From April 2020 to October 2021, patients suspected of lower respiratory tract infections (LRTI) were screened for SARS-COV-2, and SARS-COV-2 negative patients were enrolled at Haikou Third People’s Hospital in Hainan, China. The inclusion criteria were as follows: abnormal imaging findings, such as pulmonary shadows, space-occupying lesions, and other signs of pulmonary infection, combined with≥1 following items: 1) fever >37°C; 2) symptoms such as cough, expectoration, or dyspnea; 3) leukocytosis; 4) clinical signs of lung consolidation or moist rales. The exclusion criteria were: 1) refuse to participate in the study or fail to provide BALF samples; 2) recent antibiotic usage within a month. The study design was displayed in Figure 1A . The diagnosis of LRTI was made by two attending physicians based on a post hoc comprehensive approach which combined clinical adjudication and microbiological tests results. Baseline characteristics of all patients were collected on enrolment, including age, gender, temperature on admission, and underlying diseases. Blood routine tests results, C-reactive protein and procalcitonin levels, imaging findings and antibiotic regimens were recorded during hospitalization. The antibiotic treatment response, length of hospitalization and patient outcomes were also investigated. This study was performed in accordance with the Declaration of Helsinki. Ethics approval for this human study was approved by Haikou Third People’s Hospital Ethics Committee ([2020]K001). All adult participants or surrogates provided written informed consent to participate in this study.

All patients received antibiotic treatment immediately after hospitalization. Before that, bronchoalveolar lavage fluid (BALF) specimens were collected with an ultrathin bronchoscope (BF-P-260F, Olympus), and then were divided into two parts and sent for culture and mNGS (Biotecan, Shanghai) in a pairwise manner. Blood samples and sputum samples were also collected simultaneously for conventional microbiological tests (CMT), including culture, smear, T.SPOT, and acid-fast bacteria stain (AFB stain).

The workflow of mNGS sequencing consisted of a wet lab pipeline and in silico pipeline ( Figure 1B ). For the web lab workflow, microcentrifuge tubes containing 0.5 mL BALF sample and 1 g glass beads of 0.5 mm were first attached to a horizontal platform on a vortex and agitated vigorously at 2800-3200 rpm for 30 min. Then, separate 0.3 mL of the sample into a new 1.5 mL microcentrifuge tube, and extract DNA using a HostZEROTM Microbial DNA Kit (D4310, ZYMO RESEARCH) according to the manufacturer’s instructions. To prepare the library, more than 5 ng DNA extraction was digested into appropriate length (200–300 bp) followed by attaching the 3’ end with a Da tail and the 5’ end phosphorylation. Next, connect the fragmented DNA to the adapter sequence with DNA ligase, and remove the splice dimers, redundant splices, and residual reagents with purification beads. Finally, all sample DNA libraries with a concentration of more than 1 nmol/L were mixed and sequenced with an Illumina NextSeq CN500 sequencer (SE strategy, read length=75). In this study, internal control, negative control, and positive control were used for each run. Before nucleic acid extraction, specific molecular tags were placed in the sample and used as internal parameters to track the whole process and control the quality of the workflow. No-template water was applied as a negative control to monitor sample-to-sample contaminations. Tests should be repeated when the negative control failed. The clinical samples with confirmed pathogens served as a positive control, which indicated failures of mNGS when the targeted pathogens failed to be detected in these samples.

For the in silico analysis, clean sequencing data were filtered with Fastp software (v0.23.1) by the following steps: 1) removing connector sequences from reads and reads that have no insert fragments due to connector self-connection; 2) removing low-quality (length <20 bp) reads at 3’ end; 3) removing reads that contain over 10% of N; 4) removing Adapter and sequences shorter than 15 bp after trimming. Seqtk_sdust (v1.3-r106) was utilized to calculate the complexity of each read and low complexity (<0.3) sequences were filtered out. By mapping the human host sequences to the human reference genome (grch38.p12) using bowtie2 (v 2.3.4.1), the remaining human sources of reads were removed. After removing the plasmid sequence and chromosome incompletely assembled sequence, species with completely assembled genomes and chromosomes were included in the pathogenic microorganism database. Then, map the microbial reads via bwa (v0.7.15) and classify pathogenic bacteria with whole genome alignment by Kraken software. The Pathogenic microbial genomic data originates from NCBI refseq (ftp://ftp.ncbi.nlm.nih.gov/genomes/refseq/) and PATRIC (https://www.patricbrc.org/). Finally, all detected pathogens were calculated for alignment sequence number, relative abundance, and genome coverage.

For Antibiotic Resistance Genes (ARGs) analysis, the sequences of pathogens were aligned to the CARD (the Comprehensive Antimicrobial Resistance Database) database, an established drug resistance gene database, by blast (v 2.2.26). Sequences with over 90% of similarity identity and more than 70 bp in length were retained. The coverage of the ARGs was calculated according to the starting position of the comparison, and ARGs with coverage ≥80% were defined as positive.

First, bacteria suspected of colonizers were excluded referring to the in-house background microbial database. Then, the remaining bacteria (mycobacteria excluded), viruses, and parasites undergo the following filtering criteria:

The responsible pathogens for LRTIs were assessed by two attending physicians in our hospital based on the microbiological results, clinical features, and response to the treatment. Treatment response was evaluated based on a comprehensive adjudication considering the patient’s symptoms, laboratory indicators, and radiological signs.

Normally distributed continuous variables are recorded as the mean ± standard deviation and nonnormally distributed were as the median and range. Categorical variables are reported as numbers and percentages. The McNemar-test was used for comparisons of the diagnostic performance between two diagnostic methods. For comparison of the clinical laboratory indicators before and after treatment, a paired t-test was applied. All tests were two-tailed, and statistical significance was set at p<0.05. SPSS software (version 27.0; IBM Corporation, Armonk, NY, USA) was used for the statistical analysis.

In this prospective observational study, a total of 276 patients suspected of LRTIs were screened. After excluding 10 cases, consisting of two non-infectious cases and eight repeated cases, there were 266 patients included for analysis. The demographic features of these patients were presented in Table 1 . The average age of all patients was 67.25 years old. Among them, 177 (66.5%) were males and 89 (33.5%) were females. Patients with immunodeficiency accounted for 12.8%. There were 63 patients presented with fever, ranging from 37.2 °C to 40.8 °C. There were 136 patients accompanied with underlying disease, most of them were bronchiectasis (22.9%) and chronic obstructive pulmonary disease (COPD, 19.9%). Empirical antibiotics were given after BALF collection and before microbiological results.

BALF samples were collected from patients and divided into two parts for mNGS sequencing (n=266) and culture (n=160), respectively. Other samples were also collected for culture if needed, including sputum (n=235), and blood (n=104). Meanwhile, other conventional microbiological tests (CMT) were also performed with sputum samples and blood samples, including smear (n=238), AFB stain (n=240), and T. SPOT (n=50). The distribution of sample types and microbiological tests is shown in Figure 2A .

As shown in Figure 2B , among 257 patients with accessible culture results, the positive rate for mNGS (87.55%) was higher than culture (39.30%) (p<0.001). Likewise, among 264 patients who were tested by CMT, the diagnostic performance of mNGS was still higher than CMT (p<0.001), with a positive rate of 87.12% versus 52.65%. Among all patients enrolled, a total of 134 cases (50.38%) were positive for mNGS only, 97 cases were positive for both mNGS and culture, and only 4 cases were positive by culture only ( Figure 2C ). Among the 97 double-positive samples, 88.66% of them showed matched results between mNGS and culture completely (44 cases) or partly (42 cases). As for the 160 cases who had both mNGS and culture tests with BALF samples, the positive rate of mNGS was still superior to culture (88.75% vs 26.88%, p<0.001). A total of 43 cases (26.87%) were positive for both mNGS and culture, and 83.72% of them showed matched results between mNGS and culture completely (22 cases) or partly (14 cases), displaying a relatively high consistency between mNGS and culture results ( Supplementary Figure 1 ).

A total of 208 microbes were isolated by mNGS, among which 36 bacteria, 12 fungi, and 5 viruses were considered underlying pathogens. For bacteria, MTB was the most detected pathogen (n=40), followed by Streptococcus pneumoniae (n=37), Pseudomonas aeruginosa (n=32), Haemophilus influenzae (n=26), Klebsiella pneumoniae (n=27), and Acinetobacter baumannii (n=15). The mNGS detected more pathogens than culture in terms of MTB (mNGS only), virus (mNGS only), S. pneumonia (13.91% vs 2.25%, p<0.001), P. aeruginosa (12.03% vs 9.02%, p<0.05), H. influenzae (9.77% vs 2.26%, p<0.001), Aspergillus fumigatus (3.00% vs 0.75%, p<0.05), and Candida albicans (26.32% vs 7.52%, p<0.001) ( Figure 2D ). Notably, in cases where culture was negative, mNGS showed advantages in detecting uncommon pathogens, including Chlamydia psittaci (n=1), Cryptococcus neoformans (n=4), Pneumocystis jirovecii (n=2), Nocardia otitidiscaviarum (n=1), Mycobaterium paragordonae (n=1) as well as Mycobaterium simiae (n=2). Besides, both mNGS and culture isolated Burkholderia pseudomallei from two patients.

Since all cultures failed to report MTB till the data was collected, the diagnostic performance for mNGS and other conventional methods, including T-SPOT and AFB stain, were compared with the final diagnosis of MTB. The final diagnoses of MTB were made by the attending physicians according to the China Clinical Diagnosis Guide for Tuberculosis (National Health and Family Planning Commission of the People’s Republic of China., 2017). The sensitivity of mNGS for diagnosing MTB was 80%, which was greatly higher than T-SPOT (28%, P<0.05) and AFB stain (26%, P<0.01) ( Table 2 ). The diagnosing specificity of mNGS was 97.4%, similar to the T-SPOT (89.6%) and AFB stain (93.0%).

Based on the mNGS result, 147 (55.26%) cases modified the initial antibiotic treatment, including 74 cases of adjusted antimicrobial regimens, 60 patients escalated the treatment by increasing dosage or medication, and 13 cases de-escalated the EAT. Among 60 patients who escalated the antimicrobial treatment, 26 added anti-MTB meds, 26 escalated with anti-fungi treatment, and 2 with anti-virus treatments ( Figures 3A, B ). The treatment responses and prognosis of patients with mNGS-guided antibiotic adjustment were displayed in Figure 3C . There were 89.12% of patients responding to modified regimens and 75.51% of patients with good prognosis. After adjusting treatments guided by mNGS results, clinical laboratory indicators, including white cell counts (WBC), neutrophils (N), and C-reactive protein (CRP), declined remarkably when compared to baseline (all p<0.001) ( Figure 3D ). Notably, all 13 patients with a de-escalated adjustment also responded to treatment and the overall prognosis were well.

However, five patients died during hospitalization despite positive results by mNGS and culture and had modified antimicrobial treatment accordingly. They responded poorly to the antibiotic treatment, and three of them had identical mNGS and culture results. Therefore, the antimicrobial resistance of these patients was analyzed retrospectively by mNGS workflow and compared with the antibiotic susceptibility test (AST) if applicable. A total of 38 antibiotic-resistant genes (ARGs) were detected in a patient confirmed of P. aeruginosa infection ( Supplementary File 1 ) and 25 ARGs in a patient confirmed of A. baumannii infection ( Supplementary File 2). No ARG was detected in a patient confirmed of C. albicans infection. For the patient with P. aeruginosa infection, AST confirmed carbapenem-resistant P. aeruginosa (CRPA) infection and eight ARGs were predicted to be resistant to carbapenems and polypeptides, including ArmR, MexA, MexB, mexP, mexQ, mexY, opmE, and OprM. Besides, for the patient with A. baumannii infection, AST confirmed carbapenem-resistant A. baumannii (CRAB) infection, and six ARGs were detected to be resistant to carbapenems and cephalosporins, including adeJ, adeK, msrE, adeI, abeM and adeN ( Figure 3E ).

All patients demonstrated abnormal chest CT images after enrollment, with 184 (69.17%) bilateral lesions and 82 (30.83%) unilateral lesions. Atypical pathogens and some less-common pathogens usually cause various appearances in CT scans ( Figures 4A–D ), but mNGS can detect unbiasedly and guide clinical treatment. A 63-year-old patient presented with cough, fever, shortness of breath, and multiple patchy rashes over the body. CT image indicated large consolidations in the upper right lobe and mNGS reported Chlamydia psittaci and MTB. The lesions were remarkably absorbed after antibiotic treatment based on mNGS results ( Figures 4E, F ). A 59-year-old patient had abnormal laboratory test results and bilateral multiple patchy consolidations in the chest CT scan. Empirical antibiotic treatment failed and was stopped after mNGS reported Cryptococcus neoformans infection, the patient was immediately changed into an anti-fungi regimen. Twelve days later, the laboratory indicators returned to normal, and the consolidated nodules shrunk obviously compared to that at hospitalization ( Figures 4G, H ).