with the viral signature (Supplementary Fig 27, 28).

Fig. 5. Validation in an external independent cohort (IBIS).

a–c Kaplan-Meier curves for survival probability between predicted pneumotypes using viral, bacterial and combined FFT features in all patients (n = 111 samples). The survival differences between clusters were assessed using the Kaplan–Meier method and compared using a two-sided log-rank test. Hazard ratios and 95% confidence intervals were estimated using a Cox proportional hazards regression model. d Relative abundance of the core healthy lung microbiome in the predicted pneumotypes (n = 50 samples). e Viral composition of the predicted pneumotypes (n = 50 samples). The size of the circle corresponds to the RPKM-normalized abundance (Log₁₀RPKM), with larger circles indicating higher abundance levels. f Gene set enrichment analysis of the upregulated differentially expressed genes in the predicted high risk pneumotype (n = 50 samples). GO enrichment was assessed using a one-sided hypergeometric over-representation test. p values were adjusted for multiple comparisons using the Benjamini-Hochberg method, with an adjusted p value threshold of 0.05.

Consistent with our previous findings, the relative transcriptional abundance of the healthy lung bacterial core metatranscriptome taxa was significantly lower in the pre-HAP samples with a high-risk pneumotype signature (using bacterial FFT) than in those with a low-risk pneumotype signature (22% vs. 38%, respectively; p = 0.02; Fig. 5d, Supplementary Data 16).

Additionally, bacteriophages dominated the predicted high-risk (log₁₀ RPKM median = 4.3, IQR = 3.9–4.5) and low risk pneumotype (predicted using viral FFT) virome composition (log₁₀ RPKM median = 3.8, IQR = 3.5–4.4; Fig. 5e; and Supplementary Data 17).

Finally, we investigated whether host immune responses could further discriminate patients’ pneumotypes in the validation cohort using samples collected before HAP onset. A partial least squares discriminant analysis (PLS-DA) using 450 inflammatory, immune-related genes distinguished between high- and low-risk pneumotype patients (Supplementary Fig. 29). The predicted high-risk group’s inflammatory profile was confirmed (Fig. 5f, Supplementary Fig. 30 and Supplementary Table S10). Overall, these findings demonstrate that viral and/or bacterial pneumotype classification schemes identify patients at high risk of mortality, whose lung dysbiosis is characterised by depletion of actively transcribed commensal bacteria, enrichment of bacteriophages, and expression of pro-inflammatory cytokines.

Causality inference and pathophysiological mechanisms of the high-risk pneumotype

To characterize bacteriophage-bacterial host interactions prior to the development of HAP, we examined annotated phages alongside their predicted bacterial hosts. We observed that the dynamics of core healthy lung bacteria and phages were either positively (for Veillonella, low risk pneumotype; p = 0.03) or negatively (for Streptococcus, in high risk pneumotype; p = 0.024) associated. Meanwhile, pathobionts predominantly showed positive temporal correlations in the high-risk pneumotype, including Staphylococcus (p = 0.037), Pseudomonas (p = 0.015) and Klebsiella (p = 0.03). Similar patterns were observed across additional taxa in both pneumotypes (Supplementary Fig. 31).

To further investigate the causal transkingdom interactions that may contribute to the high-risk pneumotype trajectory, we applied Transkingdom Network Analysis (TkNA) by integrating data from both PREV-HAP and IBIS cohorts to ensure statistical power. The input dataset comprised bacterial (22 bacterial species), viral (69 viral genera) and host signatures. Because host genes had very dynamic expression following ICU admission, the host signature was expanded to include genes changing with time (i.e., found differentially expressed regardless of association with HAP onset day), yielding in a total of 24 DEGs (Supplementary Table S4). The resulting network comprised 38 nodes (including seven bacterial species, eleven bacteriophages, and twenty human inflammatory genes) and 157 edges (Fig. 6a; Supplementary Table S11 and Supplementary information). Thirty-four nodes (five bacterial species, eleven bacteriophages, and eighteen inflammatory genes) exhibiting a median importance score of 0.2 (IQR = 0.05–0.9) (Fig. 6a, Supplementary Fig. 32 and Supplementary Table S12), supporting their potential for causal involvement in severe pre-HAP pneumotype development. Among these nodes, we observed negative correlations between several bacteriophages (i.e., Eowynvirus, Fletchervirus, Brussowvirus and Salasmaviridae) (Supplementary Fig. 33) and the bacterial genera Streptococcus_unclassified and Alloprevotella_unclassified, which were correlated with each other but anticorrelated with Klebsiella pneumoniae. Moreover, Streptococcus_unclassified and Alloprevotella_ unclassified showed strong negative interactions with a cluster of potent inflammatory genes encoding cytokines and receptors involved in immune activation including TLR8, IL18R1, IL18RAP, LY96, PTX3, CCL20 and CHI3L1 (Fig. 6a and Supplementary Table S11).

Fig. 6. Pathophysiological mechanism of the inflammatory and high risk pre-HAP functional pneumotype.

a Transkingdom network illustrating significant correlations between differentially transcribed bacteria, bacteriophage abundance, and inflammatory immune gene expression before HAP onset. Edges are colored according to the TkNA correlation type: blue indicates negative correlations and red indicates positive correlations. Edge width reflects the node importance score, and edge transparency indicates the level of statistical significance (p value) of the Spearman test after the false discovery rate (FDR) adjustment. Rectangles represent bacteria, circles represent bacteriophages, and octagons represent immune genes (n = 40 samples). b–g Boxplots plots showing the abundance (in log10 RPKM) of Streptococcus_unclassified (p = 0.0095, b Brussowvirus (p = 0.00006, c virus-to-host ratio in Streptococcus_unclassified (p = 0.0001, d Klebsiella pneumoniae (p = 0.0012, e, Eowynvirus (p = 0.0001, f virus-to-host ratio in Klebsiella (p = 0.18, g) between pneumotypes (n = 40 samples). Statistical significance was calculated by the two-sided Wilcoxon test. Data are presented through box plots, showcasing the 25th to 75th percentiles along with the median, while the whiskers extend to the minimum and maximum observed values. h–j Longitudinal trajectories showing the expression (in log10 TMM-normalised) of ADM (h), IL18R1 (i) and PTX3 (j) genes across pneumotypes over days related to HAP onset (n = 40 samples). The shaded area represents 95% confidence interval. Symbols used: Statistical significance was calculated by the Wilcoxon test. ns for not significant (p > 0.05), * for p ≤ 0.05, ** for p ≤ 0.01, *** for p ≤ 0.001, and **** for p ≤ 0.0001.

Two potential bacteria-bacteriophage interactions were identified within the network. In the high-risk pneumotype, we observed a significant reduction in Streptococcus_unclassified transcriptional activity (p = 0.0095; Fig. 6b) and anincrease in the abundance of virulent Streptococcus bacteriophage (Brussowvirus; Wilcoxon test, p = 0.00006; Fig. 6c), leading to high Streptococcus_unclassified virus-to-host ratio (VHR) (Fig. 6d). In contrast, there was an enrichment of Klebsiella pneumoniae (p = 0.0012; Fig. 6e) and its associated temperate bacteriophage (Eowynvirus; p = 0.0001; Fig. 6f) with no differences in Klebsiella VHR among the groups (p = 0.18; Fig. 6g). Furthermore, multivariate GLM modeling of gene expression across groups (Supplementary Fig. 34) revealed sustained elevated expression over time of ADM (p = 0.001; Fig. 6h), IL18R1 (p = 0.04; Fig. 6i) and PTX3 (p = 0.01; Fig. 6j) in high-risk patients. This temporal regulation was associated with Streptococcus-Brussowvirus and Klebsiella-Eowynvirus interactions.

Overall, these results support the hypothesis that candidate phage-Streptococcus antagonistic interactions may contribute to bacterial dysbiosis, promoting heightened inflammatory responses in patients who progress toward severe, high-risk HAP.

Discussion

In this study, we used an unsupervised and multi-omic approach to analyse endotracheal aspirates of critically ill patients. We integrated metatranscriptomic and viral metagenomic profiles in order to investigate lung microbiome heterogeneity. This strategy revealed two pre-HAP functional pneumotypes associated with all-cause mortality. The high-risk pneumotype showed increased mortality, a strong proinflammatory signature, and loss of healthy lung bacterial transcriptional activity. It also displayed elevated transcriptional activity of pathobionts such as K. pneumoniae. Additionally, this pneumotype showed a virome convergence driven by increased absolute abundance of bacteriophages. Finally, we defined two predictive lung signature scores derived from active bacteriome and virome features. These scores were validated in an external prospective cohort, predicting both pneumotypes across all ICU patients. Together, these distinct functional pneumotypes identified within the first days of ICU admission support plausible pathophysiological mechanisms underlying their divergent clinical outcomes.

The observation of two pre-HAP patient groups was notable, as it mirrors previous reports of two subphenotypes among 3889 critically ill patients at HAP diagnosis. These subphenotypes, which are linked to mortality and inflammation, were defined by clinical characteristics and routine biological tests⁶. In the present study, RNA expression profiling revealed that high-risk pneumotype overexpressed lymphocyte and leukocyte migration, tumor necrosis factor superfamily cytokine production and IL-18 regulation pathways, sustaining a proinflammatory profile related to poor outcomes^14–17. Deconvolution suggested that pneumotype definitions are mainly driven by transcriptional activity differences in immune cells; however, this does not exclude the possibility that increased immune cell abundance may contribute to the phenotype definitions. As these pneumotypes precede HAP onset, early lung inflammation may predispose patients for clinical deterioration and high-mortality HAP. This underscores the role of lung viral-microbial communities in both disease susceptibility and progression^8,10, highlighting the potential of longitudinal bedside microbiome profiling to guide personalized interventions.

High-risk patients exhibited viral convergence and enrichment with virulent bacteriophages as recently documented¹⁰. In this latest work, we reported significant viral convergence in ETA samples from patients who later developed HAP compared with those who did not¹⁰. In the present study, we observed an even stronger viral convergence in the high-risk pneumotype viromes than in the low-risk within the upcoming HAP group. Additionally, we identified an enrichment of virulent Caudoviricetes bacteriophages, which are anticorrelated with bacteria belonging to the core healthy respiratory group. Since limited immigration of exogenous phages during intubation may occur, these phage blooms could originate from endogenous reservoirs within the resident respiratory virobiota present on the lung mucosa or from prophages already integrated within bacterial genomes. ICU-associated stressors such as oxidative stress, inflammation, and nutrient limitation are recognized triggers of prophage induction and the transition toward the lytic cycle^18,19. This mechanism of “activation” allows phage DNA to exit the bacterial chromosome and proliferate en masse, potentially explaining the virulent phage blooms observed in high-risk pneumotypes¹⁰. Members of this tailed bacteriophage class are described among the key contributors to HAP onset in the lung viromes of critically ill patients¹⁰. We next derived a bacteriophage-based predictive score using viral metagenomic data. The decision tree incorporated two bacteriophages enriched in the high-risk pneumotype. Cuauhnahuacvirus is a bacteriophage targeting Rhizobium, a bacteria previously associated with prolonged mechanical ventilation in ICU patients⁸, and Lidleunavirus, a virulent bacteriophage increased in the ileal mucosal virome of IBD patients¹⁸ of Lactobacillus that was reported in ICU patients with early successful extubation⁸. The link between virome and lung disease has been described previously. Sputum metagenomes from 99 COPD patients and 36 controls revealed disrupted virus-bacteria ecological dynamics and progressive loss of bacteriophage diversity, particularly Porphyromonas-associated bacteriophages, in patients with frequent exacerbations¹⁹. Similarly, metatranscriptomic analysis of 278 bronchoalveolar lavage samples from 229 paediatric HCT patients identified viral enrichment as a key driver of fatal lung injury and high-risk in-hospital outcomes²⁰. In our previously described pathophysiological model, we hypothesised that virulent bacteriophages, which are not known to be strictly harmful and may even exert protective effects by controlling pathogens (i.e., phage therapy), could nevertheless contribute to HAP pathogenesis in specific ecological contexts. This would occur through predator-prey interactions, whereby the bacteriophages would target commensal core lung genera, such as Prevotella and Streptococcus. Depleting these populations of bacteria associated with health could disrupt lung microbial homeostasis and promote HAP onset¹⁰.

Consistent with this model, we observe a significant increase in the Streptococcus virus-to-host ratio suggesting that virulent Streptococcus bacteriophage (Brussowvirus) showed patterns consistent with possible antagonistic interactions with their bacterial hosts, in line with a potential kill-the-winner dynamic. This recurrent pattern further supports our hypothesis that bacteriophage-bacteria interactions contribute to dysbiosis associated with worse clinical outcomes. However, this study identifies the presence of K. pneumoniae exhibiting an opposing pattern: when bacterial transcription activity was high, its temperate bacteriophage (Eowynvirus) multiplied as a result of bacterial division and adopted the lysogenic lifestyle following the Piggyback-the-Winner hypothesis²¹. This mechanistic scenario provides a plausible explanation for the observed lung microbiome destabilization that has led to enhanced inflammatory responses. In line with these patterns, both DNA and RNA profiles showed elevated bacteriophage activity in the high-risk pneumotype. These results suggest that high-risk lung virome is actively undergoing functional remodelling, which is associated with shifts in bacteriophage abundance and potentially with changes in the genetic composition of individual bacteriophages.

Using metatranscriptomics, we defined the transcriptionally active bacteriome in lung fluids and identified a bacterial dysbiosis in the high-risk pneumotype, reflected by reduced activity of core healthy pulmonary taxa. This mirrors the dysbiosis previously identified using 16S profiles in the severe HAP subphenotype⁶ and highlights the added resolution of metatranscriptomics in distinguishing active pathogens and identifying microbial drivers of infection^20,22–24. Besides, metatranscriptomic data here contributed to the identification of the functional lung pneumotypes in critically ill patients before HAP induction. Recent work has demonstrated that, at pneumonia diagnosis and under antibiotic treatment, the lung microbiota segregates into four distinct 16S rRNA gene-based pneumotypes reflecting the disruption of the microbial landscape during pneumonia¹². Notably, a pneumotype enriched in oral-associated taxa, including Streptococcus, was associated with successful treatment of HAP and with upregulation of IL-1 signalling pathways in alveolar macrophages, partially mirroring our low-risk functional pneumotype. We further derived a four-factor bacterial transcriptomic signature, in which high-risk status was defined by low Alloprevotella expression, a commensal bacteria in ventilated ICU patients²⁵ also linked to treatment-responsive lung masses²⁶ and improved COPD outcomes²⁷, and by elevated Paracoccus and K. pneumoniae. Paracoccus is enriched in ARDS-associated microbiomes⁸, while K. pneumoniae is a common VAP pathogen²⁸ associated with higher mortality, disease severity, and respiratory failure in ICU cohorts²⁹.

Viral- and bacterial-derived scores effectively identified pre-HAP pneumotypes in an external ICU cohort, even within the first days of admission, underscoring their utility for early risk stratification. These results support the concept that stabilizing the lung microbiome through prudent antimicrobial use, probiotic approaches, or targeted bacteriophage-based strategies may help prevent the dysbiosis preceding severe HAP³⁰.

This study has limitations: The discovery cohort was small, with only 39 pre-HAP ETA samples from patients who later developed HAP used for multi-omic pneumotype identification. Despite stringent decontamination procedures, ETA samples may be subject to upper airway microbial contamination due to the sampling process, variability introduced by differences in suctioning practices, and the impact of heterogeneous cellular composition on microbial DNA recovery. Even when adding IFN-γ treatment as a covariate, given the small size of the treated subgroup, IFN-γ treatment remains a potential confounding factor that cannot be completely excluded. Nevertheless, pneumotype definition was independently validated in an observational cohort in which no patients received interferon-γ-1b treatment, further supporting that they can be extrapolated to populations independently from immune therapy. The predictive model could not be fully validated externally, as pneumotypes are inferred rather than being clinically recorded. Finally, causality cannot be confirmed without in vivo experimentation. Nevertheless, a key strength is the validation of our findings in a prospective cohort.

In conclusion, our multi-omic analysis revealed substantial heterogeneity in the transcriptomic activity of the lung virome and bacteriome of ICU patients prior to HAP. The distinct functional pneumotypes associated with inflammation and increased mortality highlight the clinical importance of identifying airway community structure early on. Machine learning-based scores incorporating four bacterial and two bacteriophage markers accurately stratified patients upon admission and identified those at elevated risk. Furthermore, we detected a bacterial-viral dysbiosis signature accompanied by inflammatory gene expression several days prior to HAP onset, suggesting that early microbial disruption may contribute to disease development. Together, these findings refine our understanding of HAP pathophysiology and lay the groundwork for future diagnostic and microbiome-targeted interventions in critical illness.

Methods

This study complied with all relevant ethical regulations and was conducted in accordance with the Declaration of Helsinki. The PREV-HAP study protocol (NCT04793568) was approved by the Ouest II Angers Ethics Committee, France, in March 2021, and the IBIS study (NCT02003196) was approved by the Comité de Protection des Personnes Ouest IV, France.

Cohorts description and ETAs collection

The discovery cohort included 94 patients enrolled in the PREV-HAP clinical trial. This was an investigator-initiated, multicenter, parallel-group, double-blind, randomized study conducted in 11 ICUs in France, Spain, and Greece. The study protocol was approved by the Ouest II Angers Ethics Committee in France in March 2021, and the trial adhered to the Declaration of Helsinki. The trial was registered on ClinicalTrials.gov (NCT04793568) in the same month. Written informed consent was obtained from each patient’s legal surrogate. In accordance with local regulations, patients could be included before surrogate consent was obtained if the next of kin could not be contacted within the allowed enrollment window. Follow-up consent was requested from the patient within 90 days of inclusion, when feasible. Eligible patients were aged 18–85 years, receiving invasive mechanical ventilation, and had at least one acute organ failure at the time of enrollment. Participants were randomized in a 1:1 ratio to receive either interferon gamma-1b (100 µg/0.5 mL vials, IMUKIN®, Clinigen®) or placebo (normal saline) in fixed blocks of six, stratified by hospitalization cause (sepsis vs. other) and country (France, Greece, Spain). Key exclusion criteria included pregnancy or breastfeeding, hypersensitivity to interferon gamma-1b, pre-existing immunosuppression (e.g., recent chemotherapy or radiotherapy, AIDS, leukopenia), severe hepatic insufficiency (Child-Pugh B or C), liver cytolysis (transaminases >5× normal), chronic renal failure (MDRD GFR < 10 mL/min/1.73 m²), persistent coma post-resuscitated cardiac arrest, cervical spinal cord injury hospitalization, previous hospital-acquired pneumonia during the current admission, and sustained hyperlactatemia (>5 mmol/L). Patients received five subcutaneous injections of either interferon gamma-1b or matching placebo from day 1 to day 9 (one injection every 48 h). Samples were collected immediately before the first study treatment injection (interferon or placebo) at day 0, then at day 3 (immediately before the 2nd injection) and day 7-8 (immediately before the 4th injection). Follow-up continued for 90 days. HAP and ARDS diagnoses were blindly reviewed by intensivists for compliance with international definitions. HAP was defined as pneumonia occurring ≥48 h after hospital admission, requiring at least two clinical signs (fever >38 °C, leukocytosis >12,000 cells/µL, leukopenia <4000 cells/µL, or purulent pulmonary secretions), accompanied by a new or worsening infiltrate on chest imaging, and confirmed with semi-quantitative or quantitative respiratory cultures obtained prior to initiating new antimicrobial therapy. All respiratory infection diagnoses were reviewed by an independent adjudication committee following European HAP guidelines. ARDS was defined as severe hypoxemia (PaO₂/FiO₂ <200 mmHg with PEEP ≥ 5 cm H₂O) and bilateral opacities on chest imaging within one week of worsening respiratory symptoms. Information on ancestry, race, ethnicity, and socioeconomic status was not available for either cohort due to legal restrictions in France.

In the validation cohort IBIS (Immunology and Brain Injury Study), 117 patients were enrolled in two French Surgical Intensive Care Units of one university hospital (Nantes, France). The collection of human samples has been declared to the French Ministry of Health (Programme de recherche “Immunologie”, DC-2017-2987), and was approved by the Comite de Protection des Personnes Ouest IV (7/04/2015 and 08/10/2020) (number clinicaltrials.gov NCT02003196). Written informed consent from a next-of-kin was required for enrolment. Retrospective consent was obtained from patients when possible. Appropriate consent was obtained for the release of information from deceased individuals. Participants received no compensation. Inclusion criteria were male or female, 18-80 years old, brain injury (Glasgow Coma Scale below or equal to 12 and abnormal brain-CT scan) and receiving invasive mechanical ventilation. Exclusion criteria were cancer with radiotherapy or chemotherapy in the last 90 days, AIDS, leukopenia in the previous five years, immunosuppressive drugs, preexisting immunosuppression and pregnancy. All patients were clinically followed up for 90 days. Intensivists blindly reviewed HAP and ARDS diagnoses for compliance with international definitions^5,31,32.

In this study, a total of 423 endotracheal aspirates (ETAs) were collected from critically ill patients enrolled in two independent cohorts: PREV-HAP (184 ETAs from 94 patients) and IBIS (239 ETAs from 117 patients), based on sample availability. ETAs were always collected with disposable, sterile tubes through endotracheal tubes, thus limiting the risk of contamination by oropharyngeal flora.

Metatranscriptomic sequencing for integrated microbiome and host analysis

Metatranscriptomic analysis was conducted using the Revelo kit (Tecan) according to the protocol described by Destras et al. 2026³³. Briefly, 50 µL of ETA samples or no-template controls (NTCs) were spiked with an RNA internal control (MS2 bacteriophage, 2800 bp) at a concentration corresponding to a Ct value of 36. Nucleic acids were then extracted using the Maxwell platform (Promega) with the DNA Blood Kit. Sequencing was performed on a NovaSeq 6000 system using an SP flow cell for PREV-HAP cohort and S2 cartridge for IBIS, with 2 × 100 bp paired-end reads.

Metagenomic sequencing for viral profiling

ETAs were preserved diluted (dilution factor of 1:7.4) in a viral transport medium (LMR1925VTM, Labomoderne) and stored at -20 °C. Samples from both cohorts were treated according to an adapted version of the protocol from Bal and colleagues³⁴ and Anani et al.¹⁰. This protocol is designed to capture both DNA and RNA viruses. Briefly, the MS2 RNA bacteriophage (MS2, IC1 RNA internal control; bioMérieux R-GENE® ref. 71–110) was spiked in each sample to validate the whole metagenomic process. The samples were passed through 0.45 µm filters, and the resulting filtrate was incubated with DNase (Life Technologies ®, Carlsbad, USA) for 90 min. Total nucleic acid extraction was performed on the Qiagen EZ1 Advanced XL Extractor using the DSP Virus Kit, Qiagen®. After extraction, both RNA and DNA amplification were performed using a Whole Transcriptome Amplification kit (WTA2, Sigma-Aldrich®, Darmstadt, Germany). Amplified DNA was purified using QIAquick spin columns (Qiagen®, Hilden, Germany). No-template controls (NTC) and MS2 spiked-in NTC were prepared for each sample batch using nuclease-free water that followed the whole process starting from dilution in the transport media. The viral transport medium was also sequenced to confirm the absence of viral contamination. Before sequencing, quality control testing was performed with MS2 qPCR (R-GENE® ref. 71-110). The libraries were prepared using the post-PCR portion of the Illumina (San Diego, USA) COVIDSeq ® Test Kit for PREV-HAP samples and the Nextera XT Kit for IBIS samples. The libraries were then sequenced on a NovaSeq 6000 using SP (PREV-HAP) and S2 (IBIS) flow cells with 2×100 bp reads.

Bioinformatic pipeline for virome analysis

Raw sequencing reads were dehosted using SraHumanScrubber v2.2.1 to remove human-derived sequences and trimmed with fastp v0.23.4 to eliminate adapters, low-complexity regions, and low-quality bases, retaining only reads longer than 30 bp. Taxonomic classification was performed with Kraken2 v2.1.2 (parameters: minimum-hit-groups = 2, confidence = 0.1) using a custom database incorporating viral taxa from RVDB (v29.0) and Inphared (v2, January 2025)³⁵, as well as archaeal, bacterial, fungal, and human references from RefSeq (downloaded on 06/01/2025)³⁶. Non-viral reads assigned to Eukaryota, Bacteria, Archaea, Fungi, and their taxonomic descendants were filtered out using Krakentools, retaining only putative viral reads. These viral reads were assembled de novo with metaSPAdes³⁷ v3.15.5, and the resulting contigs were dereplicated using CD-HIT³⁸ v4.7 to generate operational taxonomic units (OTUs) at ≥95% nucleotide identity and ≥85% coverage, following MIUViG (Minimum Information about an Uncultivated Virus Genome) standards. OTUs shorter than 500 bp were discarded. The length-filtered OTUs were analyzed with PhaBOX³⁹ v2.1.13 to identify viral OTUs (vOTUs), assign taxonomy according to the latest International Committee on Taxonomy of Viruses (ICTV) release, and predict host range and lifestyle for bacteriophages using CHERRY³⁹ and PhaTYP³⁹, respectively. Finally, quality-filtered viral reads were realigned to the vOTU sequences using the Burrows–Wheeler Aligner (BWA)⁴⁰ v0.7.17, and count tables were generated with samtools. Viral decontamination was performed following the recently published MS2 internal control based method described by Anani and colleagues¹⁰. Viral metagenomic contaminants are listed in Supplementary Data 18.

The following sequencing quality thresholds were applied to select samples for virome analysis: MS2 internal control detection and a minimum sequencing depth of 10,000 viral RPKM (Supplementary Fig. 12). This resulted in 133/184 and 236/239 interpretable virome samples for PREVHAP and IBIS cohorts, respectively.

Bioinformatics analysis of metatranscriptomic data

Metatranscriptomic data were analyzed using a pipeline that processed each kingdom independently. Prior to downstream analyses, all reads (FASTQ files) were trimmed for quality using fastp v0.23.4 (reference). First, human transcriptomic analyses were performed on non-dehosted, trimmed reads using RASflow⁴¹ v2.0, HISAT2 v2.2.1, and the GRCh38 reference genome with default parameters. Duplicate reads were removed using Samtools, and FeatureCounts v2.0.3 was used to generate a gene count matrix. Only protein-coding genes detected in at least 10% of the samples were retained for further analysis. Second, for bacterial metatranscriptomic analyses, trimmed reads were first dehosted using SraHumanScrubber v2.2.1 with the GRCh38 human reference genome. Bacterial reads were then identified using Kraken2³⁵ v2.1.2 (minimum-hit-groups = 2, confidence = 0.1) against the previously described custom database. Reads assigned to the bacterial taxon (taxid = 2) were extracted using Krakentools. From these, 16S rRNA reads were removed using SortMeRNA⁴² v4.3.6 with the SILVA database (v138.2), and the remaining reads were assembled de novo with rnaSPAdes³⁷ v3.15.5. Assembled bacterial transcripts were dereplicated using CD-HIT³⁸ v4.7 to generate operational taxonomic units (OTUs) with ≥95% sequence identity and ≥95% coverage. OTUs shorter than 1 kb were discarded. The remaining length-filtered OTUs were taxonomically assigned using DIAMOND⁴³ v2.1.6.160 with the lowest common ancestor (LCA) algorithm against the NCBI RefSeq Bacteria database (version 4). Quality control of bacterial OTUs (bOTUs) was performed to assess completeness and contamination. Functional annotation was carried out with Prokka⁴⁴ v1.14.6, and bOTUs lacking coding sequences were excluded. Quality-filtered bacterial reads used for de novo assembly were realigned to the curated bOTUs using the Burrows–Wheeler Aligner (BWA)⁴⁰ v0.7.17, and count tables were generated with samtools. Finally, potential bacterial contaminants were identified using the decontam⁴⁵ R package v1.20.0, applying the prevalence method with a threshold of 0.5 on read count tables normalized to bacterial RPKM.

Third, for viral metatranscriptomic analyses, we performed preprocessing, quality assessment, and dehosting of raw paired-end reads using fastp v0.23.4 and SraHumanScrubber v2.2.1, applying the same parameters used in the viral metagenomic pipeline. All subsequent steps followed the viral metagenomic workflow parameters, except for the de novo assembly step. For this step, rnaSPAdes was used instead of metaSPAdes. Viral decontamination was carried out as described for viral metagenomic analyses. Bacterial and viral metatranscriptomic contaminants are listed in Supplementary Data 19.

The following sequencing quality thresholds were applied to select samples for metatranscriptome analysis: MS2 internal control detection and a minimum sequencing depth of 10,000 bacterial and 10,000 viral RPKM after decontamination (Supplementary Fig. 12). This resulted in 116/122 and 239/239 interpretable active bacteriome samples, 116/122 and 239/239 interpretable active virome samples and 115/122 and 136/239 interpretable host transcriptome samples for PREVHAP and IBIS cohorts, respectively.

Functional microbiome and virome analysis

For the bacterial component, high-quality reads obtained after ribosomal RNA filtration were analyzed using the HUMAnN⁴⁶ v3.9 pipeline from Biobakery to infer bacterial metabolic pathways. For functional annotation, all bacterial transcripts or bOTUs were annotated with Prokka⁴⁴ v1.14.6 (using the bacterial kingdom option), and the resulting protein sequences were searched against the latest NCBI database of Clusters of Orthologous Genes (COGs)⁴⁷ using BLASTp. For the viral component, derived from viral metagenomic or metatranscriptomic datasets, all vOTUs were annotated with Prodigal v2.6.3. The predicted viral proteins were then compared against latest UniProt SPROT curated viral databases, and Gene Ontology (GO) terms were retrieved from UniProtKB to classify viral proteins into functional categories following Cao and colleagues¹⁸.

Statistics and reproducibility

Unsupervised clustering analysis

Microbiome data characterized by virome metagenomes, virome and bacteriome transcriptomes and their corresponding identified proteins in addition to the bacterial metabolic pathway dataset were analyzed all using MEFISTO⁴⁸ (Method for the Functional Integration of Spatial and Temporal Omics data) to reduce dimensionality and identify a core set of factors over time by launching the python/R coupled library MOFA2⁴⁹ v1.18.0. This method effectively handles different data structures and distributions while being robust to collinearity. The data were filtered to retain features present in more than 50% of the samples. This threshold allowed the most stable and reproducible patient stratification across independent clustering criteria (Supplementary Fig. 14) and therefore clustering is based on differences in the core microbiome. Data were subsequently subjected to a log₁₀(x + 1) transformation for MEFISTO analysis. MEFISTO identified 15 latent factors. The factors 1 and 2 were then subjected to unsupervised hierarchical clustering, and two clusters were determined to be optimal based on Silhouette, Elbow, and Gap statistics. Then, each patient was assigned principal component coordinates derived from the first two components. These coordinates were visualized in a two-dimensional space using the FactoMineR v2.12 R package. For multi-omic analyses (MEFISTO), we indexed time relative to the onset of HAP, counting backwards from the onset of HAP (Fig. 2a and Supplementary Fig. 5). In the Survival analysis, we have selected one sample per patient (Fig. 2b) which was the closest to the HAP diagnosis day (n = 39 samples corresponding to 39 patients). As some patients have several samples collected before HAP onset, for the signatures (Figs. 2d, e, 3 and 4), we have added the patient identity as fixed covariate to control multiple samples (n = 64 samples corresponding to 39 patients; Supplementary Fig. 4–5).

Clinical characteristics

We fitted a Cox proportional hazards model to assess the association between clinical variables and ICU mortality. This model included cluster assignment, demographic factors, and key clinical parameters. We verified model performance and proportionality assumptions and visualized relative risks (RRs) with 95% confidence intervals using a forest plot generated with the ggforest function from the survminer v0.5.1 R package. We conducted Kaplan–Meier survival analyses to evaluate in-hospital mortality and the probability of invasive mechanical ventilation across the identified clusters. To account for competing risks, death prior to extubation was treated as a competing event rather than censored, since censoring death would overestimate the probability of extubation. The survival v3.8.3 R package was used to generate survival curves and compare them using the log-rank test to assess differences in survival distributions between clusters. Survival time was defined as the period from study enrollment to either ICU mortality. Survival probabilities were estimated across pneumotypes using Kaplan-Meier analyses, and differences in mortality risk were quantified using Cox proportional hazards models.

Host gene expression

Differentially expressed genes (DEGs) were identified using edgeR v4.6.3 within each cluster based on normalized counts. We visualized the DEGs using volcano plots to illustrate significance and fold-change distribution. Gene set enrichment analysis (GSEA) was then performed on upregulated DEGs of each cluster using the clusterProfiler v4.16.0 and org.Hs.eg.db v3.21.0 R packages. Interferon, corticosteroids, antibiotic treatment (with antibiotic class and timing), HAP onset day, ICU ventilation duration and patient identity were added as covariables in the GLM model to determine DEGs. Finally, the MCP-counter tool (https://github.com/ebecht/MCPcounter/tree/master) was used for the deconvolution analysis.

Microbiome comparisons and predictions

Alpha diversity, as measured by the Shannon index and richness, was calculated at the species level for bacteria and at the genus level for viruses. This calculation was performed using the vegan v2.7.1 package in R (v4.5.1). Diversity distributions were visualized using the ggplot2 v3.5.0 package, and longitudinal trends were modeled with loess regression and 95% confidence intervals. The ANCOM-BC⁵⁰ v2.10.1 was applied to identify features significantly associated with the identified clusters and to account for sampling bias (The analysis was performed on the full dataset without applying a prevalence filter). Differentially abundant viruses and expressed bacteria signatures were adjusted with all covariates used to determine host DEGs.

To identify bacterial/viral functional categories associated with each cluster, a generalized linear model (GLM) with a binomial family was applied. Briefly, normalized features (RPKM values) were compared between clusters by fitting individual GLMs for each function, modeling cluster membership as the binary outcome. For each function, log-odds estimates and p values were extracted, and features with p < 0.05 were considered significant. In addition, a multivariate PERMANOVA based on Bray–Curtis dissimilarities was performed using the adonis2 function of vegan v2.7.1 to assess the overall functional compositional differences between clusters. PLS-DA analyses were performed using the mixOmics R package v6.32.0. Comparisons between the virome and bacteriome compositions in both the discovery and validation cohorts were performed using the UpSetR v1.4.0 R package for visualization of shared and unique features. Machine learning analyses were conducted using the FFTrees⁵¹ v2.1.0 and XGBoost⁵² v1.7.11.1 R packages to evaluate the predictive performance of the selected signatures. For binary classification (high vs. low risk), a probability threshold of 0.5 was used to assign class labels. CARET v7.0-1 was used to assess model predictions performance (receiver operating characteristic (ROC)), to perform tenfold nested cross-validation and calculate features stability. For the decision tree analysis, all samples collected before HAP onset were treated as independent observations.

Transkingdom Network Analysis (TkNA⁵³ v1.2.1) was applied to perform causal inference using log₁₀(relative abundance data) from the viral/bacterial signatures and the log₁₀ (TMM-normalized counts) from the host genes signature, with the resulting correlation network visualized in Cytoscape⁵⁴ v3.10.2. To identify the key nodes in the network, we computed an importance score by integrating three topological features: node degree, betweenness-based bipartite centrality (BiBC), and the probability of randomly selecting each node. Each metric was normalized between 0 and 1, and the probability was inverted so that lower values indicated higher importance. The final importance score was obtained by averaging the normalized degree, BiBC, and probability values. For comparisons of numerical and categorical variables, the Wilcoxon test, two-tailed Mann–Whitney U test, and Fisher’s exact test were applied. Statistical significance was defined as p ≤ 0.05, and symbols were used as follows: ns, not significant (p > 0.05); *, p ≤ 0.05; **, p ≤ 0.01; ***, p ≤ 0.001; ****, p ≤ 0.0001.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

H.A., C.B.B., A.R. and L.J. conceived the study. S.B., G.B., F.P.M., C.P., M. P., E.O., D.S. and Q.S. performed the clinical data curation, sample preparations and sequencing. H.A. performed bioinformatic and statistical analysis. H.A., G.D., F.P.M., A.M., L.C., E.M., V.G., C.B.B., A.R. and L.J. interpreted results and drafted the manuscript. H.A., G.D., F.P.M., C.P., C.B.B., A.R. and L.J revised the manuscript. All authors reviewed and approved the final version of the manuscript.

Peer review

Peer review information

Nature Communications thanks Leopoldo Segal, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

The study was funded by the European Union’s Horizon 2020 research and innovation program under grant agreement number 847782 (HAP2 project) and MSD Avenir grant (Phenomenon project). A.M. has received funding from the European Respiratory Society and the European Union’s H2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 847462 and the European Union’s Horizon Europe research and innovation programme under the Marie Skłodowska-Curie grant agreement number 101198435 (RESTORE). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data availability

The metatranscriptomic and the viral metagenomic raw data have been deposited in the BioProject repository PRJNA1373167 and PRJNA1371704, respectively.

Code availability

All code used for the analyses in this study is available on GitHub at https://github.com/genepii/HAP2-PREVHAP-HAP_pneumotypes and on Zenodo (https://doi.org/10.5281/zenodo.20552016).

Competing interests

L.C., E.M. and A.R. have a patent on respiratory microbiome composition in ICU patients (W). The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-78146-z.