
Cancer drugs are typically tested against tumor cells under controlled laboratory conditions. But inside the body, those cells exist within a far more complicated ecosystem—one that includes nutrients, metabolites, immune cells, and, particularly in colorectal cancer, a diverse community of microbes.
A new computational framework aims to bring some of that complexity into cancer drug discovery by predicting how the microbiome may alter the effectiveness of cancer therapies.
Researchers combined machine learning with genome-scale metabolic modeling to identify potentially effective drug combinations while accounting for interactions among colorectal cancer (CRC) cells, therapeutics, and microbes. The approach, called Onco-Microbiome-GEM-ML (OMG-ML), could ultimately help researchers identify therapies tailored not only to a tumor’s molecular characteristics but also to a patient’s microbiome. This research is published in Cell Reports Methods.
“Cancer cells are not just on their own,” senior author Sriram Chandrasekaran, PhD, of the University of Michigan, told Inside Precision Oncology. “There’s a rich set of microbes that are in the colon environment, the microbiome, and then the metabolic environment itself is pretty complex.”
The goal was to develop a virtual model that better captures the complex interactions among cancer cells, microbes, their metabolic environment, and the drugs used to treat the tumor.
Adding the microbiome to drug prediction
Microbiome composition varies considerably among patients and has been linked to differences in cancer progression, chemotherapy resistance, and treatment outcomes. One organism of particular interest in CRC is Fusobacterium nucleatum (F. nucleatum or Fn), an oral commensal that is uncommon in the healthy colon but frequently enriched in colorectal tumors.
Because microbes can alter cancer-cell metabolism, the researchers hypothesized that these metabolic changes could reveal therapeutic vulnerabilities that might not be apparent when cancer cells are studied alone.
OMG-ML combines genome-scale metabolic models, which computationally represent metabolic reactions within cells and microbes, with machine learning. The model was trained on 6,514 drug combinations tested in microbe-free CRC cell lines and then extended to predict drug interactions when specific microbes were present.
“For this paper we looked at more existing therapies to see, can you better match existing therapies to the cancer, and also take into account the microbiome involved,” Chandrasekaran said.
The model identified several potentially synergistic combinations, including cabazitaxel plus megestrol acetate. It also predicted that methotrexate and fluorouracil could become more effective in the presence of F. nucleatum. Other drugs, including thalidomide and pralatrexate, showed effects that varied depending on which microbial species were present.
Such differences could eventually support more personalized treatment selection by incorporating microbiome composition alongside other tumor and patient characteristics.
A surprising role for Fusobacterium
One of the more unexpected findings involved F. nucleatum itself. Although the bacterium has been associated with CRC progression and treatment resistance, the model predicted that under certain conditions it could increase sensitivity to therapy.
“This bacteria is bad. But then, when you add chemotherapy drugs, this bacteria actually enhances the potency of chemotherapy drugs,” Chandrasekaran said. “So it could be good in some ways.”
To test the computational predictions, the researchers used an asymmetric co-culture system designed to overcome a challenge of studying CRC cells and anaerobic bacteria together: Cancer cells require oxygen, whereas F. nucleatum grows without it.
The bioengineered system created an oxygen gradient that allowed CRC cells and bacteria to grow under their preferred conditions while remaining in contact, more closely mimicking conditions in the colon.
Experiments in HCT116 colorectal cancer cells confirmed predicted synergistic drug combinations in the presence of F. nucleatum. The researchers also investigated other microbes, including probiotics, and found that their effects on treatment could differ.
Mechanistic analyses identified phospho-inositol metabolism and cysteine transport as key determinants of F. nucleatum-dependent drug synergy, offering potential biological explanations for the observed effects.
The framework also generalized beyond chemotherapy. Under immunotherapy-associated conditions, the researchers found that F. nucleatum could enhance the effects of certain anti-PD-1 immunotherapy conditions.
Toward microbiome-guided precision oncology
The work remains a proof of concept and is not ready to guide patient treatment. But Chandrasekaran envisions a future in which a patient’s microbiome could be characterized from a stool sample and that information incorporated into computational models to help predict which therapies may work best.
Microbiome sequencing itself is already commonly performed, he noted. What is largely missing is the connection between microbial profiles and specific cancer therapies.
“People are looking at microbiome fields. They are characterizing the bacteria in the stool samples,” Chandrasekaran said. “But people haven’t connected the bacteria to therapy and specific cancers. I think that’s sort of what we put together here.”
The approach could also have applications in drug development. Pharmaceutical companies could potentially use the framework to investigate whether a new therapy’s effectiveness changes in the presence of particular microbes or identify patient subsets whose microbiomes might make them more or less likely to respond.
Chandrasekaran’s group is now working to incorporate immune cells into the models, adding another major component of the tumor microenvironment. The framework could also potentially be adapted to other cancers influenced by microbial communities, including skin cancer.
Ultimately, the research points toward a broader view of precision oncology—one that considers not only tumor genetics but also the biological ecosystem surrounding the cancer.
“It’s not just about the cell you’re trying to kill, but also the environment it’s in,” Chandrasekaran said.
The longer-term vision, he said, is something resembling a tumor “digital twin” that incorporates an individual patient’s microbiome and immune state to make more precise predictions about which therapies are most likely to work.





