This study retrospectively collected clinical data from 170 patients with suspected pulmonary infections who were treated at Hefei Cancer Hospital, Chinese Academy of Sciences, between January 2023 and December 2024. In the diagnostic cohort (n=120), sputum samples were tested in parallel by conventional culture, mNGS, and TNPseq for pathogen detection, while samples from the therapeutic cohort (n=50) underwent conventional culture and TNPseq testing. The inclusion criteria were: (1) clinical signs of pulmonary infection (fever, cough, sputum, dyspnea) or imaging abnormalities (pulmonary shadows, lesions); (2) available sputum samples for pathogen assays; and (3) complete clinical records. Exclusion criteria were: (1) non–LC tumors; (2) absent sputum samples; (3) hypersensitivity, known drug resistance, unresolved extrapulmonary infections, complex comorbidities, or poor sample quality. After excluding 26 non–LC patients and one without sputum, the final diagnostic cohort included 111 patients (63 LC and 48 BPD), and the therapeutic cohort included 32 patients (14 LC and 18 BPD) (Supplementary Figure S1).

This study was approved by the Medical Ethics Committee of the Hefei Cancer Hospital, Chinese Academy of Sciences (PJ-KY2023-092) and conducted in accordance with the principles of the Declaration of Helsinki (as revised in 2013) and relevant ethical and legal requirements. The requirement for written informed consent was waived because the study involved a retrospective analysis of residual specimens from standard clinical care.

Patients fasted for ≥1 hour prior to sample collection. Fresh expectorated sputum was aseptically transferred into sterile tubes (Feigelman et al., 2017). A qualified lower respiratory sputum specimen meets the following criteria: Squamous epithelial cells ≤10/Low Power Field (LPF) and Leukocytes ≥25/LPF. Culture specimens were processed within 30 minutes in the institutional microbiology lab; sequencing aliquots were stored at −80 °C and shipped to Hangzhou Shengting Medical Technology Co., Ltd. Both sites followed national pathogen-detection protocols.

Standard culture techniques were employed to identify potential pathogens. Sputum specimens were inoculated onto selective media (Babio Biotechnology, Jinan, China) under aseptic conditions: blood agar, MacConkey agar, and chocolate agar for bacterial culture; sabouraud dextrose agar for fungal culture. Bacterial cultures were incubated at 36.5 ± 0.5 °C with 5% CO₂ for 18–72 hours, whereas fungal cultures were incubated at 28.0 ± 0.5 °C for up to 7 days. Pathogen identification was performed with the VITEK 2 Compact system. Viral cultures were not conducted due to biosafety constraints.

Nucleic acids extracted from sputum samples were used as templates. Sputum samples were processed according to the following protocol: 400μL of each sputum sample was liquefied with 2% NaOH, centrifuged, then washed twice with PBS before resuspension in lysis buffer. Mechanical lysis was performed with grinding beads, followed by enzymatic digestion. DNA was extracted using the QIAamp DNA Microbiome Kit (QIAGEN), purified with magnetic beads, and eluted in nuclease-free water. The extracted DNA was quantified using a Qubit 4 fluorometer (Thermo Fisher Scientific, Massachusetts, USA). No-template controls (NTCs) with blank elution buffer were run in parallel.

For mNGS, libraries were prepared using the KAPA EVOplus Kit (Roche, Pleasanton, CA, USA) following the manufacturer’s instructions. The amplification was carried out on an ABI 2720 thermocycler with the following conditions: initial denaturation at 98 °C for 45 s, followed by 9 cycles of 98 °C for 15 s, 60 °C for 30 s, 72 °C for 30 s, and a final extension at 72 °C for 1 min. Libraries were quantified with the Qubit 4 Fluorometer (Thermo Fisher Scientific, Waltham, MA, USA) using the Qubit 1×dsDNA HS Assay Kit, and sequenced on the NovaSeq 6000 System (Illumina, San Diego, CA, USA) with a 2×150 bp paired-end protocol.

For TNPseq, PCR amplification targeting the 16S rRNA gene of bacteria, the 18S rRNA genes and ITS region of fungi, specific viral fragments (e.g., SARS-CoV-2 ORF1ab), and drug resistance genes was conducted using an ABI 2720 thermocycler. The amplification was carried out with the following conditions: initial denaturation at 95 °C for 3 min, followed by 30 cycles of 95 °C for 30 s, 62 °C for 60 s, 72 °C for 60 s, and a final extension at 72 °C for 3 min. Purified PCR products were quantified and prepared for library preparation using the PCR Barcode Expansion Pack 1-96 (EXP-PBC096, Oxford Nanopore Technologies, Oxford, UK). Sequencing was carried out on a GridION platform (Oxford Nanopore Technologies, Oxford, UK), followed by barcode demultiplexing with Porechop, short-read filtering via NanoFilt (2.7.1), and taxonomic classification of high-quality reads through alignment against pathogen reference databases using NCBI BLAST after NTCs filtering. The coverage includes 28,699 microbial species such as bacteria, fungi, viruses, and atypical pathogens, along with the detection of over 100 common antibiotic resistance genes and macrolide drug resistance genes based on SNP point mutations.

Contaminants were identified by comparison with blank controls, including the sequencing blank, ward environmental, and biosafety cabinet environmental controls. Subsequently, any contaminating organisms detected in the samples were filtered out following established protocols (Langelier et al., 2018; Li et al., 2021). Different pathogen identification thresholds are established based on the distinct principles and data characteristics of the two sequencing technologies: for mNGS data, ≥5 reads for fungi and bacteria, and ≥1 read for viruses and atypical pathogens (e.g., Mycoplasma Pneumoniae, Mycobacterium, Nocardia, and Aspergillus); for TNPseq data, ≥3 reads for bacteria and ≥1 read for other pathogens (including Mycobacterium and Nocardia).

Sequences derived from protozoa and parasites identified in sputum samples were systematically excluded. Filtered reads were aligned to the human reference genome (GRCh38/hg38) for host DNA removal. The remaining reads were subsequently aligned against relevant pathogenic sequences in NCBI-BLAST to analyze potential pathogens present in the sample.

To distinguish between colonizing microorganisms and pathogens, we utilized established pathogenic risk detection thresholds, considering read counts, relative abundance, frequency of occurrence, and relevant patient physiological indices (Maddi et al., 2019; Chen et al., 2022). We also referred to the Manual of Clinical Microbiology (11th edition) to determine pathogenicity and classification level of the pathogens.

A literature review was conducted to select 62 microorganisms (excluding RNA viruses) that are commonly detected in pulmonary infections of lung cancer patients (Supplementary Table S1) (Gazzoni et al., 2014; Maddi et al., 2019; Guo et al., 2021; Drokow et al., 2023).

Patients were classified based on their clinical diagnosis, which was primarily determined using microbiological examination, bronchoscopy, and radiography. Patients were categorized into LC and BPD groups. Within the LC group, patients were further classified into initial diagnosis lung cancer (IDLC) or non-initial diagnosis lung cancer (NDLC), depending on whether their LC diagnosis was newly established at our institution. Treatment responses were evaluated and classified according to the Response Evaluation Criteria in Solid Tumors (RECIST) as partial response (PR), stable disease (SD), or progressive disease (PD). Based on antibiotic treatment strategy, patients were stratified into either an empirical therapy group or a TNPseq-guided therapy group.

To evaluate diagnostic efficacy, sequencing results were compared against traditional diagnostic standards, including microbial culture and composite reference standards (CRS).

CRS status was independently adjudicated by two senior specialists: a chief physician in respiratory and critical care medicine and a chief technologist in clinical microbiology. They reviewed all conventional diagnostic data, including medical records, imaging, culture reports, serology/pathology, and treatment response. If the initial adjudications regarding a case’s infection status were discordant, a consensus discussion was initiated, strictly following current clinical guidelines. If disagreement persisted, a third senior clinical expert, who did not participate in the initial adjudication, was invited to serve as an arbitrator and make the final determination. A case was defined as CRS-positive (microbiologically confirmed infection) if it met the clinical diagnosis of pulmonary infection, satisfied ≥1 of the following criteria, with documented clinical improvement following targeted therapy (De Pauw et al., 2008; Rotstein et al., 2008): (1) Pathogenic bacteria are cultured from qualified lower respiratory tract secretions, protected specimen brushes via bronchoscopy, bronchoalveolar lavage fluid, lung tissue, or sterile body fluids, and the findings are consistent with the clinical presentation; (2) Fungal elements are observed through histopathology, cytopathology, or direct microscopy of lung tissue specimens, accompanied by relevant evidence of tissue damage; (3) Seroconversion from negative to positive for specific IgM antibodies against atypical pathogens or viruses, or a fourfold or greater increase in specific IgG antibody titers between acute and convalescent paired serum samples. Additionally, we provide an illustrative case of a patient who was culture-negative but was adjudicated as having Mycoplasma Pneumoniae infection based on CRS criteria (Supplementary Figure S2).

The α-diversity of microorganisms was evaluated using the Shannon index, ACE, Chao1, Richness, and Simpson index. β-diversity was analyzed using unweighted UniFrac, weighted UniFrac, Jaccard, and Bray-Curtis distances computed with QIIME2, and visualized using principal coordinate analysis (PCoA). Comparative pathogen composition across taxonomic levels was statistically analyzed using Statistical Analysis of Metagenomic Profiles (STAMP, v2.1.3) and the linear discriminant analysis effect size (LEfSe) method.

Statistical analyses were conducted using SPSS 17.0 (IBM, USA), GraphPad Prism 5.0 (GraphPad Software, USA), and R Studio 4.3.0. Continuous variables were reported as median with interquartile range (IQR). Categorical variables were summarized as frequencies with percentages. Comparisons between two groups utilized Student’s t-test for normally distributed variables and the Mann-Whitney U test for non-normally distributed variables. Categorical data were analyzed using Fisher’s exact test or chi-square test as appropriate. Differences among multiple groups were assessed using one-way ANOVA. A two-tailed p-value <0.05 was considered statistically significant.

A total of 143 patients were included, divided into diagnostic (n=111) and treatment (n=32) cohorts (Table 1). The diagnostic cohort had a median age of 66 years (IQR: 55.5-72), comprising 70.27% males, with LC present in 56.76%. Common comorbidities included hypertension (18.92%), diabetes (7.21%), and cardiopathy (6.31%). The primary clinical symptoms were cough (34.23%) and expectoration (27.03%). The treatment cohort had a median age of 70 years (IQR: 61-73.5), 68.75% were male, and LC was diagnosed in 43.75% of patients. Hypertension (18.75%) and diabetes (15.63%) were common. Predominant symptoms included cough (59.38%) and expectoration (56.25%).

To evaluate the diagnostic performance of TNPseq, we first compared it with mNGS and conventional culture. Despite generating a lower total volume of sequencing data, especially for bacteria (Figures 1A, S3), TNPseq exhibited a decisive advantage in clinical applicability. It identified a significantly higher proportion of clinically relevant potential pathogens (48.76%, 296/607) compared to mNGS (16.80%, 896/5,334; P<0.001; Figure 1B).

Operationally, TNPseq achieved a 40.7% faster turnaround (16 hours for TNPseq vs. 27 hours for mNGS), facilitated by rapid real-time sequencing (7 hours for TNPseq vs. 22 hours for mNGS; Supplementary Table S2).

Culture tests were performed on all samples, with 11 yielding positive results. Both sequencing methods demonstrated 72.73% sensitivity for culture-positive samples. Within the cohort of 16 CRS-positive cases, TNPseq showed higher sensitivity (81.25%) than mNGS (68.75%; Supplementary Table S3), highlighting TNPseq’s strength in rapidly and precisely identifying clinically relevant pathogens. Additionally, TNPseq detected a slightly higher number of pathogens compared to mNGS (Supplementary Table S4).

Despite the difference in data volume, the two sequencing methods showed strong agreement in detecting common and critical respiratory pathogens. Both methods identified common microorganisms including Prevotella, Streptococcus, and Haemophilus (Figure 1C). Concordance between methods was high for lung cancer-associated pathogens such as Haemophilus influenzae, Streptococcus pneumoniae, and human gammaherpesvirus 4 (Figure 1D). This high concordance confirms that TNPseq retains the accurate detection capability of broad-spectrum sequencing while offering enhanced speed and specificity.

We examined microbial diversity and composition across patients with BPD, IDLC, and NDLC. Rank abundance curves (RACs) differed by sequencing method: mNGS showed broader taxonomic diversity and evenness, whereas TNPseq had steeper slopes, indicating dominance by fewer pathogens (Figures 2A, E). Despite these methodological differences, both approaches exhibited consistent trends. Significant differences in fungal α-diversity emerged between BPD and NDLC groups using both mNGS (P = 0.023) and TNPseq (P = 0.0098), as assessed by the Shannon index (Figures 2B, F). However, no significant differences were observed in the ACE, Chao1, Richness, or Simpson indices(Supplementary Figure S4). Furthermore, β-diversity did not differ significantly among the three groups (Figures 2C, G). Taxon abundance rankings aligned closely between methods, with Prevotella, Streptococcus, and Rothia consistently dominating across all patient groups (Figures 2D, H).

LEfSe analysis identified distinct microbial biomarkers associated with specific disease groups. Both sequencing platforms detected significant enrichment of Lactobacillus in NDLC patients (mNGS: P<0.01; TNPseq: P<0.001) (Figures 3A, C), which was further confirmed through heatmap visualization (Figures 3B, D).

To investigate the microbiome differences between the tumor response group (PR/SD) and the progression group (PD), we assessed microbial diversity and composition in these two groups. RAC revealed that mNGS detected greater taxonomic breadth and evenness, whereas TNPseq exhibited steeper slopes, indicating the predominance of dominant pathogens (Supplementary Figures S5A, E). Both platforms demonstrated no significant differences in overall α-diversity (Supplementary Figures S5B, F, S6) or β-diversity (Supplementary Figures S5C, G). The concordance between methods was further evidenced by taxon abundance rankings, which identified Prevotella and Streptococcus as the dominant genera in both groups (Supplementary Figures S5D, H).

Given the limited sample size, LEfSe analysis was conducted as an exploratory investigation into bacterial species differences. The results suggested that Neisseria, Veillonella, and Prevotella may be associated with the PR/SD group, showing significant enrichment (P < 0.05) by both methods (Figures 4A, C). Visualization using heatmaps confirmed their relatively elevated abundance in PR/SD (Figures 4B, D). Additionally, LEfSe analysis of TNPseq data also revealed significant enrichment of Streptococcus intermedius in PD group (P < 0.01) (Figure 4C).

We evaluated clinical outcomes among 32 patients (14 with LC, 18 with BPD), comparing initial empirical antibiotic therapy (n=19) with direct TNPseq-guided therapy (n=13). Clinical improvement was ultimately achieved in all patients. However, among those who initially received empirical therapy, 68.4% (13/19) showed no improvement and required subsequent treatment adjustment based on TNPseq findings (Figure 5A). When directly guided by TNPseq from the outset, patients experienced a median treatment duration that was 53.8% shorter than that of the empirical-to-TNPseq therapy group (6 days vs. 13 days; P<0.01) (Figure 5B). For instance, a patient presenting with cough, yellow-white sputum, and chest tightness showed no improvement after 12 days of empirical cefuroxime. Subsequent TNPseq analysis identified Stenotrophomonas maltophilia (missed by conventional culture), and the patient improved after 7 days of targeted trimethoprim-sulfamethoxazole therapy.

Within the LC subgroup (n=14), 35.7% (5/14) required switching from empirical to TNPseq-guided therapy due to initial therapeutic failure (Figure 5C). We preliminarily found that direct TNPseq-guided therapy significantly shortened the median treatment duration (5 days vs. 15.5 days; P<0.05) compared to the empirical-to-TNPseq transition (Figure 5D). TNPseq effectively reduced unnecessary antibiotic use. For example, an elderly patient with suspected pneumonia failed initial empirical levofloxacin therapy; TNPseq testing revealed human herpesvirus 4 without bacterial or fungal pathogens. Antibiotics were discontinued, and lung adenocarcinoma was later confirmed by biopsy. Another culture-negative patient underwent TNPseq, detecting Haemophilus influenzae, human herpesvirus 4, and the active β-lactam resistance gene bla_TEM-1. Prompt targeted therapy with levofloxacin and acyclovir led to symptom resolution within 7 days, avoiding broader-spectrum antibiotics. These findings underscore TNPseq’s efficacy in optimizing infection treatment and supporting antimicrobial stewardship in patients with cancer.