Myeloid malignancies are blood cancers that arise when bone marrow cells fail to develop normally. This group includes acute myeloid leukemia (AML), myelodysplastic syndrome (MDS), and myeloproliferative neoplasms (MPN). Combined, they affect hundreds of thousands of people annually, and survival rates vary widely across subtypes.
The human microbiome, the community of bacteria, viruses, and fungi living in and on us, has been linked to cancer in other organs. Studies of solid tumors have shown that each cancer type harbors a distinct microbial signature, and that blood-based microbial signals can distinguish cancer patients from healthy individuals. However, whether a meaningful microbiome exists at the site of blood cancers, the bone marrow and bloodstream, was unexplored.
This study analyzed DNA sequencing data from 1,870 myeloid malignancy patients and 12 healthy donors to systematically characterize the circulating microbiome in leukemia and related blood cancers, exploring whether microbial content varies by disease subtype and whether it carries prognostic value.
The researchers used existing whole-genome sequencing data generated from blood and bone marrow samples of 1,870 patients collected at the Munich Leukemia Laboratory between 2005 and 2017. These samples had originally been sequenced to profile cancer mutations in human DNA. The team repurposed the sequencing reads that did not align to the human genome as a window into the microbial content of each sample.
Microbial DNA was identified using PathSeq, a bioinformatics tool that separates human and non-human sequence reads and classifies the latter by species. Because low-biomass samples like blood are highly prone to contamination artifacts, the team applied extremely strict filtering, ultimately retaining fewer than 1% of microbial reads after removing known contaminating taxa and species showing signatures of cryptic human origin.
The remaining reads were used to quantify bacterial, fungal, and viral content per sample. Statistical comparisons were made across disease subtypes (AML, MDS, MDS/MPN, and MPN), and machine learning classifiers were built to test whether microbial content could predict disease subtype or patient outcomes.
Visualizing microbial composition using dimensionality reduction showed that healthy donor samples clustered separately from all disease cases, confirming that the bone marrow microbiome of cancer patients is genuinely different. Furthermore, patients with the same disease subtype were more similar to each other than to patients with other subtypes, suggesting that each myeloid malignancy carries a distinct microbial signature.
AML patients had the highest relative abundance of Proteobacteria and the lowest bacterial diversity of all subtypes. MDS patients showed the highest overall viral burden, with Epstein-Barr virus (EBV) detected in 34% of patients. Trichosporon asahii, a fungal species known to cause mortality in blood cancer patients, was the most common fungal organism, found in 18% of all patients.
Strikingly, microbial content also correlated with specific cancer-driving gene mutations. Mutations in FLT3 and NPM1, two of the most common mutations in AML, were each independently associated with higher Proteobacteria abundance, while DNMT3A mutations were linked to lower bacterial diversity.
The current gold standard for predicting outcomes in MDS patients is the Revised International Prognostic Scoring System (IPSS-R), which assigns patients to five risk categories based on blood counts and chromosomal abnormalities. However, this system leaves considerable uncertainty, especially for patients classified as low risk.
The researchers found that EBV detection in the blood refined survival prediction within the low-risk IPSS-R group. Low-risk MDS patients with detectable EBV had survival outcomes statistically indistinguishable from the intermediate-risk group, while those without EBV performed as well as the very low-risk group. This suggests EBV status splits what appears to be a homogeneous risk group into two with meaningfully different prognoses.
Importantly, EBV status was independent of the standard IPSS-R components like hemoglobin, platelets, and chromosomal abnormalities, meaning it provides genuinely new prognostic information not captured by existing clinical tools. This finding was validated using PCR on a subset of MDS samples.
To test whether the microbial differences observed across disease subtypes could be used diagnostically, the team built random forest machine learning classifiers using only the bacterial content of each sample as input features. These models were trained on 70% of patients and tested on the remaining 30%.
The classifiers achieved an AUROC of 0.87 for distinguishing AML from all other subtypes using bacterial genus abundances alone, with strong performance also for MDS (AUROC 0.84). A pan-subtype classifier that attempted to assign all four disease labels simultaneously achieved an overall AUROC of 0.84, demonstrating that the bacterial microbiome contains substantial subtype-specific information.
Among the most informative bacterial genera were Dermabacter and Kytococcus, both known to cause infections in immunocompromised individuals, as well as organisms associated with bacteremia. The finding that blast cell percentage tracks with microbial features suggests that the microbiome differences may partly reflect the extent of bone marrow infiltration by cancer cells rather than being fully independent phenomena.
This is one of the largest studies to characterize the circulating microbiome in hematological malignancies. The findings establish that the blood and bone marrow of leukemia patients contain a distinct microbial landscape that differs from healthy individuals and varies meaningfully across disease subtypes, gene mutation profiles, and patient outcomes.
The discovery that EBV status improves survival prediction in low-risk MDS has direct clinical implications. Adding EBV testing to standard risk stratification could identify patients who need more aggressive treatment despite apparently low-risk disease, potentially improving care decisions in this population.
More broadly, the study raises fundamental questions about the biological relationship between microbes and blood cancer. Whether microbes influence cancer development and progression, or whether the cancer environment shapes which microbes thrive, remains to be determined by future studies. Either way, the circulating microbiome emerges as a rich, underexplored source of clinically actionable information.