This retrospective analysis was performed for the patients with suspected lung infections after cardiac surgery in Ruijin Hospital. In this study, the postoperative patients were enrolled based on abnormal imaging changes represented by chest X-ray, including chest infiltration, consolidation, cavity, and pleural effusion. Pulmonary infection was determined if at least two of the following three options were satisfied: 1) fever greater than 38℃, 2) the number of white blood cells beyond the routine detection limits, 3) occurrence of purulent airway secretion. Sputum specimens of the patients were collected from December 2020 and June 2021 ( Table S1 ).

According to the cultivation procedure for bacteria and fungi in the diagnostic laboratory of Ruijin Hospital, the routine isolation media were used, including blood agar, eosin methylene-blue (EMB) agar, chocolate agar, Mueller-Hinton agar, sabouraud dextrose agar, and cooked meat medium (Oxoid, UK). Chocolate agar and EMB agar plates were incubated in 5-10% CO₂ at 37°C for 48-72 hours; blood agar, Mueller-Hinton agar, and cooked meat medium plates were incubated at 37°C for 24-48 hours. Sabouraud dextrose agar plates suitable for the isolation of Candida albicans were incubated at 28°C and 35°C for 24-48 hours, respectively. Species identification of isolated strains was then carried out using the VITEK-2 Compact Instrument (bioMérieux, France) (Melhem et al., 2014).

For the preprocessing of specimens, sputum samples were homogenized by the dithiothreitol (DTT) treatment as proposed by Terranova et al. (Terranova et al., 2018). Approximately 200 μL sputum per sample was resuspended by 600 μL of DTT solution (1% concentration) and 20 μL of proteinase K (20 mg mL^-1), followed by incubation at 56°C for 20 mins. Cells were removed by centrifugation to reduce host-background nucleic acid. Total DNA was extracted using HostZERO Microbial DNA Kit (Zymo Research, USA) following the manufacturer’s instructions. DNA extraction yield was quantified using a Quant-iT dsDNA HS Assay Kit and Qubit 3.0 Fluorometer (Thermo Scientific, USA). Enzymatic shearing was employed for the fragmentation (~200 bp) of DNA molecules and the libraries were then constructed using the Nextera XT DNA Library Preparation Kit (Illumina, USA). A certain number of PCR cycles (5-7) were applied to produce 1 µg of the metagenomic library. The quality of the libraries was assessed by a 2100 Bioanalyzer using the High Sensitivity DNA Assay (Agilent Technologies, USA). Metagenome shotgun sequencing in a single-end 75-bp mode was performed using the NextSeq 500/550 High Output Kit (92 cycles) on an Illumina NextSeq 550 sequencer. The samples as the No-Template Control (NTC) were sequenced simultaneously to assess contaminations during the wet-lab experiments.

Raw sequencing data were first subjected to a quality control process for trimming adapter sequences, removing low-quality tails and reads by Trimmomatic v0.36 (Bolger et al., 2014). Next, the reads mapping to the human reference genome GRCh37 were excluded using the short-read alignment tool Bowtie v2.2.6 (Langmead and Salzberg, 2012). To remove identical duplicated reads considered as technical artifacts by PCR (Martin et al., 2018), de-duplication was performed using an in-house Perl script. Taxonomic classification of microbial reads was conducted using Kraken v2.0.9-beta (Wood et al., 2019) and a custom k-mer database constructed using 51,543 genomes of ~27,000 species from the NCBI assembly databases (Kitts et al., 2016). To produce species-level abundance estimates, the number of reads in the Kraken classification report was further estimated by the Bayesian algorithm implemented by Bracken v2.2 (Lu et al., 2017). The microbial taxa bearing putative contamination during the experiments were removed from the final profiles based on the information of NTC samples in each batch: the ratio of read count (sputum sample)/read count (NTC) less than 10 for a specified species. Using a similar way to TPM (transcripts per million) described by Li et al. (Li et al., 2009), the estimates of percentage relative abundance of each species were computed for bacteria, viruses (excluding phages), and fungi, respectively.

To investigate pathogenic signals in the sputum microbiome, we proposed a rule of thumb to evaluate detection limits of pathogenic risks for individual organisms through species-level abundance profiling. Briefly, the values of the mean and standard deviation of relative abundances were calculated according to the estimates of each species across the samples. In this study, limits of detection (LOD) for prompting pathogenic potential of a specified species were estimated through the filtering criteria below: read count > 100; relative abundance > 0.05; presence in at least three samples; the mean plus the standard deviation of the abundance values as the LOD of the pathogenic risk.

Strain-level profiling analysis was conducted based on single-nucleotide variants (SNVs) using the StrainPhlAn v3.0 package (Truong et al., 2017). Briefly, metagenomic reads per sample were first aligned to the marker gene database of MetaPhlAn v3.0 (Beghini et al., 2021). The sample-specific consensus sequences from the markers were reconstructed for each species’ most abundant strain and the species-specific markers were extracted from MetaPhlAn for identifying the homologous sequences in the reference isolate genomes using blastn (Altschul et al., 1990). Based on sequence alignment, maximum-likelihood phylogenetic trees were reconstructed using RAxML (Stamatakis, 2014) and then visualized using MEGA10 (Kumar et al., 2018). Genome sequences of the isolated strains were retrieved from the NCBI Assembly database and were used for phylogenetic reconstruction together with the strains detected in the metagenomic samples. The program MetaMLST v1.2.2 was employed to predict the known and novel Sequence Type (ST) of the most abundant strains in metagenomic samples (Zolfo et al., 2017). Since it is complicated to resolve the phylogeny of the diploid fungi, an alternative pipeline was used to build the phylogenetic tree according to multilocus sequence typing (MLST) analysis of seven loci (Hirakawa et al., 2015b). Seven MLST allele sequences were reconstructed from the metagenomes using MetaMLST, and the related gene sequences from the isolate genomes were processed using FastMLST (Guerrero-Araya et al., 2021). Multiple sequence alignment was performed with MUSCLE (Edgar, 2004), and the phylogeny was inferred with RAxML.

The analysis of α-diversity and two sample T-test were conducted using the package fossil v0.4.0 (Chao, 1987) based on the Chao1 index representing community richness. To estimate the community compositional variation of the samples between groups, β-diversity was measured by calculating the Bray-Curtis distance matrix using Vegan v2.5-7 (Dixon, 2003). Differences in the rank dissimilarities within and between groups were inferred using the non-parametric ANOSIM test of 1,000 permutations. All statistical analyses and visualizations were performed in R v4.1.0 (R Core Team, 2021).

In this retrospective study, we analyzed the data of microbiological testing on sputum samples from postoperative patients that were diagnosed with pulmonary infection. During the 7-month study, the specimens were collected from 50 adults aged from 34 to 81 years with a median of 64 years. Approximately three quarters (n = 36; 72%) were male. Participant characteristics are listed in Table S1 , including body mass index (BMI), dates of both cardiac surgery and specimen sampling, patient outcomes, National Early Warning Score (NEWS) (Smith et al., 2013), and Sequential Organ Failure Assessment (SOFA) scores upon admission (Vincent et al., 1998). The median time of sample collection was day 3 after the cardiac surgery. The samples were then subject to conventional culture and mNGS testing for species identification. DNA mNGS testing yielded an average of 22.9 million reads for each specimen ( Table S2 ).

As shown in Table S1 , the culture testing reported 12 bacterial species and two fungal species across all the samples. It resulted in one species detected in each of the 16 sputum samples, two species in five samples, and three species in one sample ( Table 1 ). None of the microorganisms was detected in the remaining tested samples (n = 28; 56%) by the culture method. On the contrary, a substantial number of microorganisms were identified by mNGS testing. The numbers of microorganisms assigned to the domains bacteria, fungi, and viruses were 682, 58, 21, respectively. Particularly, the numbers of bacterial and fungal species detected by mNGS testing were much higher than those by culture in the individual samples ( Table 1 ). The average number of bacterial species in each sample was 90 with a range from 4 to 219. The average number of fungal species in each sample was 5 with a range from 1 to 17. Besides, the viral species were identified in more than half (n = 31; 62%) of the samples through mNGS; whereas the routine culture approach used in our study can support isolation and identification for bacteria and fungi, but not for viruses.

To assess the pathogenic risk of individual microorganisms, we proposed the abundance LOD for screening candidate pathogens based on mNGS assays and species-level abundance profiles of bacteria, fungi, and viruses, respectively ( Table S3 ). Most of the bacterial and fungal species detected by the culture approach were present in the list of the pathogens that were screened out in the individual samples ( Table 2 ; the full list shown in Table S1 ). For instance, both culture and mNGS testing for patient #2 identified B. cenocepacia and C. albicans, which were predominant pathogens in the bacterial (86.3% abundance) and fungal community (98.7%), respectively. For patient #44, the mNGS approach identified five pathogens C. striatum (56.2%), S. aureus (39.2%), C. albicans (93.1%), A. flavus (5.5%), and Human alphaherpesvirus 1 (99.8%). Of these, S. aureus and C. albicans were also detected by culture. According to the LOD (30.8%) of K. pneumoniae, this bacterium was assessed as a pathogen in the four patients: 85.4% for #1, 87.1% for #4, 97.7% for #16, and 72.1% for #40. The sputum culture was positive for K. pneumoniae in these patients except for #4. The mNGS approach for patient #4 identified K. pneumoniae (87.1%) and C. albicans (97.8%) as pathogens; whereas, none of them was detected by culture. C. albicans, an opportunistic pathogenic yeast, was the most prevalent fungal pathogen detected in the sputum metagenomes of 15 patients ( Table S1 ). In addition, human alphaherpesvirus 1 was the most frequently detected pathogen that was identified in six patients. Generally, according to our proposed mNGS screening criteria, the predicted microbial targets bearing pathogenic risk not only well matched the ones detected by culture, but also provided additional and prioritized pathogens missed by culture. It should provide complementary evidence for microbiological diagnosis in the clinical setting.

To explore organismal composition in the sputum microbiota, the mNGS dataset containing all sequenced metagenomes was analyzed. We focused on the diversity of bacterial communities bearing high species richness compared to the fungal and viral communities ( Table 1 ). The species accumulation curve tended to be smooth, indicating that the number of bacterial species would vary slowly with the increase in sample size ( Figure 1A ). The profile of percentage taxonomic abundances across the sputum metagenomes was displayed in Figure 1B . According to the average values of abundances, the top five abundant species were K. pneumonia (7.3%), S. maltophilia (5.6%), S. epidermidis (4.9%), S. infantis (4.8%), and A. baumannii (3.6%) ( Figure 1B ). It was observed that certain species that were remarkably dominating in the community of the individual samples were estimated to be etiologic agents mentioned above, such as K. pneumonia in #1/#4/#16/#40, A. baumannii in #25/#36, N. flavescens in #9/#17/#46, and S. anginosus in #13/#19/#21/#24/#26 ( Table S1 ). Next, we compared the community structural dynamics of the sputum microbiota between two clinical outcomes: Recovery and Worse. The α-diversity analysis of community richness indicated that no significant difference (p = 0.125) was found between the two groups ( Figure 1C ). Based on the β-diversity ANOSIM test (R = 0.173, p = 0.008), it was apparent that the mean of ranked dissimilarities between groups was lower than that within the group Worse ( Figure 1D ). Meanwhile, the low R-value suggested a weak effect of clinical outcomes on the microbial communities. Both α- and β-diversity analyses indicated that the bacterial community diversity was less associated with clinical outcomes of postoperative infected patients, probably due to diverse pathogens dominating the individual samples.

Strain identification is of significant importance for better understanding microbial pathogenicity and transmission of clinically relevant microorganisms (Nayfach et al., 2016). According to the pathogen list assessed by mNGS testing ( Table S1 ), strain-level profile analysis was performed for four opportunistic pathogenic species K. pneumoniae, C. striatum, S. aureus, and C. albicans. StrainPhlAn analysis displays strain-level phylogenetic trees visualizing the relatedness between the metagenome-recovered strains and clinical bacterial isolates with publicly genome sequences ( Figure 2 ). K. pneumoniae strains were reconstructed from the metagenomic samples of three patients #1/#4/#16 ( Figure 2A ). The analysis of metaMLST showed that the strain from patient #4 was K. pneumoniae of ST29 and the other two strains were represented by novel STs. Consistently, a K. pneumoniae strain derived from #4 was most closely related to a multidrug-resistant ST29 strain (CP024489/INF249) isolated from urinary tract infection (Gorrie et al., 2018) ( Figure 2A ). The strain from #16 was closely related to the K. pneumoniae strain NTUH-K2044 (AP006725) causing liver abscess and meningitis (Wu et al., 2009). In addition, two C. striatum strains from #27/#44 were phylogenetically placed together ( Figure 2B ), and their closest relatives were the two strains (VCOZ00000000 and VCOY00000000) of C. striatum isolated from influenza patients with secondary lower respiratory tract bacterial infections (Qin et al., 2020). Besides, the S. aureus strain from #44 was neighboring to the strain 55/2053, which caused the global pandemic of Panton-Valentine leukocidin-producing S. aureus in Europe and the United States in the 1950s (Zuo et al., 2021) ( Figure 2C ).

Figure 3 shows the phylogeny of three metagenomic strains of C. albicans from #13/#20/#29 together with the isolated strains of C. albicans, which is a well-known human fungal pathogen responsible for painful mucosal infections (Kim and Sudbery, 2011; Gulati and Nobile, 2016). Phylogenetic relationships of these fungal strains were investigated by the StrainPhlAn-based tree and the MLST-based tree, respectively. Both trees show that the C. albicans strain from #20 was closely related to the strains CHN1, Ca529L, A123, and A67.