This retrospective, monocentric cohort study was conducted in the general surgery ICU of Xuanwu hospital (January 1, 2023, to November 30, 2025). Inclusion and exclusion criteria specific criteria are detailed in Figure 1 and Supplementary Table S1. Briefly, inclusion required: (1) Age ≥18; (2) Diagnosis of septic shock per sepsis-3 criteria; (3) mNGS testing within 24 h of ICU admission. Exclusion criteria were operationalized as: (1) Pregnancy; (2) ICU stay <24 h; (3) Missing data for primary outcomes or any baseline laboratory variables. Septic shock was defined according to the sepsis-3 criteria and identified via ICD-10 codes. To ensure independence of observations, only the first ICU admission was analyzed for patients with multiple stays. Patients were categorized into two groups: those with underlying malignancies (gastrointestinal and hepatobiliary cancers) and those without. For patients with ICU admissions for septic shock, each stay was analyzed independently. Epidemiological, clinical, and biochemical data were extracted from electronic medical records and standardized, with key parameters spanning demographics (age, sex, weight), comorbidity and health status [assessed using the Charlson Comorbidity Index (CCI) and APACHE II score], therapeutic interventions (including pre-ICU and ICU-based therapies such as antibiotic administration, mechanical ventilation, renal replacement therapy, and transfusion support), and clinical outcomes (including ICU length of stay, ICU mortality, in-hospital mortality). Clinical samples for mNGS (including blood, bile, ascitic fluid, and bronchoalveolar lavage fluid) were consistently collected within 24 h of the onset of septic shock or upon ICU admission. In most cases, sampling was performed immediately following the initial clinical suspicion of infection and prior to any significant modifications to the empirical antimicrobial regimen. Definitions and bias control the cancer status was defined as histologically confirmed gastrointestinal or hepatobiliary malignancy, with details on TNM stage and chemotherapy or immunotherapy status extracted to account for severity. Selection bias was acknowledged due to the surgical specialty of our center. The primary outcome was in-hospital all-cause mortality. Secondary outcomes included ICU length of stay and antibiotic duration. Antimicrobial stewardship was guided by institutional protocols based on IDSA guideline (Barlam et al., 2016), incorporating mNGS results for targeted adjustments.

To compare the microbial landscape, clinical samples including peripheral blood, ascitic fluid, bile, and bronchoalveolar lavage fluid (BALF) were collected. In our clinical practice, empirical antimicrobial therapy is initiated immediately upon the clinical suspicion of septic shock. Consequently, mNGS sampling was performed concurrently with or within the first few hours of antibiotic initiation. BALF and other body fluid sample DNA was extracted using the TIANamp Magnetic DNA Kit (Tiangen, Beijing, China) according to the manufacturer’s protocols. Plasma was prepared from blood samples and circulating cell-free DNA was isolated from plasma follow the instructions in QIAamp Circulating Nucleic Acid Kit (Qiagen, Germany). The quantity and quality of DNA was assessed with the Qubit 2.0 (Thermo Fisher Scientific, United States) and NanoDrop 8000 (Thermo Fisher Scientific, United States), respectively. DNA library construction was performed using the Hieff NGS C130P2 OnePot II DNA Library Prep Kit (Yeasen Biotechnology, Shanghai, China), adhering to the manufacturer’s guidelines. Agilent 2100 system was used for quality control of the DNA libraries, and sequencing was carried out on the DIFESEQ-200 platform (Dinfectome, Nanjing, China) employing a 50-base single-end, approach yielding approximately 20 million reads for the metagenomic workflow.

We use in-house developed bioinformatics pipeline for pathogen identification (Lv et al., 2024). Briefly, high-quality sequencing data were generated by removing low quality reads, adapter contamination, duplicated and shot (length <36 bp) reads. Human host sequences were identified by mapping to human reference genome (hs37d5) using bowtie2 software (version 2.2.6). Reads that could not be mapped to the human genome were retained and aligned with microorganism genome database for pathogens identification. Our microorganism genome database contained bacteria, fungi, virus and parasite genomic sequences (download from https://www.ncbi.nlm.nih.gov/).

Subsequently, we applied specific criteria to ascertain positive results from mNGS, as referenced in prior studies (Lv et al., 2024; Xu et al., 2023). The criteria for a positive detection were as follows: (1) At least one species-specific read was required for the detection of Mycobacterium, Nocardia, and Legionella pneumophila; (2) For other bacteria, fungi, viruses, and parasites, a minimum of three unique reads were necessary; (3) Pathogens were considered negative if the ratio of microorganism reads per million to the corresponding no-template control was less than 10. In the absence of a universally accepted standard for mNGS interpretation, we utilized a stringent, multi-step validation protocol to define clinical relevance, specifically tailored for the high-risk septic shock population besides bioinformatics thresholds. Specifically, two senior clinicians independently reviewed each candidate pathogen alongside the patient’s clinical presentation, imaging, and concurrent culture results. A result was only recorded as positive if it was deemed the likely causative agent of the septic shock episode, thereby ensuring clinical relevance and minimizing the impact of potential commensal or environmental DNA.

Descriptive statistics were used to summarize the data: categorical variables are presented as frequencies and proportions, while continuous variables are expressed as medians with interquartile ranges (IQR). Continuous variables were compared using the Mann–Whitney U test. Categorical variables were assessed using the chi-squared test or Fisher’s exact test. Metagenomic correlation: microbial diversity (alpha and beta diversity) and the abundance of ARGs were compared between the cancer and non-cancer groups. Cox regression models were constructed to identify independent risk factors for ICU and in-hospital mortality, incorporating statistically significant (p < 0.05) and clinically relevant variables. Before performing univariate Cox regression analysis, model fitting checks were conducted separately for the tumor group and non-tumor group. For variables with the warning message “Loglik converged before variable 1; coefficient may be infinite”, this indicated the presence of perfect or quasi-perfect separation, where the model failed to stably estimate the hazard ratio (HR). To evaluate whether the prognostic impact of key variables was modulated by malignancy status, we performed formal interaction tests. Interaction terms between cancer vs. non-cancer group and each candidate predictor were incorporated into the Cox proportional hazards models. A p-value for interaction (p_interaction) <0.05 was considered to indicate a significant or suggestive differential effect, which provided the statistical rationale for performing stratified multivariable analyses within the cancer and non-cancer cohorts. All statistical analyses and visualization were performed using R software (version 4.0.2) and PRISM 9.0. A two-sided p-value <0.05 was considered statistically significant. Heatmaps were generated to visualize the correlation patterns of the studied variables using the pheatmap package in R (version 1.0.13). Rows and columns were hierarchically clustered using average linkage hierarchical clustering with dissimilarity measured by 1 minus Pearson correlation coefficient.

A total of 155 patients with suspected septic shock were screened. After excluding repeat admissions (n = 17) and those with incomplete clinical or mNGS data (n = 31), 107 patients admitted to the general surgery ICU with septic shock (SSh) were included in this study. The study population comprised 52 cancer patients (48.6%) and 55 non-cancer patients (51.4%). The baseline characteristics of all participants are detailed in Table 1. The median age of the total cohort was 60.00 years (IQR 48.00–70.00), with a slight male predominance (51.4%). Hypertension was the most prevalent comorbidity, affecting 51.4% (n = 55) of the patients. The median body mass index (BMI) was 24.49 kg/m² (IQR 22.27–27.34 kg/m²). Regarding disease severity at admission, the median APACHE II score was 5.00 (IQR 4.00–7.00), and the median Charlson Comorbidity Index (CCI) was 6.00 (IQR 4.00–10.00).

Significant differences in baseline demographics and clinical status were observed between cancer and non-cancer patients (Table 1; Supplementary Table S2). Cancer patients were significantly older than non-cancer patients (median 68.00 years vs. 51.00 years, p < 0.0001). Conversely, non-cancer patients exhibited a higher prevalence of hypertension (71.15% vs. 32.73%, p < 0.0001) and a higher median BMI (25.22 vs. 23.34 kg/m², p = 0.0021). Severity scores were notably higher in the cancer group, including both the APACHE II score (6.00 vs. 5.00, p = 0.0019) and the CCI (7.00 vs. 4.00, p = 0.006). In terms of clinical and laboratory findings during the ICU stay (Table 2), cancer patients demonstrated a higher neutrophil ratio (87.00% vs. 82.60%, p = 0.0088) and lower platelet counts (161.0 vs. 268.0 × 10⁹/L, p = 0.0157) compared to non-cancer patients. In the cancer group, the primary malignancies were predominantly located in the gastrointestinal tract and (n = 33, 63.46%) hepatobiliary system (n = 16, 30.77%) (Supplementary Table S3). Regarding the TNM clinical stage, a significant proportion of patients presented with advanced disease: 20 patients (38.46%) were classified as Stage III and 11 patients (21.15%) as Stage IV, reflecting the high disease burden in this ICU population. 21 patients (40.38%) were at Stage I or II. The majority of patients (n = 28, 53.85%) developed sepsis following major surgical resection (within 7 days of ICU admission), highlighting the complex surgical background of oncological sepsis. Active chemotherapy or immunotherapy within the past 30 days was reported in 19 patients (36.39%).

All 107 patients (100%) underwent multi-site sampling for mNGS, including blood, with additional samples obtained from ascitic fluid (63.55%), bile (12.15%), and bronchoalveolar lavage fluid (11.21%). The overall positivity rate for pathogen detection was 70.09% (n = 75). Interesting disparities emerged in the infectious profiles between the two groups. Non-cancer patients had a significantly higher rate of positive samples compared to cancer patients (85.45% vs. 53.85%, p = 0.0006). Similarly, the incidence of bacteremia was nearly double in the non-cancer group (45.45% vs. 25.00%, p = 0.0426). While the presence of resistance genes was numerically higher in non-cancer patients (61.82% vs. 48.08%), this did not reach statistical significance (p = 0.1767).

Interventions and outcomes are summarized in Table 2. Regarding respiratory and renal support, 30.84% of patients required mechanical ventilation, and 34.58% received continuous renal replacement therapy (CRRT), with no significant differences between groups. However, the duration of antibiotic therapy was significantly shorter in cancer patients compared to non-cancer patients (7.00 days). To identify independent predictors of prognosis, we performed univariate and multivariate Cox regression analyses for the two cohorts separately, variables were selected based on a predefined criterion (univariable p < 0.05 and clinical relevance). In the non-cancer group, univariate analysis revealed that age (HR 1.159, 95% CI 1.043–1.288, p = 0.0059) was significantly associated with survival (Table 3). In the multivariate model, age remained a highly significant independent risk factor (HR 1.18, 95% CI 1.02–1.38, p = 0.027). In the cancer group, univariate analysis showed that IL-6 levels (HR 1.001, 95% CI 1–1.001, p = 0.0109) and WBC count (HR 1.208, 95% CI 1.043–1.400, p = 0.0118) were associated with mortality. However, after adjusting for potential confounders in the multivariate analysis, only IL-6 reached statistical significance (HR 1.00, 95% CI 1.00–1.00, p = 0.018) (Table 4). To evaluate the modulatory effect of malignancy on prognostic drivers, formal interaction tests were performed between grouping status and candidate predictors. The analysis revealed significant interactions for age (p_interaction = 0.004), along with a suggestive interaction for the IL-6 (p_interaction = 0.055) (Figure 2). These results indicate that the prognostic weight of systemic inflammation and baseline physiological reserves differs substantially between the two cohorts, suggesting a trend that IL-6 may have a significant impact on the outcome; however, no statistically significant difference was observed, which may be limited by the small sample size. In contrast, no significant interactions were observed for CCI (p_interaction = 0.388), bacteremia (p_interaction = 0.109), suggesting their effects on mortality remained relatively consistent.

The microbial landscape, as determined by mNGS across multiple sample types (blood, bile, ascitic fluid, and BALF), is visualized in the hierarchical clustering heatmap (Figure 3 and Supplementary Table S4). Top 15 pathogens were categorized into four major classes: Gram-negative bacilli, Gram-positive cocci, anaerobes, and fungi and other opportunistic pathogens. In term of pathogen distribution, Gram-negative bacilli, including Klebsiella pneumoniae, Escherichia coli, Pseudomonas aeruginosa, and Acinetobacter baumannii, were frequently detected across both groups. Gram-positive cocci were primarily represented by Enterococcus faecium, Enterococcus faecalis, and Staphylococcus aureus. Anaerobic species such as Bacteroides fragilis and Bacteroides vulgatus were notably present, particularly in ascitic fluid samples. Fungal pathogens, specifically Candida albicans and Candida glabrata, were also identified. The clustering analysis integrates clinical metadata, demonstrating the relationship between the microbial infectome and patient outcomes and ICU duration. While both groups shared common ICU pathogens, the non-tumor group showed a higher overall pathogen positive rate (85.45% vs. 53.85%, p = 0.0006) and a higher incidence of bacteremia (45.45% vs. 25.00%, p = 0.0426), as detailed in the clinical comparisons.

To further elucidate the molecular basis of antimicrobial resistance in the General Surgery ICU of Xuanwu Hospital, we integrated taxonomic findings with ARGs annotation using the mNGS data. The resulting pathogen-resistome map, organized by microbial groups and resistance gene categories, provides a high-resolution view of the genetic determinants of resistance (Figure 4 and Supplementary Table S5). The mNGS analysis identified a broad spectrum of ARGs across several major resistance gene groups, including aminoglycoside resistance genes, macrolide-lincosamide-streptogramin resistance, tetracycline resistance, and various beta-lactamases. Specific resistance across different microbials was also found from resistomes. Klebsiella pneumoniae and Escherichia coli exhibited the most extensive resistome profiles, frequently harboring genes across almost all resistance genes, particularly those conferring resistance to aminoglycosides and beta-lactams. Acinetobacter baumannii and Pseudomonas aeruginosa also showed targeted resistance patterns, aligning with their clinical roles as multidrug-resistant (MDR) pathogens. In Gram-positive cocci, Enterococcus faecium demonstrated a high density of resistance genes, particularly in the macrolide and aminoglycoside groups. Staphylococcus aureus was similarly associated with multiple resistance determinants. Although the non-tumor group had a higher overall pathogen detection rate, the pathogen-resistome reveals that once an infection is established, the complexity of the resistome is significant in both groups. This underscores that for cancer patients in the surgical ICU, even a lower microbial load can be associated with highly resistant genetic signatures, complicating empiric therapy. Additionally, we examined the distribution of virulence factors (VFs) and found them in only 6 patients in the cancer group and 2 in the non-cancer group. Most identified VFs were derived from Gram-positive cocci, and all VFs, including iutA, exoT/U/Y, and hysA, are summarized in Table 5.