**Background:** Candida species are among the most common causes of fungal bloodstream infections, with high mortality in immunocompromised patients. While Candida overgrowth in the gut is considered a prerequisite for invasive candidiasis, the specific bacterial communities that promote or restrict this overgrowth in humans remain poorly understood. Most prior research has relied on murine models, which differ substantially from human gut communities in composition and immune regulation. This study aimed to characterize the intestinal ecological signature associated with Candida species expansion in the human gut using an integrative multi-omics approach in lung cancer patients.
**Methods:** The study enrolled 75 advanced-stage lung cancer patients (40 adenocarcinoma, 28 squamous cell carcinoma, 7 other) from two Hungarian hospitals, all receiving anti-PD-1 immunotherapy (nivolumab n=44, pembrolizumab n=31). Stool samples were collected after therapy initiation. Fungal communities were characterized via ITS2 sequencing (average 78,332 high-quality reads/sample), while bacterial and viral communities were profiled using shotgun metagenomic sequencing (average 26,106,952 reads/sample). Patients were stratified into high-Candida (HC, n=38, mean TSS=33.4%) and low-Candida (LC, n=37, mean TSS=0.6%) groups based on median CLR-normalized Candida genus abundance. Machine learning models (SIAMCAT with data augmentation) were trained to predict Candida levels using bacterial features and validated in an independent cohort of 11 patients. In vitro competition experiments were performed using four C. albicans gut isolates and a human-associated S. cerevisiae strain (YJM128) under varying carbon sources (glucose vs. glucose+lactate) and oxygen conditions (aerobic vs. microaerobic).
**Key Results:** Beta diversity analysis showed significant separation between HC and LC groups for both fungal (PERMANOVA P=0.001, R²=5.7%) and bacterial communities (P<0.05). Bacterial species explained approximately 13% of fungal species beta diversity. The HC group showed significantly lower obligate anaerobe abundance (ΔR²=8%, P=0.017) and a trend toward increased oxygen-tolerant/intolerant ratio (ΔR²=5%, P=0.058). Machine learning models using bacterial species abundance classified patients with high accuracy (main cohort CV auROC=77.9%; validation cohort auROC=78.6%), while bacterial functional abundances (EC) achieved slightly higher accuracy (CV auROC=80.4%; validation auROC=82.1%). Species-level models for C. albicans, C. sake, and C. glabrata showed similar performance (validation auROC: 86.7%, 86.7%, and 79.2%, respectively). Differential abundance analysis revealed that LC-predictive bacteria were predominantly SCFA producers (e.g., Bifidobacterium adolescentis, Eubacterium rectale, Anaerotruncus colihominis), while HC-enriched bacteria included lactate producers (Lactobacillus gasseri, Lactococcus lactis) and Enterobacteriaceae species. MetaCyc pathway analysis showed 78 pathways more abundant in HC vs. only 11 in LC (P<0.05, FDR<0.15), including lactate production pathways and aerobic respiration pathways. D-lactate dehydrogenase (D-LDH) was significantly increased in HC (P=0.029). Urine metabolomics in a subset of patients confirmed significant positive correlation between lactate and C. albicans abundance (partial Spearman P=0.015, R²=0.57). In vitro competition experiments demonstrated that C. albicans significantly outcompeted S. cerevisiae in glucose+lactate medium under aerobic conditions (GLM P=0.004) and microaerobic conditions (GLM P=0.019), with 3 of 4 C. albicans strains showing increased growth advantage in glucose+lactate vs. glucose alone under microaerobic conditions.
**Clinical Implications:** This study provides a comprehensive ecological framework for understanding Candida overgrowth in the human gut, identifying a dysbiotic signature characterized by reduced SCFA-producing anaerobes and increased lactate-producing, oxygen-tolerant bacteria. The high accuracy of machine learning models in predicting Candida levels from bacterial data alone suggests potential for microbiome-based risk stratification. The finding that C. albicans can utilize lactate as a carbon source to outcompete other fungi under low-oxygen conditions offers a mechanistic explanation for Candida expansion in dysbiotic guts. These insights could inform prophylactic microbiome-driven strategies for patients at high risk of candidiasis, such as those receiving chemotherapy or broad-spectrum antibiotics. However, the study is limited by the lack of direct oxygen measurements in the gut, the absence of systemic candidiasis cases during the study period, and potential confounding by immunotherapy treatment. Longer follow-up studies are needed to delineate the transition from overgrowth to dissemination.