A prevalent and culturable microbiota links ecological balance to clinical stability of the human lung after transplantation

Combined culture-dependent and culture-independent approach identifies the prevalent and viable bacterial community members of the human lung post-transplant

To characterize the bacterial community composition of the lung microbiota post-transplant, we performed 16S rRNA gene amplicon sequencing of 234 longitudinal BALF samples from 64 lung transplant recipients collected over a 49-month period (Fig. 1a, Supplementary Table 1). A total of 7164 operational taxonomic units (OTUs) were identified, excluding OTUs contributing to reads in 11 negative control samples³² (see “Methods”, Supplementary Fig. 1a, Supplementary Data 1 and 2). In accordance with previous studies on BALF samples from healthy non-transplant individuals^4,5,6,26, we found that Bacteroidetes and Firmicutes followed by Proteobacteria and Actinobacteria are the most abundant phyla in the post-transplant lung (Fig. 1b). Prevalence analysis across all BALF samples showed that the community composition is highly variable with only 22 OTUs shared by ≥50% of the samples (Supplementary Fig. 1b, Supplementary Data 3). However, these 22 OTUs constituted 42% of the total number of rarefied reads, indicating that they are predominant members of the post-transplant lung microbiota (Fig. 1c, Supplementary Fig. 1c, Supplementary Table 2, Supplementary Data 3). They belonged to the genera Prevotella 7, Streptococcus, Veillonella, Neisseria, Alloprevotella, Pseudomonas, Gemella, Granulicatella, Campylobacter, Porphyromonas and Rothia, the majority of which are also prevailing community members in the healthy human lung^3,5,7,26, suggesting a considerable overlap in the overall composition of the lung microbiota between the healthy and the transplanted lung.

Differences in bacterial loads between samples can skew community analyses when based on relative abundance profiling alone. Therefore, we used qPCR to determine the total copies of the 16S rRNA gene as an estimate for bacterial counts, and normalized the abundances of each OTU across the 234 samples (absolute abundance). We found that the bacterial counts vastly differed between samples, ranging between 10¹ and 10⁶ gene copies per ml of BALF (Supplementary Fig. 1d). The number of observed OTUs increased with decreasing counts (Fig. 1d) suggesting that a large fraction of the OTUs were detected in samples of low bacterial biomass and hence represent either transient or extremely low-abundant community members, or sequencing artefacts and contaminations. In turn, 19 of the 7164 OTUs constituted >75% of the total bacterial biomass detected across the 234 BALF samples (Fig. 1e). This included 11 of the 22 most prevalent OTUs (see above) plus eight OTUs that were detected in only a few samples but at very high abundance (Staphylococcus; OTU_2, Corynebacterium 1; OTU_16 and OTU_24, Anaerococcus; OTU_49 and OTU_234, Haemophilus; OTU_78, Streptococcus; OTU_6768, Peptoniphilus; OTU_63, Supplementary Table 2). It is important to differentiate these opportunistic colonizers from other community members with low incidence, as they reached very high bacterial counts in some samples with potential implications for lung health.

To demonstrate the viability of prevalent lung microbiota members and to establish a reference catalogue of bacterial isolates from the human lung for experimental studies, we complemented the amplicon sequencing with a bacterial culturing approach (Supplementary Fig. 2). We cultivated 21 random BALF samples from 18 individuals, on 15 different semi-solid media (both general and selective) in combination with 3 oxygen concentrations; aerobic, 5% CO₂, and anaerobic (See “Methods” and Supplementary Table 3), representing 26 different conditions. We cultured fresh BALF immediately upon extraction (within 2 h), as we observed loss in bacterial diversity upon cultivating frozen samples. This resulted in a total of 300 bacterial isolates, representing 5 phyla, 7 classes, 13 orders, and 17 families from which we built an open-access biobank called the Lung Microbiota culture Collection (LuMiCol, Supplementary Data 4, https://github.com/sudu87/Microbial-ecology-of-the-transplanted-human-lung).

To examine the extent of overlap between bacteria in LuMiCol and the diversity obtained by amplicon sequencing, we included 16S rRNA gene sequences from 215 isolates that passed our quality filter into the community analysis, which allowed for the retrieval of OTU-isolate matching pairs³² (Methods). We found that 213 isolates matched to 47 OTUs (Fig. 2a, c, Supplementary Data 5), including 17 of the most prevalent and abundant bacteria (Fig. 1e, Supplementary Table 2). As expected, bacteria with high abundance in the amplicon sequencing-based community analysis were isolated more frequently, with Firmicutes revealing the highest isolate diversity (Fig. 2a–c, Supplementary Data 4, 5) and being recovered under the most diverse culturing conditions.

In summary, our results from the combined culture-dependent and culture-independent approach show that the lung microbiota post-transplant is highly variable in terms of both bacterial load and community composition with many transient and low-abundant bacterial taxa. However, a few community members display relatively high prevalence and/or abundance suggesting that they represent important colonizers of the human lung.

LuMiCol informs on the diversity and metabolic preferences of culturable human lung bacteria

We characterized the culturable community members of the lower respiratory tract contained in LuMiCol by testing a wide range of growth conditions and phenotypic properties (see “Methods”). The majority of the cultured isolates could taxonomically be assigned at the species level based on genotyping of the 16S rRNA gene V1-V5 region. However, the limited taxonomic resolution offered by this method does not allow to discriminate between closely related strains, which can include both pathogenic and non-pathogenic bacteria. Hence for Streptococcus, we additionally tested for type of hemolysis (alpha, beta, or gamma) and resistance to optochin, which differentiates the pathogenic pneumococcus and the non-pathogenic viridans groups (Fig. 2a, Supplementary Fig. 2b, c). This demonstrated that the 16 Streptococcus OTU-isolate pairs belong to the viridans group of streptococci (VS)³³. Interestingly, these isolates exhibited the highest genotypic and phenotypic diversity throughout our collection and belonged to five OTUs among the 22 most prevalent community members, with Streptococcus mitis (OTU_11) present in 93.6% of all samples.

BALF from healthy individuals contains amino acids, citrate, urate, fatty acids, and antioxidants such as glutathione but no detectable glucose³⁴, which is associated with increased bacterial load and infection^35,36,37. To get insights into basic bacterial metabolism, we assessed the growth of all 47 isolates matching an OTU under different oxygen concentrations. We used undefined rich media (Todd-Hewitt Yeast extract) and defined low-complexity liquid media (RPMI 1640), including a glucose-free version to mimic the deep lung environment (see “Methods”). Despite the presence of oxygen in the human lung, the majority of the isolates were either obligate or facultative anaerobes (Fig. 2a), including some of the most prevalent members (Prevotella melaninogenica (OTU_3), Streptococcus mitis (OTU_11), Veillonella atypica (OTU_6) and Granulicatella adiacens (OTU_17). A similar trend was also observed in liquid media under anaerobic conditions, with the exception of the genera Prevotella, Veillonella and Granulicatella. Most streptococci from the human lung grew best in complex liquid media containing glucose under anaerobic conditions, including the most prevalent species in our cohort, S. mitis (OTU_11) (Fig. 2b). However, noticeable exceptions were S. vestibularis (OTU_34), S. oralis (OTU_3427 and OTU_1567), and S. gordonii (OTU_10031), which grew equally well in the presence of oxygen and in low-complexity liquid medium (Fig. 2b). Most Actinobacteria grew best on rich medium in the presence of 5% CO₂, with an exception of Actinomyces odontolyticus (OTU_39), which required anaerobic conditions. Some Actinobacteria grew equally well in anaerobic conditions as in the presence of 5% CO₂, i.e., Corynebacterium durum (OTU_501), Actinobacteria sp. oral taxon (OTU_328 and OTU_228).

The two most predominant opportunistic pathogens in our lung cohort, P. aeruginosa (OTU_1) and S. aureus (OTU_2), grew best in rich liquid medium in the presence of oxygen (Fig. 2c), although these also grew to lower degree under anaerobic conditions. These results indicate that changes in the physicochemical conditions in the lung may favor the growth of these two opportunistic pathogens. In summary, our observations from the bacterial culture collection provide first insights into the phenotypic properties of human lung bacteria and will serve as a basis for future experimental work.

Identification of four compositionally distinct pneumotypes post-transplant using machine learning based on ecological metrics

To detect and characterize differences in bacterial community composition between BALF samples from transplant patients, we clustered the samples using an unsupervised machine learning algorithm based on pairwise Bray–Curtis dissimilarity³² (beta diversity, See “Methods”, Supplementary Data 6). This segregated the samples into four partitions around medoids (PAMs) at both phylum and OTU level (Fig. 3a, b, Supplementary Fig. 3a, b). We refer to these clusters as “pneumotypes” PAM1, PAM2, PAM3, and PAM4 (Supplementary Table 4). PAM1 formed the largest cluster consisting of the majority of samples (n = 115) followed by PAM3 (n = 76), PAM2 (n = 19), and PAM4 (n = 24) (Supplementary Data 7). Examination of various diversity measures (Species occurrence, OTU diversity, OTU richness, Fig. 3c–e), distribution of the dominant community members (Fig. 3f), and bacterial counts (16S rRNA gene copies, Fig. 3g) revealed distinctive characteristics between the four pneumotypes.

PAM1 showed the highest similarity in community composition between samples (Species occurrence/Sorenson’s Index, Fig. 3c), and had intermediate levels of diversity and bacterial load (Fig. 3d, e and g, Supplementary Fig. 3c). Twenty of the 22 most prevalent community members were enriched in incidence and abundance in PAM1 when compared to the other PAMs (ART-ANOVA, FDR, abundance P < 0.01, Fig. 3h, Supplementary Table 5) with five OTUs occurring in >90% of the samples (incidence); P. melaninogenica (OTU_3, 97.4%), S. mitis (OTU 11, 99.1%), V. atypica (OTU_6, 93.9%), V. dispar (OTU_30, 93%) and G. adiacens (OTU_17, 93%). Contrastingly, two OTUs (P. aeruginosa; OTU_1 and P. fluorescens; OTU_15) had neither a higher incidence nor a higher abundance in PAM1 (Fig. 3h, Supplementary Table 5). Thus, PAM1 samples harbor balanced bacterial communities of relatively high similarity composed of the most prevalent bacteria across our data set. Henceforth, we refer to this PAM as the ‘balanced pneumotype’ (Pneumotype_Balanced).

In contrast to Pneumotype_Balanced, PAM2 and PAM4 harbored lower bacterial diversity (Fig. 3d) and OTU richness (Fig. 3e), were dominated by a single community member (Fig. 3f), and had higher bacterial loads (Fig. 3g). In these two PAMs, the taxa associated with Pneumotype_Balanced had a low sample incidence and absolute abundance compared to the other PAMs (Supplementary Fig. 4a, b). S. aureus (OTU_2), Corynebacterium (OTU_24) and Anaerococcus (OTU_49) were enriched in PAM2 (ART-ANOVA, FDR, P < 0.001, Supplementary Fig. 4a), while Haemophilus (OTU_78) and P. aeruginosa (OTU_1 & OTU_15) dominated PAM4 (ART-ANOVA, FDR, P < 0.001, Supplementary Fig. 4b). We refer to these as ‘Pneumotype_{Staphylococcus}’ (PAM2) and ‘Pneumotype_Pseudomonas’ (PAM4), with the major species known to be potential pathogens that proliferate rapidly in lung, under a variety of pathological respiratory conditions^38,39, in line with the high loads observed in our setting (Fig. 3g).

The fourth cluster identified, PAM3, exhibited the lowest between-sample similarity in species composition (Fig. 3c), the highest OTU diversity and richness (Fig. 3d, e), and lowest dominance (Fig. 3f). The samples in this PAM were characterized by remarkably low bacterial loads, up to two orders of magnitude below samples in other PAMs (Fig. 3g, Supplementary Fig. 3c), suggesting a depauperated microbiota that has been associated with dysbiotic physiological states of the gut microbiota⁴⁰. Consequently, the high OTU richness detected in PAM3 samples is likely due to over-sequencing of rare or transient species, or sequencing artefacts. This is also supported by the fact that the 30 predominant microbiota members in PAM3 were significantly reduced in their incidence and abundance compared to the other PAMs (ART-ANOVA, FDR, P < 0.001, Fig. 3i). We refer to this PAM as the ‘microbiota-depleted’ pneumotype (Pneumotype_MD).

Qualitative culture results obtained from matched BALF and bronchial aspirate (BA) on selective media further reinforced the genuine existence of the four pneumotypes (Fig. 3j, Supplementary Fig. 5). BALF samples from Pneumotype_Balanced had the highest percentage of matches to the oropharyngeal microbiota, which consists of many of the bacteria predominant in this pneumotype (e.g., Streptococcus or Veillonella). Similarly, culture results of BALF samples with Pneumotype_{Staphylococcus} and Pneumotype_Pseudomonas were most frequently positive for S. aureus/Corynebacterium spp. and P. aeruginosa, respectively, while those obtained for BALF samples with Pneumotype_MD were often culture negative (Fig. 3j, Supplementary Fig. 5a). A similar picture was observed for BA (Fig. 3j, Supplementary Fig. 5b), where however a higher percentage of positive cultures for oropharyngeal flora was observed compared to BALF, especially for Pneumotype_MD. This suggests differences in the microbiota between the two sample types, despite the known topographic continuity of microbial communities in the airways^1,3,7. This is supported by the fact that for Pneumotype_{Staphylococcus} and Pneumotype_Pseudomonas, there was no association between pre- and post-transplant colonization by S. aureus (χ² = 0.047, P = 0.82) and P. aeruginosa (χ² = 0.2, p = 0.65) (Supplementary Fig 5c, d), respectively, in the upper respiratory tract.

Many of the OTUs of Pneumotype_MD could not be cultured in our bacterial culturing approach (Supplementary Table 3, 5, Supplementary Fig. 5), which together with the low bacterial abundance in corresponding samples, questions their relevance/existence as lung microbiota members. In contrast, most of the major community members characteristic of the other three pneumotypes were represented by isolates in LuMiCol, including the two opportunistic pathogens, P. aeruginosa (OTU_1) and S. aureus (OTU_2), providing the basis for future experimental work on the predominant species of these pneumotypes. Taken together, we identified four distinct bacterial communities in transplanted lung, which we refer to as pneumotypes, and validated them by culturing of BALF samples.

Bacterial pneumotypes are linked to distinct host gene expression patterns

The existence of bacterial pneumotypes with distinctive community composition suggests differences in the microenvironmental conditions of the human lung post-transplant, which could be echoed in other constituents of the lung ecosystem. We compared the median expression levels of 31 host genes belonging to seven functional categories across the four pneumotypes. These genes are involved in inflammation, immunoregulation, tissue remodeling and detection of bacteria and viruses (See “Methods”, Fig. 4a), and their expression patterns differed across the four pneumotypes, with particularly high transcriptional activity in Pneumotype_{Staphylococcus} (Fig. 4a). To identify the genes with the greatest power to discriminate between the four pneumotypes and between samples differing by bacterial counts, we applied a machine learning approach (Random Forest, See “Methods”) based on the host gene expression in 234 BALF samples. The Pneumotype_Balanced was predicted with highest accuracy (92%), followed by Pneumotype_Pseudomonas (83.4%) and Pneumotype_MD (81.4 %), while no accuracy was achieved for Pneumotype_{Staphylococcus} (Supplementary Table 5). We identified 6 of the 31 genes to have a particularly high predictive power IFNLR1, MRC1, IL10, IL1RN, LY96, IDO (Importance score >10; 99% Confidence Interval, Fig. 4b). IFNLR1 encodes interferon lambda receptor 1, which is involved in antiviral defence and epithelial barrier integrity⁴¹. This gene was up-regulated in samples with Pneumotype_Balanced compared to the other three pneumotypes (Fig. 4c). MRC1 (Mannose Receptor C-Type 1)⁴² and LY96 (Lymphocyte Antigen 96)⁴³ encode microbial polysaccharide and lipopolysaccharide recognition proteins⁴³, respectively. Compared to samples with Pneumotype_Balanced, these two genes were up-regulated in Pneumotype_{Staphylococcus} and Pneumotype_MD, and down-regulated in samples with Pneumotype_Pseudomonas (Fig. 4d, e). Samples with Pneumotype_Balanced further differed from those linked to the other three pneumotypes by higher expression of genes involved in immune modulation and peripheral immune tolerance (IL-10/Interleukin 10 and IDO1/Indoleamine 2,3-Dioxygenase 1, Fig. 4f, Supplementary Fig. 6), and a lower expression of IL1RN (Interleukin 1 Receptor Antagonist, Fig. 4g), produced as part of the inflammatory response to control the potentially deleterious effects of Interleukin-1 beta (IL-1β)⁴⁴.

Similarly, we found five genes with high discriminating power (Fig. 4i, importance score >10) for bacterial counts, of which two were particularly good predictors: PDGFD and IFNLR1. PDGFD encodes the D isoform of platelet-derived growth factor, which promotes the proliferation of cells of mesenchymal origin such as fibroblasts⁴⁵. Expression of this gene was negatively correlated (Fig. 4j, AIC 62.4, p < 0.001) with bacterial abundance. In contrast, IFNLR1 expression positively correlated with bacterial abundance (Fig. 4k, AIC 35.7, p < 0.05). Accordingly, PDGFD expression was higher while IFNLR1 expression was lower in Pneumotype_MD (Fig. 4c, h) as compared to the other pneumotypes, suggesting a link between the normal presence of bacteria in the lower respiratory tract and homeostatic levels of tissue remodeling, epithelial barrier integrity and host response to viruses. In summary, these results show that host-specific gene expression markers align with distinct bacterial states, highlighting the existence of complex associations between different lung ecosystem characteristics.

Anellovirus dynamics is associated with bacterial community and host physiology in lung

The observed links between pneumotypes and antiviral defence prompted us to look into the tripartite interactions between lung bacteria, viruses, and host. To this end, we quantified the load of the three genera of anelloviruses identified in humans (Alphatorquevirus, Betatorquevirus, and Gammatorquevirus) across the 234 BALF samples. In accordance with a previous study⁴⁶, we found that the transplanted lung contains high levels of anelloviruses, with Gammatorquevirus predominating. Viral loads of the three genera peaked between 1.5 and 6 months after transplantation and decreased at later time points (Fig. 5a). Anellovirus load varied substantially between pneumotypes (Fig. 5b–d). All three viral genera were lowest in samples with Pneumotype_MD and Alphatorquevirus showed particularly high levels in samples with Pneumotype_Pseudomonas (Fig. 5b). Intra-individual pairwise analysis revealed a particularly strong decrease in load for longitudinal transitions from Pneumotype_Balanced to Pneumotype_MD and a corresponding increase for the inverse (Fig. 5e, Supplementary Fig. 7). We further identified four human genes: TLR3, IGF1, RSAD2, IFITM2, as important predictors of anellovirus loads in BALF (Fig. 5f, See “Methods”). Of these, Toll-like Receptor 3 (TLR3) was positively correlated with total viral load (Fig. 5g, AIC 73.9, p < 0.001). This is consistent with the low viral load observed with Pneumotype_MD, where TLR3 was down-regulated (Fig. 5h). These findings link changes in the lung microbiota composition to changes in viral loads and host gene expression indicating possible implications for allograft outcome.

Pneumotypes are linked to differential risk of post-transplant clinical complication

A large set of clinical data (Supplementary Data 7) enabled us to associate differences in bacterial community composition, host gene expression, and anellovirus load to allograft and patient health status. Immunosuppression as well as prophylactic and therapeutic antibiotic usage were anticipated as major confounding factors. However, we found no association between the different pneumotypes and the main immunosuppressive drugs (ANOVA, prednisone; P = 7.68 × 10⁻¹, tacrolimus; P = 9.1 × 10⁻¹) used in our cohort, at the time of BALF sampling (Supplementary Fig. 8a, b). In contrast to what has been reported for blood plasma after transplantation⁴⁷, we also did not observe a correlation between anellovirus load and immunosuppressive drug levels (Lm, prednisone; P = 8.0 × 10⁻¹, tacrolimus; P = 9.1 × 10⁻¹, Supplementary Fig. 8c, d, See “Methods”). However, we observed a negative relationship between the number of antibiotics administered at the time of BALF sampling and the fraction of samples in Pneumotype_Balanced, and a positive relationship with the fraction of Pneumotype_MD samples (Fisher’s test, P = 2.0 × 10⁻³, Fig. 6a). These observations thus suggest a link between intensive antibiotic use and a disturbance of the most balanced and compositionally stable lung microbiota profile.

We observed that a clinical diagnosis of infection was rare in the presence of Pneumotype_Balanced and Pneumotype_MD, compared to Pneumotype_{Staphylococcus} and Pneumotype_Pseudomonas (Binomial linear model, Yes or No, P < 0.001 and P = 0.016, respectively; Fig. 6b). This confirms the results of our 16S rRNA gene analysis, which showed that Pneumotype_{Staphylococcus} and Pneumotype_Pseudomonas are dominated by opportunistic pathogens, S. aureus and P. aeruginosa, respectively. It is also consistent with the finding of lower numbers of neutrophils (Fig. 6c), but not macrophages (Fig. 6d), in Pneumotype_Balanced and Pneumotype_MD as compared to Pneumotype_{Staphylococcus} and, to a lesser extent, Pneumotype_Pseudomonas, emphasizing that pneumotypes are associated with local conditions that differ in terms of recruitment of pro-inflammatory cells.

Lung transplant recipients face risks of allogeneic responses against the graft, notably promoted by clinical infection. Our study did not have the statistical power to dissect the links between pneumotypes and different types of rejection, limited by the number of samples per rejection category in our dataset. Therefore, we grouped 29 samples from 17 patients with either CLAD, acute cellular rejection grade ≥2, or the presence of donor-specific antibodies (mean fluorescence intensity >1000), as these all indicate a suboptimal control of host immune competence and thus an increased probability of allograft injury (Fig. 6e, See “Methods” for clinical definitions). The majority of these samples were associated with Pneumotype_{Staphylococcus} (41.7%) and Pneumotype_MD (26.2%), followed by Pneumotype_Pseudomonas (15.4%) and Pneumotype_Balanced (13.1%), suggesting that this latter microbiota profile is associated with a lower risk of clinical complications. This was further corroborated by the count of circulating B lymphocytes in peripheral blood, suggesting more active humoral immunity in the presence of Pneumotype_{Staphylococcus}, Pneumotype_Pseudomonas and Pneumotype_MD, compared to Pneumotype_Balanced (Tukeys test, P = 2.7 × 10⁻², P = 2.6 × 10⁻¹ and P = 1.2 × 10⁻¹ respectively; Fig. 6f). In addition to bacterial composition, anelloviruses were also linked to CLAD through a lower load of Gammatorquevirus (Wilcoxon test, P = 7.0 × 10⁻³, Fig. 6g), while no significant association was observed with Alphatorquevirus or Betatorquevirus (Wilcoxon test, P = 1.5 × 10⁻¹ and P = 9.0 × 10⁻² respectively, Fig. 6g).

Finally, we used the measurement of ‘Forced Expiratory Volume in 1 s’ (FEV1) to search associations between lung ecology and pulmonary function testing. This assessment was made irrespective of the diagnosis of CLAD, which requires an irreversible drop in FEV1 below 80% of the baseline value, with prior exclusion of alternative confounding diagnosis (See “Methods”). Pneumotype_{Staphylococcus} and Pneumotype_Pseudomonas were associated with lower FEV1 values overall, with a frequent substantial decline below 80% predicted (Dunn’s test, P = 3.0 × 10⁻², Fig. 6h), while Pneumotype_Balanced, along with Pneumotype_MD, was linked to preserved lung function.

Pneumotype_Balanced shows the highest temporal stability and resilience in the transplanted lung

Taking advantage of the longitudinal sampling, we explored the dynamics of the pneumotypes after transplantation. We analyzed transitions between pneumotypes in up to eight BALF samples per transplant, collected within five consecutive time windows (Fig. 7a). There was no significant difference in the distribution of pneumotypes across the different time windows (χ² test, P = 6.0 × 10⁻¹). Although most BALF samples were associated with Pneumotype_Balanced, transitions between two different microbiota pneumotypes occurred for about half of all consecutive sample pairs (Supplementary Fig. 9). The transition dynamics were explained by Markov chain properties, i.e., the pneumotype of a given sample only depends on the state of the previous sample in the chain (χ² test, P = 3.3 × 10⁻¹, Fig. 7b). The transitions were irreducible, aperiodic and recurrent, and none of the pneumotypes behaved as an absorbing state (See “Methods”). Pneumotype_Balanced exhibited the greatest stability, with the highest probability of recurrence (63%), fitting the Markov probabilities, followed by Pneumotype_MD (42%; Fig. 7b). In addition, the large fraction of transitions toward Pneumotype_Balanced between the first four time windows indicated a substantial resilience capacity for this profile. Accordingly, the transitions between Pneumotype_{Staphylococcus}, Pneumotype_Pseudomonas and Pneumotype_Balanced occurred mainly in the direction of this latter profile (Fig. 7b), while in contrast to model prediction, Pneumotype_{Staphylococcus} and Pneumotype_Pseudomonas appeared to be virtually disconnected.

Finally, we illustrate the relationship between the temporal dynamics of pneumotypes and clinical outcomes using a case study (Fig. 7c). Patient 35 diagnosed with pulmonary fibrosis received two transplants, providing 12 serial samples and presenting each of the four pneumotypes (Fig. 7c). Disruption of the Pneumotype_Balanced occurred from month 25, followed by transition to Pneumotype_{Staphylococcus} at month 30. This was accompanied by a positive culture for Corynebacterium spp. in line with our enrichment analysis and clinical culture tests (Supplementary Fig. 4, Fig. 3j), increased BAL neutrophilia, and a concurrent increase in host immune gene expression (heatmap in Fig. 7c). Thereafter, the patient was repeatedly exposed to antibiotics, and respiratory function started declining irreversibly leading to the diagnosis of CLAD at month 49 with Pneumotype_MD. Overall decrease in bacterial and anellovirus loads in the lung, suggested an increasing selection pressure on microbes most likely due to a combination of antibiotic treatment and poorly controlled host immune competence. The second transplant at month 50 was linked to a re-establishment of Pneumotype_Balanced, which aligned with preserved lung function, intermediate loads of lung bacteria and anelloviruses, decrease in neutrophil counts and change in host immune gene expression. However, a transition to Pneumotype_Pseudomonas was observed later until the end of sampling, with increased bacterial counts but no decrease in lung function (barplot in Fig. 7c). Taken together, these observations highlight the potential of integrating pneumotype with clinical and molecular data, with the primary goal of tracking disruption of Pneumotype_Balanced beneficial to clinical stability.

Source: Ecology - nature.com