Bronchoalveolar Lavage Fluid (BALF) samples from PLWH with pneumonia were submitted to Probe-Capture Metagenomics, which were retrospectively reviewed from the Zhongnan Hospital of Wuhan University from 1 may 2022 to 30 December 2022. Patients who met the following criteria were considered as pneumonia: (1) exhibited at least one of the following manifestations: fever; cough, expectoration, dyspnea or other respiratory symptoms; (2) chest imaging suggested pneumonia; (3) determined by the consensus of two experienced senior clinicians based on the clinical manifestation, laboratory tests, chest imaging and response to antimicrobial treatment. The exclusion criteria were as follows: (1) Probe-Capture Metagenomics of BALF were not performed; (2) age < 18 years old; (3) Hospital stay less than 24 h (Figure 1). At the same time, CMTs, including chest imaging, smear and culture of bacteria and fungi, 1,3-β-D glucan (G) assay (G test) (β-D-glucan detection kit, Shanghai Fuxing Changzheng Medical Science Co., Ltd., China) and galactomannan (GM) assay (Platelia^™ Aspergillus Ag, Bio-Rad, United States), polymerase chain reaction (PCR) test [including cytomegalovirus (CMV), EB virus, Pneumocystis jirovecii (P. jirovecii), Mycobacterium tuberculosis complex (MTB)], tuberculin skin test (TST), interferon-gamma release assay (T-SPOT.TB kit, TB.300, Oxford Immunotec, UK), acid-fast stain (TB Auramine-Rhodamine kit, K12345, Thermo Fisher, United States) and GeneXpert MTB/RIF (Xpert) (Cepheid. Xpert^® MTB/RIF Assay, Cepheid, United States), were performed for pulmonary infection. The sequence and conditions of the primer pairs used for P. jirovecii PCR was described in one previous study (Chen et al., 2024). The primers for CMV PCR were as follows: Forward: 5’-CGCGACTTGACCTGGACTAC-3′, Reverse: 5’-TGGTGGCAGTGTACCCGTAA-3′. The conditions used were: 45 cycles of denaturation for 5 min at 95°C; annealing for 15 s at 95°C; and extension for 30 s at 60°C. The primers for EB PCR were as follows: Forward: 5’-GCCAGAGGTAAGTGGACTTT-3′, Reverse: 5’-TCTTGGTGAGCGGTAGTAAA-3′. The conditions used were: 40 cycles of denaturation for 3 min at 95°C; annealing for 10 s at 95°C; and extension for 30 s at 58°C. The primers for MTB PCR were as follows: Forward: 5’-CCTGCGAGCGTAGGCGTCGG-3′, Reverse: 5’-CTCGTCCAGCGCCGCTTCGG-3′. The conditions used were: 50 cycles of denaturation for 10 min at 95°C; annealing for 15 s at 95°C; and extension for 60 s at 62°C. The date of collection, symptoms and test results were included in the clinical data. All data was anonymized in our study. These patient clinical data was included as Supplementary Material. Every fungi/bacteria/virus evaluated in CMTs were present in the databases of Probe-Capture Metagenomics. The diagnosis was made based on clinical symptoms, laboratory findings, microbiologic examination, chest radiology, and the response to treatment. This study was approved by the institutional ethics committee of Zhongnan hospital of Wuhan university (2022101). All subjects have voluntarily taken part in the research and signed the informed consent form. The Declaration of Helsinki’s standards and any applicable laws were followed when conducting the study.

Clinical data were retrospectively obtained from medical records. These following data were extracted: age, gender, HIV infection status, antiretroviral therapy (ART), antibiotic use within 3 months, lymphocyte count, CD4⁺ T lymphocyte count (CD4 count), HIV viral load, C-reactive protein (CRP), Interleukin-6 (IL-6), procalcitonin (PCT), erythrocyte sedimentation rate (ESR), D-dimer, lactate dehydrogenase (LDH), alanine aminotransferase (ALT), albumin, creatinine, results of CMTs, chest imaging, and Probe-Capture Metagenomics tests.

To optimize probe design, we employed HUBDesign, a bioinformatics pipeline that generates probes from representative sequences at multiple taxonomic levels. This method allows for precise targeting and enrichment of specific clades, facilitating the detection of both known and novel organisms. The default parameters of HUBDesign were applied in this study. We designed a comprehensive probe library targeting up to 167 clinically relevant pathogenic microorganisms. And the pathogen-specific probe set includes 62 bacteria and 41 antimicrobial resistance genes, 16 fungi, 14 parasites, 15 DNA viruses, and 60 RNA viruses, collectively covering >95% of pathogens detected in clinical samples. A total of 640 probes were designed including 517 probes targeting pathogens and 183 probes targeting resistance genes. Pathogens used for probe design are provided in Supplementary Materials. Leveraging sequence homology and DNA hybridization, this pipeline has been successfully demonstrated in designing and validating probe sets for two key applications: (1) comprehensive detection of all sequenced coronaviruses, and (2) identification of bacterial pathogens associated with sepsis. The Probe-Capture assay developed in this work has been granted an invention patent by the China National Intellectual Property Administration (patent name: Group, method, kit and application of Probe Capture for detecting pathogenic microorganisms, patent No. ZL2020116367800).

DNA was extracted from all samples using a QIAamp^® UCP Pathogen DNA Kit (Qiagen) following the manufacturer’s instructions. Human DNA was removed using Benzonase (Qiagen) and Tween20 (Sigma). Using a Qubit dsDNA HS Assay Kit (Invitrogen, United States), the isolated DNA was quantified. Total RNA was extracted with a QIAamp^® Viral RNA Kit (Qiagen) and ribosomal RNA was removed by a Ribo-Zero rRNA Removal Kit (Illumina). cDNA was generated using reverse transcriptase and dNTPs (Thermo Fisher). Libraries were constructed for the DNA and cDNA samples using a Nextera XT DNA Library Prep Kit (Illumina, San Diego, CA). An aliquot of 750-ng library from each sample was used for hybrid capture-based enrichment of microbial probe one rounds of hybridization (SeqCap EZ Library, Roche, United States). Purification and size selection were carried out following the double-sided bead purification procedure. Library was quality assessed by Qubit dsDNA HS Assay kit followed by High Sensitivity DNA kit (Agilent) on an Agilent 2,100 Bioanalyzer. Library pools were then loaded onto an Illumina Nextseq CN500 sequencer for 75 cycles of single-end sequencing to generate approximately 20 million reads for each library. For negative controls, we also prepared sterile deionized water in parallel with each batch to serve as non-template controls (NTC), using the same protocol.

Raw reads were adapter trimmed and quality filtered using Trimmomatic (v0.39). All reads longer than 40 base pairs with a minimum phred quality score of 30 were kept for further analysis (Bolger et al., 2014). Low-complexity reads were removed by fastp (v0.23.4) default settings (Chen et al., 2018). The human sequence data were identified and eliminated by mapping them to the hg38 reference genome utilizing Burrows-Wheeler Aligner software (BWA-0.7.17) (Li and Durbin, 2009). Taxonomic classification of sequencing reads was performed using Kraken2 (v2.1.3) with default parameters, including a confidence threshold of 0.2 (-confidence 0.2) and 40 threads (-threads 40). The Kraken2 reference database was constructed using the kraken2-build –standard pipeline, incorporating complete genomes and assemblies from the NCBI RefSeq archaea, bacteria, viruses, plasmids, human (GRCh38), UniVec_Core, protozoa, fungi and plants, all downloaded on December 28, 2024. To enhance microbial representation, genomes of Pneumocystis jirovecii (GenBank accession no. GCA_001477535.1) were manually incorporated into the database, as detailed in Supplementary Table 1. To mitigate spurious alignments arising from PCR duplicates or low complexity regions, microbial reads were subsequently aligned against the same reference database using the Burrows–Wheeler Aligner (BWA, v0.7.17) with default parameters. Only alignments exhibiting a nucleotide identity of ≥96% to the reference genome were considered valid. Furthermore, reads aligning to multiple loci within the same genus were excluded from downstream analyses to prevent ambiguous taxonomic assignments. To remove the mistakes brought on by different sequencing depths between samples, we normalized the sequencing reads using reads per million (RPM). The clinical reportable range (CRR) for pathogens was established according to the following three references indicated in a previous study (Jing et al., 2021): I. Johns Hopkins ABX Guide,¹ II. Manual of Clinical Microbiology, and III. clinical case reports or research articles published in peer-reviewed journals.

The clinical reportable range (CRR) for pathogens was established according to a previous study (Sun et al., 2025). The suspected pathogens were determined according to the following rules (Wilson et al., 2019): A virus was considered positively detected if RPM ≥ 3 covered at least three non-overlapping regions in the genome. For identification of bacteria, fungi, and parasites, we developed a RPM ratio metric, defined as RPM-r = RPM sample/RPM_NTC, or RPM ratio = RPM sample if RPM_NTC = 0. A positive detection of a bacterium, fungus, or parasite required an RPM ratio ≥10. The results can be considered positive when they are consistent with the pathogenicity of positive microorganisms, clinical characteristics, and therapeutic efficacy.

The Stanford HIV Genotypic Resistance Interpretation Algorithm² was used to identify mutations associated with HIV drug resistance. HIVDR mutations linked with integrase strand transfer inhibitor (INSTI), reverse-transcriptase inhibitor (RT), and protease inhibitor (PI) were reported with threshold of mutation detection threshold (MDT) ≥ 1%, and divided into three categories: major, accessory, and other. The estimated drug resistance based on the total point score was designated: susceptible, potential low-level resistance, low-level resistance, intermediate resistance, and high-level resistance.

To identify clinically important antimicrobial resistance bacteria, such as carbapenem-resistant Enterobacterales, carbapenem-resistant Acinetobacter baumannii, carbapenem-resistant Pesudomonas aerugeinosa, third-generation cephalosporin-resistant Enterobacterales, extended-spectrum β-lactamases (ESBL) producing Enterobacteriaceae, vancomycin-resistant Enterococcus faecium, methicillin-resistant Staphylococcus aureus, and so on, the specimen’s sequence was screened and aligned with sequences of antibiotic resistance genes in the Comprehensive Antibiotic Resistance Database (CARD; https://card.mcmaster.ca) (Jia et al., 2017). This analysis was only performed when the targeted bacterial sequence surpassed a threshold of 1,000. And the relationship between antibiotic resistance genes and the detected bacteria was evaluated using the CARD database (Alcock et al., 2023).

SPSS 25.0 software (IBM Corp., Armonk, NY, United States) was used for statistical analysis. The numerical variables were presented as medians and interquartile ranges and nominal variables as counts and percentages. Comprehensive clinical diagnosis and determination of microbial aetiology were used as reference standards. The McNemar test was used to compare the diagnostic performance of CMTs and Probe-Capture Metagenomics. Fisher’s exact test was used to test for association of pathogen with the categorical variables of CD4 count (≥200 vs. < 200 cells per μL). All tests were two-tailed, and the p value of <0.05 was considered statistically significant.

During the study period, a total of 91 HIV patients (mean age: 48 years) including 78 men (85.7%) met the inclusion criteria. Upon admission to hospital, the median CD4 count of 91 patients was 57 cells/μL. 48 (52.7%) were taking antiretroviral therapy, and 16 (17.58%) had been receiving P. jirovecii pneumonia prophylaxis. 10 (11.24%) out of 89 patients were receiving empirical antifungal therapy, and 62 (68.13%) out of 91 patients were receiving empirical antibiotic therapy. 57 patients (62.64%) had fever, and 72 patients (79.12%) of CT scan of the lung were abnormal. Detailed clinical characteristics are shown in Table 1.

Of 91 samples, excluding the EB and Torque teno virus, 6 had single pathogen and 85 had at least two microorganism (Figure 2A) detected by Probe-Capture Metagenomics combined with CMTs. In addition, the top six organisms were CMV, P. jirovecii, MTB, HHV-7, Candida albicans (C. albicans) and Aspergillus (Figures 2B–E).

As shown in Figure 2B, Tropheryma whipplei (T. whipplei), Enterococcus faecium, Klebsiella pneumoniae, Pseudomonas aeruginosa and Streptococcus pneumoniae were more prevalent in the detected bacteria spectrum among PLWH with pulmonary infection.

As shown in Figure 2C, P. jirovecii and C. albicans were more common in the detected fungus spectrum. Among the detected fungus, two kinds (Aspergillus and Cryptococcus) of them were detected by CMTs only, five kinds (Candida krusei, Cryptococcus neoformans, Candida glabrata, Sedosporium citrifolia and Candida tropicalis) of them were detected by Probe-Capture Metagenomics only. In the cases of P. jirovecii detection, 5 cases were diagnosed by Probe-Capture Metagenomics alone, and 6 cases were diagnosed by CMTs alone. In the cases of C. albicans detection, 12 cases were diagnosed by Probe-Capture Metagenomics alone, and 1 case was diagnosed by CMTs alone. In the cases of Aspergillus fumigatus detection, 5 cases were diagnosed by Probe-Capture Metagenomics alone. In the cases of Aspergillus flavus detection, 6 cases were diagnosed by Probe-Capture Metagenomics alone and 2 cases were diagnosed both by Probe-Capture Metagenomics and CMTs. In the cases of T. marneffei detection, 4 cases were diagnosed by Probe-Capture Metagenomics alone. In the cases of Aspergillus niger detection, 2 cases were diagnosed by Probe-Capture Metagenomics alone, and 1 case was diagnosed by CMTs alone. In the cases of Candida tropicalis detection, 1 case was diagnosed by Probe-Capture Metagenomics alone.

As shown in Figure 2D, CMV and HHV-7 were more common in the detected virus spectrum. In the cases of CMV detection, 26 case was diagnosed by Probe-Capture Metagenomics alone. In addition, except for CMV, the remaining viruses were detected only by Probe-Capture Metagenomics.

As shown in Figure 2E, MTB and Mycobacterium avium complex (MAC) were more common in the detected mycobacterium spectrum. In the cases of MTB detection, 5 case was diagnosed by Probe-Capture Metagenomics alone, 13 cases were diagnosed both by Probe-Capture Metagenomics and CMTs, and 5 cases were diagnosed by CMTs alone. In addition, all of the nontuberculous mycobacteria (NTM), Mycoplasma hominis, Chlamydia trachomatis, Ureaplasma and Legionella pneumophila were detected only by Probe-Capture Metagenomics.

As shown in Table 2 and Supplementary Table 2, Probe-Capture Metagenomics demonstrated superior detection performance compared to CMTs for CMV and C. albicans infections in PLWH with pulmonary infections. Using the McNemar test to compare paired binary outcomes (positive/negative) from Probe-Capture Metagenomics and CMTs, Probe-Capture Metagenomics showed significantly higher detection rates for CMV (p < 0.001) and C. albicans (p < 0.001). For CMV viremia, Probe-Capture Metagenomics achieved a sensitivity of 100% (95% CI, 94.65–100%) compared to 67.65% for CMTs (95% CI, 55.84–77.56). Similarly, for C. albicans infection, Probe-Capture Metagenomics sensitivity was 94.74% (95% CI, 75.36–99.73), significantly higher than CMTs at 31.58% (95% CI, 15.36–53.99; p < 0.001). In contrast, BALF Probe-Capture Metagenomics showed comparable detection performance to CMTs for P. jirovecii, MTB, and Aspergillus infections (p > 0.05, McNemar test).

We further compared the lung microorganisms of PLWH with pulmonary infection between patients with CD4⁺T count < 200 cells/μL to those with a count ≥ 200 cells/μL. As shown in Table 3, the detection ratios of CMV and P. jirovecii in patients with CD4⁺T count < 200 cells/μL were significantly higher than that of CD4⁺T count ≥ 200 cells/μL (p < 0.05). However, there are no significant differences of the detection ratios of MTB, C. albicans, NTM, Aspergillus, T. marneffei, HHV-7 and EBV between patients with CD4⁺T count < 200 cells/μL and CD4⁺T count ≥ 200 cells/μL (p > 0.05).

Furthermore, 91 samples were successfully sequenced to detect DRMs using Probe-Capture Metagenomics. HIV sequences were detected in 26 BALF samples. Detailed results are presented in the Supplementary Materials. As shown in Figure 3, DRMs (V179E, V179D) for nonnucleoside reverse transcriptase inhibitors (NNRTI) were found in the BALF of 3 patients. The scores of DRMs associated with HIV resistance were all 10 in each cases, indicating that HIV in these three patients had potentially low-level resistance to NNRTI (Figure 3B). In addition, the BALF of 27 patients exhibited the presence of antimicrobial resistance genes (ErmB, blaTEM, ErmC, mecA, blaCTX-M, blaSHV), which were identified by Probe-Capture Metagenomics (Table 4).